Our progress has been faster than I ever would have imagined, and this was a giant year for us. Across the Shaper ecosystem, one company went from 7-figures to 9-figures of revenue this year, one company won a $25-million contract as its first major deal, and one company had a successful exit that yielded a nearly 100x return for Shaper. Across the companies, we’ve raised >$150 mm. and hired 300+ employees. And I think we’re just getting started!

In each of the last two years, I’ve written a set of reflections on what I learned (see the posts from 2024 and 2025). Here are 10 lessons from Year 3.

Lesson 1: In the early stages, create surface area.

I’ve watched several early-stage Shaper companies ramp at meaningfully different speeds.

The companies that made the most progress finding product-market fit maximized their surface area to learn from the market: publicly building, iterating quickly, getting in front of customers, and trying to sell something from day 1.

Early-stage companies can spend a lot of time debating product, pricing, network effects, or long-term strategy. Those questions matter, but early on it is usually impossible to answer them in the abstract. Motion creates information.

There are lots of tactics to create surface area and motion, and it often involves decisions that feel “short-termist.” Go mid-market for faster sales cycles, even if they aren’t your ideal long-term customer. Do some things that aren’t automated, even if there won’t be a way to scale it up. Sell through channels if they get you to more customers, even if it might one day create channel conflict. Build in public and get hype, even at the cost of competitors taking your ideas. All of these tactics increase your speed of getting market feedback, and get your business into fast-iteration mode sooner.

At an extreme, within 2.5 years, Protege became a 9-figure revenue business — which is the fastest a company I’ve been involved in has ever made it to this stage. The key was creating surface area.

Lesson 2: Services businesses can be strong businesses too.

There is a lot of groupthink against services in the startup world. Some of it is reasonable: services businesses can be hard to scale and hard to defend. But I think a lot of this thinking is outdated, and the bias against services creates interesting opportunities to simply do the thing customers actually want.

Fractional AI was the clearest example. Most VC-backed companies entering the forward-deployed engineering world were searching for the productized angle. Customers, however, wanted exceptional engineers to help them solve important AI problems. We focused on what customers wanted, and we believed there was a $100 billion opportunity to get this right. Fractional AI still had to think carefully about scalability, margins, and defensibility, but it was fine to build a services-focused business with eyes wide open about how to build around those, and we ultimately got a home-run by staying true to this vision of the world.

Lesson 3: Know when to ignore VCs.

Investors can be the worst culprits in perpetuating groupthink. When you fundraise, you’ll get 50 different opinions on your business, and more often than not, rejections. It is always worth understanding investor feedback and understanding the signals and insights that come from it, but founders should remember that most VCs are quite often wrong, especially when you have a novel point of view.

Personally, I started to formulate this view with the LiveRamp and Datavant journeys. They were both home-runs for the VCs that invested, but were not especially popular businesses with investors. Both were category-creation plays, and 99% of VCs looked at a nonexistent market and concluded “small TAM,” rather than seeing the opportunities I believed existed to create a winner-take-most, network-effect-intensive new market.

I gained more conviction watching several fundraises this year: businesses that fit the zeitgeist often had an easy time, while some companies with phenomenal underlying business models found it surprisingly hard to raise.

Fun fact: six months before Fractional AI was acquired, we struggled to raise a round. Most VCs wanted to push us toward productization and didn’t believe in a services vision. Listening to the pushback, pressure-testing it, and then deciding where we still had conviction was valuable. VC feedback is a useful signal on the business model, but that’s it.

Lesson 4: Hire someone focused on “ops” early.

Most founders are focused on all their first hires being either “builders” (engineers, product) or “sellers.” Within the first ~5 hires at most Shaper companies, we try to place a generalist Head of Business Operations who owns some mixture of recruiting, people, operating cadence, and making sure the trains run on time. I’ve come to see this as one of the highest-leverage early hires, and have never seen the hire happen too early.

I’ve now watched several Shaper companies go from startup to ~100 people in roughly two years. That is a hard ramp, and companies will generally break if they don’t think methodically about recruiting, culture, onboarding, and basic process before they need it. A superstar in this seat figures out how to build enough infrastructure so the company can keep moving quickly as it grows.

Lesson 5: Build a culture of irrational over-delivery, even when it’s hard.

There are always rational reasons not to overdeliver: the customer is annoying, the deal was underpriced, the company has too much on its plate, or the product is harder than expected. It is appropriate to debate all this when deciding whether to take a deal or sell a particular product. Once you’ve sold it, though, I think the best CEOs take an uncompromising view: you have to overdeliver.

This becomes cultural. If quality is continuously traded off against other priorities, every team develops an explanation for why it is okay that a particular customer is unhappy or another deployment is late, and eventually none of your customers actually love you. The best companies do the opposite: going above and beyond becomes part of the culture, and the team rallies around delivery even when it is hard or not perfectly scoped.

Lesson 6: Keep burn low.

We are in another period where many startups can raise very large rounds relatively easily. Even in this environment, having a low burn rate (and the optionality to flip toward profitability or use the balance sheet) is incredibly valuable.

Low dilution shows up in returns: “skipping” a round dramatically changes founder returns from companies (and many of the wealthiest entrepreneurs, eg. Bezos, Gates, Zuckerberg, etc., were able to raise relatively little in their early days and effectively skip rounds).

But beyond the financial logic, there’s a much bigger strategic reason: operating discipline is hard to install late in a company’s culture, and disciplined companies run more effectively. Lean companies prioritize better, avoid simply throwing people at problems, and often move faster because they have fewer handoffs and owners. Treat headcount growth as unfortunate, even when capital is unlimited.

Lesson 7: Know when to stop working on an idea.

Over and over again, I’ve seen that the worst trap entrepreneurs get into is “slight/medium” success — not a hockey-stick journey, and not a clear failure that is easy to walk away from. This can lead to entrepreneurs spending years of their life stuck, with a massive opportunity cost. This is compounded by guilt (letting down investors, employees, customers) and an intertwined sense of identity. The standard motivational stories (“we bootstrapped for 12 years, then found success!”) make it even more difficult to get to no.

Earlier this year, the CEO of Zenith Health decided to shut down her business. This was a very hard, and very correct decision: she is a phenomenally talented entrepreneur with a huge opportunity cost, and — while there were some signals of progress — the timeline to scale and the amount of early traction just didn’t justify more investment. So we navigated a shutdown, and helped the team land as gracefully as possible.

In this situation, my job is to listen, and help the CEO with their reflections. But making it an acceptable option to move on is an important part of my job.

Lesson 8: My job is recruiting.

Last year, I wrote that ~80% of my job is recruiting. This is still the case.

Over and over, the highest-leverage thing I can do for a company isn’t a stroke of strategic genius; it is getting the right person into the right company, in the right seat, at the right moment. This year, we’ve increasingly turned Shaper itself into a recruiting machine — helping companies build great early teams and maintain very high talent density as they scale. That is usually more valuable than anything I can contribute in a meeting. [By the way — We’re hiring!]

Lesson 9: Build a good business and exits will come.

Founders are often asked, “Who might acquire this?” or “What is the exit strategy?” I don’t think either is a particularly relevant question to spend much time on. While it is good to know the market, and it can be useful to build relationships over time with big companies that are looking at the space, that’s not what drives M&A.

Both Datavant and Fractional AI reinforced that the timing of an exit, and even the identity of the buyer, can be a total surprise. What the two had in common is that we were building really good businesses, thinking long term, and not in a hurry to sell. That also creates the best negotiation leverage that leads to the best outcomes: if you don’t need to sell, a buyer has to pay a price that convinces you to change course, and you can genuinely walk away if the deal doesn’t make sense.

Lesson 10: Alpha lies between business models.

One of my pet peeves is people trying to put a clean label on Shaper. We aren’t exactly a fund, a family office, a venture studio, or an operating company. We sit somewhere between all four, and in different situations we lean in different directions.

I increasingly think that refusing the label has created alpha. Labels come with rules and conventional playbooks; staying in the gray lets us choose the right structure for a particular company and remain patient across a wide range of outcomes. Fractional AI is a good example: we were comfortable with different financing paths, levels of involvement, and exit scenarios, and that flexibility helped us make better long-term decisions.

* * *

Overall — what an incredible third year. Thank you to everyone across the Shaper ecosystem who made it possible. It still feels like we are at the beginning, and I’m even more excited for the year ahead.

A few thoughts on defensibility right now:

1. Most enterprise software companies rely on “stickiness” as their moat, which is a very weak moat. From ERPs to CRM companies to vertical SAAS, the dominant thinking over the past decade has been to move fast, get “customer lock-in”, and customers won’t migrate off you because the switching costs are too great (workflow stickiness, data stickiness, training stickiness, compliance stickiness, etc.). The issue with this logic is, while high switching costs prevent companies migrating for a small improvement in ROI, when there are step-functions in ROI, migration is worth it. AI is one of those step-function moments for many industries, so anyone with “switching cost” or “stickiness” as their primary moat is vulnerable.
2. Stronger forms of moats exist, but are rare in enterprise tech. In the consumer world, most of the biggest tech businesses have strong moats — either a network business model (eg. Facebook) or a marketplace business model (eg. Amazon) — that are extremely difficult for a new company to displace. These are much more rare in B2B — where the market has traditionally rewarded more end-to-end workflow ownership than defensibility. But they do exist; for example, as I wrote in this piece, in the data world, “data currencies” and “data marketplaces” have relatively strong network effects, but good curation and dashboards don’t. As a result, some companies are much more subject to disruption from AI-native approaches.
3. The companies that have strong moats will be beneficiaries of AI. So far, the market has punished all enterprise tech in its first reaction. But my bet is that while the companies with weak moats will be automated away over time, the companies where there are strong intrinsic network effects will perform even better. In the immediate term, AI will reduce their costs (while the network effect protects their revenue); but in the long-term, strong moats that provide durable distribution advantage will be one of the most valuable assets in the AI boom.
4. This is all true for startups too. I’m fairly skeptical of a large number of AI-native businesses that have had explosive growth, but limited defensibility (eg., does anything in the AI coding space actually have defensibility?). I’m even skeptical that the foundational models have real defensibility; “switching costs” as a model gets to know a person/enterprise/workflow seem to be the main argument for how the models will not be commoditized over time. To endure, startups either need to build with a strong network effect from day 1, or quickly have a “phase 2” with a moat after they quickly establish market share — or the forces of creative destruction will come for them too.

In the long-term, efficient markets will win, and things that don’t have moats will see their margin disappear. This has taken a while for enterprise software that sees “switching costs” as their main moat, and is finally about to happen. But it’s not a blow to the whole enterprise software industry, as the companies with real moats will thrive in this era.

Background on the Explosion of the Connectors Space

A decade ago, very few startups were in the business of building data connectors; MuleSoft was seen as a contrarian with a small TAM for building in the space, and the other middleware players (eg. TIBCO) were all considered “legacy” businesses. That has changed recently: just this week, the largest data connector deal ever was announced, and there has been an explosion in both horizontal connector businesses (ranging from Zapier to Airbyte) as well as connector businesses in every vertical (from Redox in healthcare to Catena in telematics to Segment in marketing). This explosion of the connector space is due to three factors:

1. The increase in the value of data
2. The explosion in the number of places where data resides
3. The explosion in the number of applications that exist that use data

Failure Scenarios of Connector Businesses

Several of these businesses are currently “hot,” but there are many more failed “Plaid for X” models than there are successes.

Most connector businesses fail, likely due to:

1. They can’t get the most valuable integrations
2. An industry doesn’t actually need lots of integrations
3. “Connectors” become a mask for low-margin services rather than repeatable technology
4. The company builds lots of integrations with cool technology, but so do 5 other connector businesses — so they fight relentlessly on price

The 6 Questions for Successful Connector Businesses

Against this backdrop, I have landed on six questions I ask about a “connector” strategy.

1. Are integrations “proprietary”? For a connectors business to be long-term viable, it needs some of its connectors to be defensible via “proprietary” integrations. This takes two forms as connector businesses get big: 1) some data sources that don’t have an “open API” allow you to establish a proprietary integration (e.g., Plaid’s integrations with banks), and/or 2) companies integrate into you (e.g., Okta’s integrations with applications). While you may start with “open APIs”, you cannot build defensibility on these alone; as you get scale you ideally take advantage of that scale to break into proprietary integrations. At LiveRamp, for example, we focused on integrating with every media platform; at the time, about half had a standard integration template, but about half would only prioritize integrations if a client pushed for them. This created a network effect in the business as winning clients helped us win more integrations.
2. Do customers use one connector tool or many? “Connector Gravity” is important for long-term defensibility; if customers decide integration-by-integration which connector tool to use, you have much less defensibility and stickiness. For Datavant and LiveRamp, customers needed a common ID, so it made sense to use one source of integrations and there was a high level of gravity; similarly, SSO from Okta has a high level of gravity because of the user experience. But many other integrations businesses have low gravity and are forced to compete connector by connector.
3. How many integrations does a typical customer have? The more integrations a customer has, the better. If it’s only a handful, it’s hard to have a large business. Most enterprises use thousands of vendors that hold or need data, though, so many spaces have room for lots of integrations.
4. How hard is it for the end customer to build connectors themselves? The upper bound of value a connectors business creates is the cost for an end customer to build an integration in house. The more complex an integration is, the better. In both the Datavant and LiveRamp businesses — the first integration was very difficult to build, the next 20, somewhat difficult, then increasingly easier, until integration 100 was almost automatic as we relentlessly templatized. This made the “build vs. buy” calculus very clear; it made no sense for a customer to build something in house.
5. Are integrations reusable? The economics of a connectors business work best if you can amortize an integration across many clients by building the integration once, then reusing it across clients. This varies based on the industry and type of integration.
6. Is there a long-tail of valuable integrations? In some industries, the dominant player has so much market share that connectors are less valuable. For example, there are parts of healthcare where Epic has so much market share, that integrating with the “long-tail” doesn’t create enough value. The more fragmented the industry, the better it is for connector businesses.

As data gets used more and more for AI applications, I expect more and more niche connector businesses will emerge. In most cases, these won’t succeed — as they won’t meet the criteria above and they’ll end up being commoditized away. But I expect at least a dozen more connector unicorns will be formed this decade when entrepreneurs are thoughtful about these questions.

Disclosures: I am a shareholder, board member, and advisor to many “connector” businesses.

It’s been a fast start and fun adventure so far, as the six companies we’re building (Arbital, Protege, Fractional, Zenith, Adelphi, and Catena) have all gained incredible traction rapidly. It’s hard to imagine what we’ll accomplish in the next two years, and even harder to imagine the next twenty!

A year ago, I laid out a few lessons from the first year running Shaper, and comparisons to my experiences leading Datavant and LiveRamp. As a follow-up, this post walks through four key takeaways from our second year.

Lesson 1: ~80% of my job is recruiting.

This shouldn’t have surprised me: every time I’ve led a company scaling extremely quickly, talent has become the bottleneck. I see it as the biggest challenge and biggest opportunity across the portfolio: recruiting world-class engineers, building out an incredible exec team, and attracting incredible generalists.

Recruiting is a long-game, and requires asking for lots of referrals, building long-term relationships, and aspiring to match the perfect person to the perfect role at the perfect company at the right moment in their life. It also requires each Shaper company to have a team that is equally obsessive about talent density, building strong cultures and pouncing on talent opportunistically.

Over the last year, we’ve experimented with a number of hacks to be able to hire more effectively across the portfolio. For example, we’ve hired at the Shaper level, with the intention of superstars rotating into (and ultimately joining) portfolio companies as they find the perfect fit. I tried to introduce this idea with a single “Special Projects” hire ~9 months ago, and we found so much incredible talent that I ended up “accidentally” hiring three people instead, and we’ve since expanded the program into a generalized “Builder-in-Residence” program. This has become a pipeline of great talent for the portfolio and I think is one of the biggest areas we’re adding value. Similarly, we’re building out “CFO-in-Residence” programs to build a pipeline of up-and-coming CFOs who can support the portfolio fractionally to start and grow into Head of Finance roles as they match within the portfolio.

Meanwhile, we’ve rushed to make recruiting a core competency — adding a talent leader at Shaper as an early hire. And everyone on the Shaper team spends a huge percentage of time talking to potential candidates. We’re obsessing over talent density, and it’s been a big investment so far, but I believe it’s already paying off.

Lesson 2: Quality over quantity.

This year, we did not launch any new companies, instead choosing to double down on the six we launched in our first year. This required discipline, and I think was the right choice given the momentum of the current portfolio.

Even though many people put us broadly in that category, I am generally a skeptic of the “venture studio” business model. The main issue is CEO adverse selection (why would a great CEO accept the equity dilution?); my approach to avoid the adverse selection is by staying in my lane and focusing on businesses where I’m a natural cofounder (i.e. businesses where I believe I can materially change the odds, size, and speed of success) and also by ensuring I can meaningfully be involved in each business. To that end, a small portfolio of exceptional businesses where I meaningfully invest time is dramatically more valuable than a large portfolio.

Lesson 3: The best companies simply move faster.

The biggest commonality in the most successful companies I’ve seen is simply the speed they move. The pace that they reply to emails, follow-up with clients, return redlines, drive interview processes, make decisions, launch products, and drive execution always surprises me, and is in sharp contrast to “average” companies. This could be correlation (Hungrier teams? Higher throughput talent?), but I expect it’s at least partly causal, for two reasons:

1. success compounds exponentially, and being a few days/weeks/months faster to close the deal/sign the candidate/win the opportunity compounds
2. most decisions in a company are actually “microdecisions”, rather than big decisions — and a culture of velocity enables many more microdecisions to be made — increasing organizational throughput materially

Lesson 4: Traditional speed limits seem to be disappearing.

One metric that astounds me: Two companies in the Shaper portfolio each crossed $10 mm. run-rate within 15 months of incorporation. For comparison: LiveRamp took six years from incorporation to $10 mm ARR, and Datavant took ~3 years from launch to $10 mm ARR. The opportunities surrounding AI are unlocking faster growth than ever as the biggest economic shift of a generation takes place.

And beyond revenue, we’re seeing a number of other milestones happen at unprecedented paces. One Shaper company signed an 8-figure TCV deal in its first year of existence (with a 4-week sales cycle!). Another company did a 10x, 7-figure land-and-expand upsell within 3 months of landing an initial deal. The velocity and magnitude of opportunity that exists in this moment is simply unprecedented in my career, and again favors companies with incredible talent density and company velocity to capitalize on it.

***

Overall — what an amazing first two years we’ve had so far, and I’m even more excited for the year ahead!

From my perspective, I see this as the beginning of a war between foundational models for the right AI training data capabilities.

This makes sense: each foundational part of AI infrastructure is currently undergoing massive capital infusion and seeing massive winners. These are the core bottlenecks to AI’s progress:

• Chips (NVIDIA is the trillion dollar player)
• Models (OpenAI and others are all aiming to be the trillion dollar players)
• Electricity & data center capacity (hundreds of billions of dollars of investment in each category)
• Training Data (no clear giant player/investment…yet)

The data market has been ignored as a trillion dollar bottleneck, largely because the first phase of training came from the open web and was therefore initially free. Scale AI has been the early leader in training data by moving from crawlable data to human-generated data, but this is still scratching the surface of the data needed for AI.

With Meta making this move, the first major capital outlay has happened around data — and I believe it’s just the first shot in a much broader war between foundational models as data is more and more widely seen as a foundational bottleneck.

Where Scale AI fits in the data map — the 4 Approaches to Training Data.

There are currently four approaches to gathering training data.

Approach 1: Crawling the web.

The oldest source of training material is the public Internet. Large language models have long feasted on massive web crawls and public databases. This has been an abundant, cheap way to get started — but is ultimately limited. Publishers and creators are clamping down on automated scrapers, and authors and publishers are pushing back. Even if the information were all available without copyright concerns, the internet is a small minority of human knowledge and not close to enough to satiate the hunger for data by LLMs (there is much more information that I’ve written that is in my emails, texts, and documents than on the open web, and that’s true for every person and organization). And in nuanced areas (legal contracts, medical images, etc.) the amount of available public data is extremely small.

Approach 2: Human Generated Data

Faced with the limitations of public data, human generated data has become a popular addition to training. Often, this comes in the form of data annotation. Machines struggle with raw, unstructured inputs — they need examples tagged by humans. Scale AI specializes in that: it combines algorithms with “human-in-the-loop” annotation to turn gigabytes of images or text into clean training sets. A self-driving car maker will pay workers to draw bounding boxes and labels on millions of street photos; a medical AI company will hire experts to highlight tumors, fractures, or other pathologies.

Other datasets may be created specifically for AI, whether that’s commented code to tackle a very specific problem or recordings of scripted prompts read in different languages, accents, or emotional tones.

This human labor is expensive but critical — hence the avalanche of money pouring in. Meta’s check to ScaleAI is in part a bet that owning this human data pipeline will pay dividends in future model improvements. But even with massive investments — the amount of data that foundational models can pay humans to generate is tiny compared to the amount of data already generated. So Scale’s current approach won’t be sufficient to feed the massive need for data.

Approach 3: Synthetic Data

As public data is drying up & human-generated data too expensive at scale, one alternative proposed has been synthetic data — letting artificially generated data for AI that looks like real data.

While this has been popular for specific applications (and works well for proofs-of-concept and early training of a new model), synthetic data won’t work to fundamentally power LLMs — the logic of AI training on top of AI is obviously circular and has the risk of model collapse. So while synthetic data has its place, it still doesn’t solve the data shortage.

Approach 4: Proprietary Data (The Ultimate Goldmine)

All the above are tiny beside the prize of proprietary data; this is the real untapped trove. The earliest deals for proprietary training data have been focused on sources like Reddit, which have massive amounts of user-submitted content, and media sites that have high-value information. There are thousands of sources of information like this.

But a massive amount of human knowledge is in the form of proprietary, non-crawlable information. The information needed to train medical AI breakthroughs is in medical records and images; the information needed to automate contracting workflows is in millions of contracts; the information needed to optimize customer support is in millions of call transcripts; the information needed to architect houses is in millions of drawings; the information needed for AI to generate a feature-length movie is in other TV and video content; the information needed to know the latest science is in paid journal articles.

As AI becomes more and more specialized, thousands of these proprietary deals will need to exist (sometimes augmented with human-generated data). While the issue to date has been illiquidity of this information & difficulty to license it at scale, as the data shortage for AI becomes more acute more and more of these deals are being structured.

Data Wars Are Just Beginning

Meta’s deal with Scale AI is merely the opening salvo. Across the industry, companies are arming themselves on all four fronts: continuing to crawl the web, funding human labeling firms, experimenting with synthetic data pipelines, and entering deals for proprietary data. Expect this to become the next trillion-dollar category as the limitations of the open web set in.

[Thanks to Bobby Samuels at Protege AI for his help drafting this]

As AI emerges, an open question remains: which business models will dominate. Will foundational models and chips take all the returns for the industry, or will they commoditize? Are vertical apps a viable approach, or do they lack defensibility? Is the best approach just to automate legacy businesses?

From private equity and large incumbents to venture capital and small startups, the market is betting on 7 core approaches to monetizing AI. This post walks through each approach, and their distinct conditions for success and failure.

Approach 1: Build “must-have” infrastructure for AI

This category has seen the biggest winners of AI so far, such as OpenAI and NVIDIA. These are companies that aim to have every AI use case built on top of them, and ultimately take a royalty from all AI usage across the economy.

This is a tough category to break into (there are only a handful of true “must-have” components of the stack), and questions around ultimate defensibility for each layer of infrastructure (will there be one foundational model that dominates or many?) — but it seems a safe bet that the single-biggest winners of AI will be a handful of foundational infrastructure players.

Approach 2: Build a vertical AI application (point-solution)

Most AI startups are vertical apps built to bring AI to X workflow, and there are hundreds of new startups being minted each month with this approach. The last few years have seen a handful of these take off with massive adoption and revenue growth — for example Abridge (healthcare), Cursor (coding), Harvey (legal services) are growing at extreme paces.

The core challenges in this category are around distribution and defensibility. So many areas are so easy to disrupt right now, that there are hundreds of companies simultaneously building “a platform to automate X.” At the same time, for most verticals, there’s an open question about whether the horizontal infrastructure players will ultimately solve most of the use cases themselves. So competition is extremely fierce. In order for vertical AI companies to ultimately thrive, they need to be in a market with a clear moat.

Approach 3: Build an AI-native company that solves a legacy business end-to-end

Rather than build a “point solution” company, some entrepreneurs are choosing to build an end-to-end service that competes with incumbents — but with a cost structure and features associated with an AI-native business, rather than a traditional business model. This tracks the dot-com boom (some of the big winners were internet-native remakes of existing business models, like Amazon.com for books).

This has been most prominent in areas like pharma, where companies focused on AI for drug discovery have chosen to become end-to-end pharma companies (that discover, develop and commercialize drugs) rather than monetize through licensing the AI itself. But it has spread beyond that; most industries are seeing a new set of AI-native competitors taking on the incumbents.

The main challenge with this approach is that legacy businesses typically have fairly strong structural advantages (such as distribution moats or customer stickiness), so even if AI can provide a 20% advantage in cost structure, that’s insufficient in many cases to overcome the existing moats of incumbents. So this only works in industries where the benefits of AI are so massive in the near-term that they can overcome the structural advantages of incumbents quickly.

Approach 4: Large incumbents automating portions of their business

One of the superpowers of generative AI so far has been the ability to automate bespoke workflows. From call center optimization to coding automation to customer service chatbots, early adopters are already seeing AI as a way to fundamentally change the cost structure of existing business workflows while keeping the distribution advantages of incumbency.

In addition to this approach being popular with existing incumbents, it is also a popular bet within private equity: acquire incumbents in industries that are ripe for AI automation, and become the first to automate. The flurry of deals around industries like accounting services, for example, are likely a bet that some of the work can ultimately be automated by early adopters.

The math on this is simple: let’s say a product generates $100 of revenue, and costs $80 to make (fully-loaded). Let’s say you can use AI to automate away 25% of the human effort; now it costs $60 to make. The company can then double its profit (from $20 to $40 per unit) — dramatically changing the business model.

The big assumption here is that prices don’t fall for incumbents as everyone takes advantage of the AI opportunities; this depends on the structure of the industry and the size of the distribution moats. The other challenges here are:

1. Navigating the change management needed to be an early-adopter is harder for large incumbents
2. Large incumbents typically struggle to get top-tier AI talent to lead a technical transformation

Approach 5: Combining large incumbent with AI-focused startup for automation

As a modification of the strategy above, another approach is to combine legacy incumbents with vertical AI startups for automation. This captures the same economics, but it can solve the talent gap by getting talent via acquisition.

This is especially attractive in industries where the AI tech quickly commoditizes so there are dozens of vertical AI players (so none of them can command a massive acquisition price).

The viability of this approach hinges on integration — it is complex to integrate AI-tech startups and legacy incumbents, and the thesis here involves retaining talent. Without exceptional integration, acquisitions typically don’t play out as well in practice as they do on paper. I suspect that the best PE firms in the coming decade will master this integration playbook and look for a number of acquisitions with this thesis in mind.

Approach 6: Combining large incumbent with AI-focused startup for distribution

Another acquisition-centered thesis is to combine an incumbent’s distribution moats with a startup’s AI technology. Rather than focusing on cost-cutting, this approach involves increased sales by offering a fast path-to-market for new tech.

I expect many big enterprise tech companies will do this (for example, Salesforce or Oracle acquiring companies to bring their tech into enterprise), but there are also many vertical-specific approaches. For example, in healthcare, scaled electronic medical record companies like Epic can bundle “AI for X workflow” products and take advantage of their existing distribution advantages.

A large amount of public company M&A and PE-driven M&A deals are built around variations of this thesis. The main challenges around this approach are:

1. making sure there really are co-sell distribution synergies,
2. making sure the integration is successful
3. moving fast enough as the AI market moves quickly, and
4. ensuring that the distribution advantages are long-term as the market changes.

Approach 7: Build a complement to AI

Some of the most interesting hypotheses for making money with AI come from theories of what products will be “complementary.” Rather than building an AI business, some companies are betting on non-AI products where demand will increase as AI demand grows.

Some examples of bets around this include data (proprietary data is more valuable in a world where AI processes it successfully), talent (elite talent is more valuable as it gets more leverage from AI), power (AI workflows will consume massive amounts of electricity), cyber insurance (risk of data breaches goes up), and real estate in San Francisco (an AI boom will bring lots of talent into SF).

Of course, there are counterarguments on many of these hypotheses — talent could become less valuable, and the world could move away from San Francisco — so the risk becomes in picking the industries that are true complements.

***

Across these strategies, the trillion-dollar question is which will be the dominant approach. Consider an area like legal services. Will the biggest winner of AI be AI-tech companies that sell products to existing law firms? Or an AI-native law firm that incorporates AI from day 1 in its own workflows? Or will NVIDIA and OpenAI be the biggest beneficiaries even in specific vertical workflows like law and crowd out all vertical players?

Personally, I have bet the most on variations of #7 (looking at data and talent as complements to AI), would bet strongly on incumbents with strong distribution doing smart M&A (approach #6), and am generally skeptical of the amount of hype around vertical AI apps in approach #2 (where most of the investment is going now — but where there is usually light defensibility). Massive amounts of capital and talent are casting their own votes right now as this debate plays out across the economy.

Working with the team at Fractional AI, we’ve put together an overview of all the different use cases in healthcare where we’re seeing AI buzz, and where the buzz is getting real adoption. This piece explores 150 AI use cases — some currently in practice, others on the horizon — across three key sectors: Life Sciences / Pharmaceuticals, Providers, and Payors. You can also download a copy of the report here.

AI Use Cases for Providers

AI can reshape how healthcare providers deliver care, offering advanced clinical decision support, improving operational efficiency, and enhancing patient engagement. From assisting physicians in diagnosing complex conditions to automating administrative workflows, AI could help providers optimize both medical and business operations. Additionally, AI-powered patient monitoring, radiology, and surgical assistance tools can enhance care and safety. This section highlights AI’s potential impact on the Provider ecosystem.

Clinical Decision Support for Providers

1. Differential Diagnosis — Given a patient’s symptoms, history, and test results, AI can suggest a ranked list of possible diagnoses. This can help clinicians consider options (including rare diseases) they might not immediately consider.
2. Medical Imaging Diagnosis — AI can rapidly analyze X-rays, MRIs, and CT scans to detect abnormalities such as tumors, fractures, or infections, significantly improving diagnostic speed, consistency, and accuracy while reducing human error.
3. Pathology Analysis — AI can enhance digital pathology by quickly and accurately identifying cancerous cells, infections, and other abnormalities in biopsy slides, helping pathologists diagnose diseases with greater confidence and efficiency.
4. Comprehensive Patient History Summarization — AI can compile and summarize a patient’s complete medical history, highlighting key diagnoses, treatments, and test results. This would support informed decision-making, help prevent missed historical data points, avoid redundant workups, and streamline care planning.
5. Prescription Optimization Based on Insurance Formulary — AI could analyze a patient’s insurance plan to suggest medications covered under their formulary, ensuring cost-effective prescriptions and reducing delays caused by prior authorizations or non-covered medications.
6. Diagnostic Interpretation in Medical Imaging — AI can analyze radiology images such as X-rays, MRIs, and CT scans to detect subtle abnormalities, including tumors, fractures, or lesions. By highlighting areas of concern and suggesting potential findings, AI can enhance diagnostic accuracy, reduce human error, and support radiologists in making faster, more informed decisions.
7. Predictive Analytics for Disease Progression — AI can forecast how diseases like sepsis, cancer, or heart failure will evolve over time, enabling earlier interventions, personalized treatment plans, and better patient outcomes by anticipating complications before they arise. These predictive insights allow healthcare providers to implement preventative strategies before conditions worsen.
8. Risk Stratification — Machine learning models can analyze health records, lifestyle factors, and genetic predispositions to identify patients at high risk for conditions like diabetes, stroke, or heart disease, enabling proactive management and preventive care.
9. Personalized Treatment Recommendations — AI systems can analyze patient-specific factors (genetics, disease subtype, past treatment responses) and recommend tailored treatment plans. For example, in oncology, ML models might match a patient’s tumor profile with the therapy that has the highest success probability.
10. Medication Management — Clinical AI systems could review a patient’s medications, diagnoses, and labs to flag potential drug-drug interactions, optimal dosing, or suggest deprescribing unnecessary medications.
11. Dose Optimization — AI can recommend optimal dosing for medications (especially with narrow therapeutic windows or complex regimens) by analyzing pharmacokinetic models and patient data.
12. Antibiotic Stewardship — Decision support tools could analyze local microbiology data, resistance patterns, and patient specifics to recommend the most appropriate antibiotic (and duration) for an infection. This would fight antimicrobial resistance and improve patient outcomes.
13. Guideline Compliance Alerts — AI integrated with the EHR can check ongoing treatment orders against clinical guidelines and alert if something important is missing or if an order deviates from best practices. Physicians could also be alerted to updates, such as new medication approvals.
14. Clinical Knowledge Retrieval — When faced with a complex case, clinicians can query an AI assistant (powered by a medical LLM) to retrieve summarized relevant knowledge: practice guidelines, similar case outcomes, or new research findings. This on-demand consultation can augment clinician decision-making with current medical knowledge.

Administrative & Operational Efficiency for Providers

1. Medical Scribe — AI-powered ambient listening and mobile applications are streamlining medical documentation by transcribing and structuring physician-patient conversations into clinical notes, reducing administrative burden and improving workflow efficiency.
2. Automated Medical Coding — NLP algorithms can read clinical notes and suggest billing codes (ICD-10, CPT) for diagnoses and procedures. This speeds up coding and improves accuracy, resulting in faster billing and fewer denials.
3. Claims & Billing Automation — On the provider side, AI tools can pre-scan claims for errors or missing information before submission to payors, preventing denials.
4. Prior Authorization Assistance — When certain tests or treatments require insurer approval, AI could compile the necessary clinical data and even draft the prior authorization requests with justification, based on guidelines and the patient’s record.
5. Claims Denial Support — ML models could identify patterns in denied claims and recommend process changes or training to prevent them, predict which claims might be denied so staff can intervene, and even estimate expected reimbursement for services.
6. Scheduling & Staffing Optimization — AI could predict patient volumes (ED visits, OR cases, clinic no-shows) and help optimize staff schedules accordingly.
7. Operating Room Utilization — Machine learning algorithms could analyze surgical schedules, case length variability, and surgeon habits to suggest improvements in block scheduling and add-on case placements. By optimizing OR calendars, hospitals would increase surgical cases and revenue without adding rooms.
8. Patient Flow and Bed Management — AI could predict discharges and patient movement in the hospital, helping bed managers proactively assign beds and avoid bottlenecks. By anticipating needs, hospitals would improve throughput.
9. Supply Chain & Inventory Management — AI systems could forecast the usage of medical supplies and automate re-ordering, avoiding both stockouts and overstock. They could also optimize the supply distribution within a health system.
10. Front-office Scheduling — Front-office voice AI scheduling agents can handle patient appointment bookings, reschedules, and cancellations through natural language conversations. The AI assistant could check provider availability, confirm necessary visit details, and finalize appointments without human intervention.
11. Document Processing (Fax/OCR) — Hospitals still receive referrals and records via fax or scanned documents. AI-powered OCR and NLP could read these and integrate the data into the EHR or task lists (e.g., referral fax triggers creating an appointment).
12. Smart Alarms — Instead of every monitor beeping for every threshold breach, AI can contextualize and prioritize alerts (for example, ignoring a transient drop in oxygen if all other signs are stable). By making alarms smarter, AI can reduce alarm fatigue for clinicians so real issues aren’t missed.
13. Dental Insurance Verification — AI could automate dental insurance verification by instantly extracting patient eligibility, coverage limits, and prior authorization requirements from payor databases and EHR systems.
14. Emergency Response Coordination — AI could enhance emergency response by analyzing real-time data from 911 calls, wearable devices, and traffic patterns to optimize dispatch decisions and resource allocation. By predicting patient severity, suggesting the fastest routes for ambulances, and coordinating hospital readiness, AI would improve response times and increase the chances of successful medical interventions.

Patient Monitoring & Safety Tools for Providers

1. Chatbots for Patient Triage — AI chatbots for triage can use natural language processing to assess symptoms and recommend care based on medical guidelines. They can direct patients to virtual visits, urgent care, or self-care, reducing unnecessary ER visits and improving access to timely care.
2. Early Warning Systems (EWS) — AI-driven alert systems could continuously analyze patient vital signs, lab results, and nurse notes to predict clinical deterioration (e.g. sepsis, cardiac arrest) hours in advance. This would alert care teams earlier to intervene sooner.
3. Remote Patient Monitoring Analytics — For patients at home or in post-acute care, AI can analyze data from wearables, home medical devices, and patient-reported symptoms to identify warning signs. This allows providers to proactively reach out before a minor issue becomes an emergency.
4. Medication Adherence Monitoring & Reminders — AI could analyze patterns from “smart” pill bottles, pharmacy refills, or even digital biomarkers (like subtle changes in voice or activity) to infer if a patient is not taking medications as prescribed. AI-enabled applications would then intervene with reminders or support.
5. Mental Health Monitoring — AI-driven apps can monitor patient text messages, voice tone in phone calls, or social media (with consent) for signs of depression or suicidal ideation. These early flags can enable timely mental health interventions.
6. Readmission Risk Prediction — Before a patient is discharged, AI could predict the likelihood of 30-day readmission by looking at dozens of factors (condition, past visits, social determinants, etc.). Those at high risk would get extra care coordination, follow-up calls, etc.
7. Postoperative Complication Prediction — AI models could identify surgical patients at risk for complications like infections or readmission based on their intraoperative data and early post-op vitals. This would allow for targeted monitoring or preventive measures in recovery, improving surgical outcomes.
8. Post-Discharge Follow-up Bots — After leaving the hospital, patients may get automated check-in messages or calls from an AI bot asking about key symptoms, pain levels, or wound status. Detected issues would trigger a human follow-up.
9. Social Determinants of Health (SDoH) Analysis — AI could analyze social determinants of health (SDoH) by integrating medical, socioeconomic, and community data to identify risks like housing or food insecurity. It would suggest interventions, such as connecting patients to assistance programs, improving health outcomes, and reducing disparities.

Patient Engagement & Education Tools for Providers

1. Personalized Education — Generative AI systems could create easy-to-understand summaries of a patient’s condition, upcoming procedure, or discharge instructions tailored to their literacy level and language.
2. Mental Health Chatbots — AI chatbots can offer cognitive behavioral therapy exercises, mood tracking, and empathetic conversation for patients with anxiety or depression. They can provide 24/7 support or coaching, augmenting mental health resources.
3. Speech Therapy — AI-powered tools could assess pronunciation, fluency, and articulation, providing real-time feedback and personalized exercises for speech disorders. They would support remote therapy, track progress, and help speech-language pathologists refine treatment plans with data-driven insights.
4. Remote Consultations — AI could improve virtual visits by assessing symptoms, analyzing medical history, and automating documentation. This would support providers with decision aid, patient summaries, and seamless telemedicine integration.
5. Chronic Disease Virtual Coach — Patients with chronic conditions could use AI-based apps that act as a “coach,” giving tailored advice each day on diet, exercise, medication reminders, and motivational feedback based on their logged data.
6. Language Translation — AI translation services (speech and text) can allow providers and patients who speak different languages to communicate effectively. Integrated tablets or apps can do real-time medical translation for dozens of languages, breaking language barriers in care with high accuracy.
7. Voice Assistants — Smart speaker devices or voice assistants in hospital rooms could allow patients to ask for help, control the TV, or get information just by speaking. This would give patients more control and instant communication (especially those with mobility issues).

Surgery / Procedural Assistance & Radiology / Imaging Tools

1. Surgical Planning — AI could assist in preoperative planning by analyzing medical images, patient history, and anatomical structures to help surgeons optimize their approach. By generating 3D reconstructions, predicting potential complications, and suggesting the best surgical pathways, AI would enhance precision, reduce risks, and improve patient outcomes.
2. Robotic Surgery — AI could enhance precision in robotic-assisted surgeries by analyzing real-time data, improving dexterity, and minimizing human error during delicate procedures.
3. Real-Time Guidance for Procedures — AI could provide real-time overlays of critical information during procedures like laparoscopy or catheterization, helping clinicians navigate complex anatomy with greater accuracy.
4. Radiation Therapy Planning — AI could automate the time-consuming task of contouring organs and tumors on imaging for radiation treatment planning. AI could also help calculate optimal radiation dose distribution.
5. Image Reconstruction — AI could improve the quality of medical images by enhancing clarity in low-dose radiation scans, allowing for safer imaging while maintaining diagnostic accuracy. By reducing noise and refining details in MRI, CT, and PET scans, AI would help radiologists interpret images more effectively without compromising patient safety.
6. Quality Assurance in Imaging — AI could ensure medical scans meet quality standards before radiologists review them, reducing the chances of misinterpretation due to poor image resolution or artifacts.

AI Use Cases for Life Sciences Companies

Over the next decade, artificial intelligence will revolutionize the pharmaceutical industry by accelerating drug discovery, optimizing clinical trials, and enhancing commercialization strategies. AI-driven models can analyze vast datasets to uncover patterns and insights, improving efficiency, reducing costs, and expediting the delivery of life-saving treatments. Here we explore key areas where AI can make a transformative impact for Life Sciences / Pharmaceuticals companies.

Drug Discovery & Preclinical Development

1. Target Identification — AI models can analyze genomics and proteomics data to rapidly identify promising disease targets, uncovering novel biological pathways and expediting early-stage research.
2. Lead Optimization & Drug Design — Machine learning could refine molecular candidates by predicting ADME properties, ensuring favorable pharmacokinetics before testing.
3. Drug Repurposing — AI could mine literature and clinical data to identify new therapeutic uses for existing drugs, reducing development risk and time.
4. Predicting Drug Toxicity — AI could assess molecular structures to predict toxicities, mitigating safety risks early in development and designing safer drugs.
5. Biomarker Discovery — AI-driven analysis of patient datasets could reveal predictive biomarkers for precision medicine, improving treatment efficacy and trial stratification.
6. Polypharmacology Analysis — AI could predict unintended drug interactions, enabling the design of safer, multi-target treatments with fewer unforeseen adverse effects.
7. Novel Excipient Development — AI could help design and optimize excipients that improve drug solubility, stability, and bioavailability. This would be particularly useful for poorly soluble drugs, ensuring they achieve therapeutic efficacy when formulated into final products.
8. Predicting Drug Solubility & Permeability — AI could forecast how well a drug dissolves and absorbs in the body by analyzing molecular properties and environmental conditions. This enhances formulation strategies, ensuring optimal bioavailability and therapeutic effectiveness.
9. Organ-on-a-Chip Technology — AI could support advanced preclinical testing and automate lab experiments, increasing research efficiency. AI-powered robotic systems could help scientists design and run high-throughput biological assays.
10. Lab Experiment Automation — Robotics combined with AI planning could design and run preclinical experiments. AI would decide optimal experiment sequences to yield maximum information, increasing laboratory throughput in modern “self-driving labs”.
11. Intellectual Property Protection in Pre-clinical Research — In pre-clinical stages, legal teams safeguard intellectual property for new drug formulations, innovative testing methods, and breakthrough discoveries. AI could analyze prior art and patent databases to strengthen filings, identify potential conflicts, and help prevent IP disputes.

Clinical Development

1. Clinical Trial Design Optimization — AI could predict trial success probabilities, optimizes endpoints, and refines study protocols, reducing amendments and improving efficiency. AI-powered simulations would allow researchers to test multiple trial designs virtually before selecting the most effective approach.
2. Site Selection & Investigator Identification — AI could analyze historical trial data, patient density, investigator experience, and site performance metrics to recommend the best trial locations.
3. Patient Identification — NLP/ML can scan electronic health records, genetic databases, and social media to find patients meeting complex inclusion criteria in real-time. Additionally, LLMs could ingest lengthy trial protocols, then help staff quickly determine patient eligibility.
4. Patient Engagement & Education — AI can enhance patient trial engagement by delivering personalized education, reminders, and support before and throughout a clinical trial. Regular check-ins and adherence prompts would help minimize dropout rates, improve trial adherence, and ensure participants remain informed and motivated.
5. Synthetic Control Arms — AI can leverage real-world data to reduce the need for placebo groups, cutting costs and ethical concerns. AI-driven synthetic controls are increasingly accepted by regulatory bodies as an alternative to traditional control arms.
6. Digital Twins — AI could create virtual patient replicas to simulate biological and clinical responses using multi-omics, real-world data, and machine learning. In trials, digital twins can optimize patient stratification, predict efficacy, and reduce control group needs by modeling treatment responses.
7. Automated Data Monitoring — AI systems could continuously monitor incoming trial data to detect anomalies, protocol deviations, or safety signals in real time. This would ensure faster identification of issues, enabling corrective actions.
8. Automated Clinical Data Entry — AI (including computer vision and NLP) could auto-extract data from source documents, like lab reports or physician notes, into electronic data capture systems. This would reduce manual data entry errors and workload, speeding up the data cleaning process with current OCR/NLP capabilities.
9. Automated Trial Amendments — AI could detect and flag all necessary changes across trial protocols, databases, and stakeholder communications when a study is modified. AI would update clinical trial management systems, alert investigators, patients, and regulatory bodies, and generate amendment-ready documentation.
10. Clinical Study Report Generation — LLMs could assist medical writers by summarizing trial results and patient narratives into study reports and submission-ready documents.
11. Intellectual Property Protection in Clinical Trials — Legal teams need to ensure that intellectual property (IP) created during clinical trials is adequately protected, including filing patents and managing IP rights. AI tools could assist by monitoring patent landscapes and helping patent attorneys draft robust patent applications.
12. Informed Consent and Ethical Compliance — Legal teams ensure that clinical trial protocols adhere to ethical standards and regulations governing informed consent. AI could help by automating consent document generation and tracking compliance.

Drug Safety / Pharmacovigilance & Regulatory Tools

1. Post-Market Drug Safety Surveillance — AI could scan post-market data (e.g. FDA FAERS database, EHRs, social media) to detect early signals of adverse drug reactions. By finding unusual clusters of symptoms or events, AI would alert safety teams faster than traditional monitoring.
2. Automated Safety Reporting — Chatbots and guided forms powered by AI could help clinicians and patients report adverse events with higher-quality data.
3. Intelligent Case Prioritization — Machine learning models could predict which adverse event reports are likely true signals (vs. noise) or which cases might escalate (e.g. potential legal impact), allowing safety teams to prioritize urgent cases.
4. Labeling Compliance Checks — AI could compare drug labels, marketing materials, and prescribing information against regulatory requirements and recent evidence. It would flag non-compliant claims or needed updates, helping ensure labels are always up-to-date and accurate.
5. Regulatory Submissions & Compliance — AI could automate the preparation and review of regulatory submissions, reducing manual workload and ensuring compliance with evolving guidelines from regulatory agencies. NLP-based AI systems would analyze regulatory documents for consistency and completeness.
6. Regulatory Intelligence — AI could aggregate and analyze global regulatory announcements, guidelines, and precedents to inform companies of changes that could affect their products.

Manufacturing & Supply Chain Optimization for Life Sciences

1. Predictive Maintenance of Equipment — AI could predict machine failures on production lines by analyzing sensor data (vibration, temperature, etc.), scheduling maintenance before breakdowns occur, and reducing downtime.
2. Process Optimization & Yield Improvement — ML models could analyze process parameters (reactor settings, feed rates) to optimize yields and batch consistency.
3. Inventory & Demand Forecasting — AI forecasting models could predict demand for drugs and supplies across markets, helping pharma supply chain managers optimize inventory levels. This would help prevent stockouts or waste.
4. Personalized Medicine Manufacturing — AI could enable precision manufacturing of personalized therapies by optimizing batch production and formulation adjustments based on patient needs.
5. Production Scheduling — AI scheduling tools could allocate production lots to manufacturing lines and schedule workforce shifts efficiently. By accounting for complexities (change-over times, cleaning cycles), AI would generate optimal production schedules, increasing throughput.
6. Anomaly Detection in Manufacturing — Real-time AI could monitor sensor streams to detect out-of-spec conditions or unusual patterns that could indicate a problem. This early warning would allow interventions to prevent batch failures.
7. Vision-Based Quality Control — Computer vision systems could inspect drugs on the production line (pills, vials, packaging) for defects or deviations. AI could detect minute flaws or particulate contamination with higher accuracy than manual inspection.
8. Warehouse & Distribution Automation — AI could enhance logistics by optimizing how products are stored and shipped. Algorithms would determine the best warehouse layouts and pick-pack routes or even guide autonomous robots.
9. Lab Validation Update Plans — AI could automate data analysis, detect anomalies, and predict compliance risks for more efficient validation. It would also improve scheduling, documentation, and risk assessment, helping labs quickly adapt to new technologies and regulations while ensuring accuracy.

Commercialization for Life Sciences

1. Sales Forecasting — AI forecasting models could predict product demand by integrating prescriptions data, epidemiology, and market trends, allowing the business to gain more accurate forecasts for manufacturing and sales targets.
2. Personalized HCP Outreach & Marketing — AI could analyze prescribing patterns and patient demographics to identify target HCPs for optimized outreach. It could also personalize marketing materials based on an HCP’s specialty, prescribing behavior, and past engagement, improving relevance and adoption.
3. Sentiment Analysis — AI could analyze HCP, patient, and caregiver sentiment across social media, journals, and conferences to assess treatment perceptions. By identifying trends and concerns, AI would refine engagement, education, and adoption strategies.
4. Field Force Optimization — ML could suggest the optimal call plan for sales reps — which doctors to visit, how often, and which message to emphasize — to maximize uptake. This would leverage data on past rep interactions and outcomes, making sales teams more effective.
5. Speech Analytics for Sales Reps — AI could analyze sales rep conversations with HCPs to identify key themes, sentiment, and engagement effectiveness. By detecting trends in objections, prescribing behavior, and competition, it would provide insights to refine messaging and improve sales strategies.
6. Pricing and Market Access Modeling — ML models could simulate how pricing changes or reimbursement policies will impact market share and revenue. By analyzing payor data and competitive landscapes, AI would provide insights for optimal pricing strategies and negotiation points that align with current market dynamics.
7. MLR Review — AI could streamline MLR reviews by analyzing product claims, marketing materials, and promotional content to ensure compliance with regulatory standards. It would detect inconsistencies, flag potential risks, and automate documentation, enabling faster approvals while maintaining compliance with legal and regulatory guidelines.
8. Chatbots for Medical Information — AI chatbots could provide pharma companies with instant, accurate medical information while ensuring compliance and easing support workloads. For HCPs, they offer quick access to clinical data and guidelines. For patients, they can deliver medication guidance and safety information.
9. Call Center Optimization — AI can optimize patient services call centers by automating responses, analyzing interactions, and routing complex cases to specialists. Using natural language processing and sentiment analysis, AI can improve response times, personalize communication, and enhance customer satisfaction.
10. Patient Adherence Programs — AI can identify patients likely to abandon therapy (using pharmacy fill data, demographic, and even sentiment info) so support programs intervene. It could also tailor interventions (e.g. timing of reminders) to each patient.
11. Competitive Landscape Analysis — AI-driven analysis could help pharma companies track competitors, drug approvals, pricing, and market trends in real time. AI would identify threats and opportunities, enabling smarter strategy adjustments.
12. Post-Launch Performance Analytics — AI could track market trends, prescriptions, and patient engagement to identify sales drivers and barriers. It would also help refine marketing, optimize distribution, and improve drug performance.

AI Tools for Payors

For health insurers and payors, AI could be a powerful tool for managing claims, reducing fraud, and optimizing cost structures. By leveraging predictive analytics, AI could enhance risk assessment, utilization management, and provider network optimization. AI could also play a crucial role in regulatory compliance, member engagement, and population health strategies, helping payors improve efficiency while maintaining quality care. This section explores AI’s potential role in transforming how payors balance financial sustainability with patient-centered care.

Claims Management & Fraud Detection Tools for Payors

1. Automated Claims Processing — AI could streamline claims processing by identifying missing information, validating eligibility, and auto-adjudicating claims. This would reduce manual intervention, speed up processing, and minimize errors, ultimately leading to faster reimbursements and improved operational efficiency.
2. Fraud Detection & Prevention — AI could analyze claims data to detect anomalies, such as upcoding, unbundling, or duplicate claims. By leveraging advanced pattern recognition and predictive analytics, AI would help payors prevent fraudulent activities and reduce financial losses while ensuring compliance with industry regulations.
3. Predictive Claim Denial Prevention — Machine learning models could analyze patterns in historical claim denials to predict which claims might be rejected. Payors could then proactively address issues prior to submission.
4. Automated Prior Authorization — AI could streamline prior authorization by checking medical necessity against payor policies and historical data. This would reduce approval wait times, improve provider workflow efficiency, and enhance patient access to timely medical care.
5. OCR for Claims Documentation — Optical Character Recognition (OCR) could extract and validate data from paper claims and supporting documents. This would reduce manual data entry errors, accelerate claims adjudication, and enhance overall accuracy.
6. Provider Billing Compliance Audits — AI could scan provider billing data to identify inconsistencies and non-compliant coding practices. By ensuring regulatory adherence and reducing billing errors, AI would minimize claim rejections and improve provider reimbursement accuracy.
7. Medical Necessity Determination — AI could evaluate claims and medical records to determine whether a procedure meets medical necessity criteria. This would help reduce unnecessary denials, ensure appropriate patient care, and support evidence-based decision-making.
8. Duplicate Claim Detection — AI models could identify duplicate claims submitted by providers, preventing overpayments and redundant billing.
9. Anomaly Detection in Claim Payments — AI could scan claims history to detect unusual payment trends that may indicate fraud or billing errors.

Member Engagement & Experience Tools for Payors

1. AI Chatbots for Member Support — Virtual assistants could answer member inquiries about benefits, claims, coverage, and provider networks. This would improve response time and enhance member satisfaction.
2. Personalized Health Plan Recommendations — AI could analyze member demographics, health history, and preferences to suggest the most suitable health plan. This would ensure members choose plans that align with their healthcare needs.
3. Omnichannel Member Engagement — AI could seamlessly integrate communication across phone, chat, email, and portals, ensuring a unified and frictionless experience for members at every touchpoint.
4. Sentiment Analysis for Member Feedback — AI could process and analyze member feedback from calls, emails, and surveys to detect dissatisfaction, uncover emerging concerns, and provide insights for service improvements.
5. Voice AI for Call Center Optimization — AI-powered speech recognition could continuously analyze call interactions to assess sentiment, identify recurring issues, and highlight opportunities to enhance customer service and operational efficiency.
6. Member Eligibility Verification — AI-driven bots could instantly confirm a member’s coverage and eligibility for services, significantly reducing provider inquiries and expediting access to essential care.
7. Member Onboarding Automation — AI could automate the onboarding process by guiding new members through plan details, benefit utilization, and available digital tools, ensuring a smooth and informative experience.
8. Voice Authentication — AI could enhance security and streamline customer service interactions by verifying member identity through advanced voice recognition technology.
9. Personalized Member Communication — AI could customize messaging and outreach based on individual member behavior, preferences, and engagement patterns to foster stronger connections and improved satisfaction.
10. Self-Service Portals — AI-powered enhancements could improve self-service portals by delivering instant, accurate responses to members’ inquiries, making information access faster and more efficient.
11. Predictive Modeling for High-Risk Members — AI could leverage advanced algorithms to identify members at risk of hospitalizations or severe medical complications, enabling proactive and timely intervention.
12. Case Management Recommendations — AI could evaluate a member’s health data to identify the most effective interventions and case management programs tailored to their specific needs.
13. ED Overutilization Alerts — AI could continuously monitor emergency department usage patterns and flags members who frequently visit the ED for non-emergent conditions. Payors and care teams could intervene in real time to redirect these individuals to more appropriate and cost-effective care settings.
14. Specialist Referral Management — AI could ensure that members are referred to the most suitable specialists based on their medical history, specific health conditions, and network preferences.
15. Automated Appeals Processing — AI could assist members in tracking, submitting, and resolving claim appeals. This would speed up the resolution process, enhance transparency, and reduce frustration for members and providers alike.

3. Provider Network Management

1. Provider Performance Scoring — AI could evaluate provider performance using a combination of patient outcomes, cost efficiency, and compliance metrics. By analyzing vast datasets, AI would generate comprehensive performance insights that enable payors to make informed, data-driven decisions about network optimization.
2. Provider Directory Accuracy Management — AI could continuously monitor and validate provider directories, detecting inaccuracies in key details such as contact information, specialties, and availability.
3. Contract Negotiation — AI could leverage historical claims data and cost trends to provide data-driven insights that assist in provider contract negotiations. By identifying cost-saving opportunities and benchmarking reimbursement rates, AI would enable payors to secure more favorable contracts, improve financial sustainability, and maintain high-quality care within their networks.
4. Value-Based Care Model Optimization — AI could play a crucial role in designing, managing, and optimizing value-based care contracts by tracking and analyzing key quality metrics.
5. Predictive Provider Network Adequacy — AI could assess geographic and specialty gaps within provider networks, ensuring members have sufficient access to essential healthcare services.
6. Provider Dispute Resolution — AI could streamline provider dispute resolution by automatically analyzing contract terms, claims data, and reimbursement discrepancies.
7. Fraudulent Prescription Detection — AI could utilize advanced pattern recognition and anomaly detection to identify suspicious prescribing behaviors that may indicate fraud, abuse, or non-compliant prescribing practices.
8. Provider Credentialing & Licensing Verification — AI could streamline provider onboarding by continuously cross-referencing credentials with licensing databases and regulatory bodies, ensuring up-to-date compliance.

Cost & Risk Management Tools for Payors

1. Risk Adjustment Analytics — AI could systematically analyze claims and clinical data to detect under-documented or miscoded conditions that impact risk scores. By identifying these gaps, AI would ensure that risk-adjusted payments accurately reflect the true health status of members.
2. Predictive Cost Modeling for Claims Forecasting — AI could leverage historical utilization patterns, demographic data, and emerging healthcare trends to predict future claims costs.
3. Medical Loss Ratio (MLR) Optimization — AI could monitor claims spending, utilization patterns, and cost-efficiency metrics to identify areas where payors can improve their Medical Loss Ratio (MLR).
4. Predictive Modeling for Stop-Loss Insurance — AI could identify high-cost claimants by analyzing medical history, claims data, and predictive health indicators. This would enable insurers to assess risk exposure more accurately and determine the most effective stop-loss coverage levels.
5. Drug Formulary Cost Optimization — AI could evaluate prescription drug utilization trends, formulary adherence, and cost variations to recommend the most cost-effective alternatives within formulary guidelines.
6. Actuarial Analysis — AI could enhance actuarial modeling by processing large-scale claims data, risk factors, and premium pricing variables at an accelerated rate.
7. Self-Insured Employer Cost Containment Strategies — AI could assess healthcare utilization patterns among self-insured employer groups to recommend targeted cost-containment strategies.
8. Alternative Payment Model (APM) Optimization — AI could analyze bundled payment models, shared savings programs, and value-based care arrangements to help payors and providers refine their alternative payment strategies.

Regulatory Compliance & Reporting for Payors

1. Automated Regulatory Compliance Audits — AI could continuously scan and monitor claims, billing processes, and operational workflows to ensure adherence to CMS, HIPAA, and other federal and state regulations.
2. Real-Time CMS Star Ratings Optimization — AI could track and analyze key performance metrics that directly impact CMS Star Ratings, such as patient satisfaction scores, medication adherence, and preventive care measures. AI would recommend targeted interventions that allow payors and providers to proactively improve their ratings.
3. Automated 1095-B & 1095-C Reporting — AI could automate the generation, validation, and submission of tax forms required for Affordable Care Act compliance, significantly reducing administrative burden for employers and payors, minimizing the risk of IRS penalties, and eliminating errors associated with manual data entry.
4. ICD-10 & CPT Coding Accuracy Verification — AI could audit medical coding practices by cross-referencing ICD-10 and CPT codes against payor policies, medical necessity guidelines, and regulatory frameworks.
5. Predictive Modeling for OIG Audit Risks — AI could analyze extensive claims data to detect anomalies, inconsistencies, and patterns that may increase the likelihood of an Office of Inspector General (OIG) audit.
6. Automated Compliance Documentation Management — AI could streamline regulatory compliance workflows by categorizing, indexing, and managing required compliance documents. It would ensure that compliance teams have real-time access to the latest regulatory updates, deadlines, and required documentation.
7. State Medicaid Reporting Automation — AI could automate the generation and submission of Medicaid reports based on state-specific requirements, ensuring payors remain compliant with complex regulatory mandates.

Population Health & Predictive Analytics Tools for Payors

1. Health Equity & Disparities Analysis — AI could uncover disparities in healthcare access and outcomes across different demographics. Payors could then use this data to optimize and tailor patient outreach programs.
2. Vaccine Uptake Optimization — AI could identify members unlikely to receive recommended vaccinations and deliver targeted engagement campaigns, boosting immunization rates.
3. Personalized Preventive Care Recommendations — AI could evaluate members’ health history, demographics, and engagement patterns to recommend personalized preventive care measures. Payors could then design member-centric engagement programs to boost adherence to preventive screenings, etc.

****

Acknowledgements:

Special thanks to Kristen Chapey for drafting this piece, to Gaurav Singal, Engy Ziedan, Chris Taylor, Annie Powers, Xerxes Sanii, Elyse Benedicto, Kyle Bryant, and Bobby Samuels for invaluable feedback, and to the many invaluable resources we consulted (select highlighted below) to pull together this work.

• ChatGPT & Perplexity
• https://www.mckinsey.com/industries/healthcare/our-insights/generative-ai-in-healthcare-adoption-trends-and-whats-next
• https://pmc.ncbi.nlm.nih.gov/articles/PMC8285156/
• https://www2.deloitte.com/us/en/pages/consulting/articles/ai-dossier-life-sciences-healthcare.html
• https://aicheckup.substack.com/p/where-generative-ai-meets-healthcare
• https://builtin.com/artificial-intelligence/artificial-intelligence-healthcare
• https://www.alation.com/blog/ai-healthcare-top-use-cases/
• https://cloud.google.com/use-cases/ai-in-healthcare
• https://empeek.com/insights/top-ai-applications-in-healthcare/
• https://health.clevelandclinic.org/ai-in-healthcare

However, despite the boom in the category — I am generally skeptical of most companies in the space, and there is a graveyard of failed privacy tech startups (with a few notable exceptions).

Over time, I’ve formed three hypotheses on the space.

Hypothesis 1: This is a golden age for privacy technology.

There are technical, regulatory, and commercial tailwinds for the space. From a technical perspective, there has been massive progress in the underlying tech for privacy preservation over the last decade, and lots of new approaches have emerged out of academia and into commercial feasibility. From a regulatory perspective, regulations like CCPA, GDPR, HIPAA, and a variety of other privacy laws have created urgency for companies to be thoughtful about privacy preserving tech. From a commercial perspective, use of data (and complex combinations of data sets) is skyrocketing.

Hypothesis 2: To be successful, privacy tech companies need to be specific about what data flows they’re unlocking.

Too many privacy tech companies are focused on tech (or privacy) for its own sake. While there are reasonable academic debates about when federated data exchange is better than a centralized clean room approach, very few customers really care — what they care about is a specific business problem for a data flow they need to unlock.

For example, Datavant was built around two very specific requirements: i) under HIPAA, companies generally have a legal obligation to de-identify health data they use for analytics, and ii) these companies want to be able to combine multiple deidentified data sets for their analyses; we made it easy to check the box on those two business requirements. No customers deeply cared about the cryptographic details of how it worked, as long as it worked and fulfilled those requirements. And because it worked, it unlocked data flows (and business outcomes) that previously weren’t possible — and we found great product-market fit as a result.

Too many privacy tech companies are generic “we make data exchange that’s already happening more privacy-safe” — but the reality is, most clients don’t care about that unless it unlocks a new data flow or business objective on their end.

Hypothesis 3: Privacy tech businesses should follow a network business model.

The other mistake privacy tech companies often make is following a pure tech playbook. “Our technology is great for data sharing in healthcare, financial services, public sector, or marketing” is a typical mantra for privacy tech businesses, and they sell to clients indiscriminately across verticals.

This is a trap — there isn’t enough defensibility or stickiness in the pure tech play as there is a features arms race between lots of privacy tech businesses.

However — the most important data flows involve multiple parties sharing data for some purpose. For example, healthcare organizations sharing de-identified data across institutions for research, banks sharing data across companies to determine creditworthiness, publishers and brands exchanging information to target ads, and e-commerce companies sharing information about fraud risk. In each of these businesses, there are very real needs for privacy-preserving technology for the parties to exchange data safely — but also a need for standardization on a particular means of sharing data across the industry; the technology is much more valuable if counterparties are already using it. That’s where the network effect lies: adoption by a critical mass of the space on a particular privacy tech makes it the de facto standard.

Viewed as a network business, privacy tech companies should be obsessed with network business dynamics — getting critical mass in specific domains, getting key nodes into the network via a land-and-expand approach, and narrowing focus as much as possible. Instead of focusing on multiple verticals, they should generally focus as narrowly as possible.

***

In summary, there are massive tailwinds for the privacy tech space as a whole — but the rewards will only go to the companies that both unlock a specific data flow and focus on building out a network. Companies that don’t do this (most privacy tech businesses), unfortunately aren’t generally great businesses.

Different companies are stronger or weaker on these two dimensions. Incumbents in a market generally have strong distribution. Meanwhile, in most industries, the nimbleness and talent profile of startups enables them to build strong products more rapidly to adapt to technology changes. But a better product alone doesn’t win, as many wide-eyed entrepreneurs have learned.

During past technological paradigm shifts, speed of building a strong product was more important in determining the winners — and startups benefited. During the dot-com boom and the mobile wave, the biggest returns went to the most successful startups rather than major incumbents, and the VC investment class had the greatest success.

This is because there were natural network effects built into most of the products that were created in those waves. During mobile, the biggest winners were companies like Uber and Instagram — which moved quickly and nimbly to create great products, well before any incumbents could enter the market. By the time the incumbents could react, the market was already locked up because of the extreme network effects.

AI businesses don’t have the same inherent network effects. Unlike in past technological paradigm shifts, the companies spawned by this shift don’t have obvious network effects to propel them if they are the first-movers. No vertical AI CEOs have a good answer at this point for why their business is defensible; it’s usually a version of “it’s a hard problem” rather than the obvious network effects of an Instagram or Uber.

So that means the best products will be commoditized. The lack of network effects means that, instead of there being one runaway category-winner, it’s likely that every category has dozens of companies with comparable products. Uber and Instagram were winner-take-all, there’s no reason that “vertical AI to automate X workflow” won’t have 20 equivalent companies fighting over the same share.

And it means that incumbents have an advantage. The lack of network effects also means that, even if a nimble startup builds an incredible product two years ahead of the incumbent, the incumbent will still be able to use its distribution advantage to win the market after it eventually catches up. Two years was an eternity because of Instagram’s network effects (there was no way Facebook could have built a competitor with that deficit, or Google Photos could have prevailed). Without network effects, a two-year headstart simply doesn’t matter that much. If a healthtech company builds “AI for XYZ,” and Epic builds it two years later at 80% the quality, I would bet on Epic winning.

The economics of AI favor incumbents: if real innovation is being produced, but the product itself is commoditized — the value of distribution will increase. AI will create massive winners: real innovation is going to come to market, and real value will be created — so buyers are eager to step forward for a variety of new capabilities. But the beneficiaries will be the incumbent with strong distribution; they can either fast-follow on product development, or (more often) acquire from amongst the 20 slight variations of startups in a given category, making the acquisition relatively inexpensive.

***

As a practical example of all this, the medical coding space for providers in healthcare (i.e. converting medical records into billing codes) is an example of a sector where i) AI is going to be transformational on the industry, ii) there are a flurry of innovative companies going after the problem, and iii) I would bet on an incumbent with good distribution ultimately winning (either by building its own product or doing a tuck-in acquisition).

I believe the AI boom will create at least as much value as the dot-com boom and mobile wave, but I think it is less clear that the traditional VC asset class will do well or that as many incredible startup successes will be minted in vertical AI plays [though I am still extremely bullish on startups that are infrastructure for AI, or that are complimentary to AI]. Instead, for vertical success, I would bet on incumbent companies that figure out how to hire well and acquire well, PE strategies of combining AI companies with incumbents, and tech-enabled rollup plays.

This post is a playbook for how to design your org at various stages of maturity, derived from lessons I learned scaling through these phases at LiveRamp (which I saw from 0 to 800 employees) and Datavant (which I saw from 0 to 8,000 employees). It’s most applicable to high-growth, enterprise B2B product companies (though many of the concepts apply to other types of businesses as well).

This guide is organized as five sections:

1. Principles throughout the journey
2. 0–10 People: The Anarchy Phase
3. 10–100 People: The Leader-Centric, Structured Chaos Phase
4. 50–250 People: The Functional Stage
5. 250+ People: Functional vs. Divisional vs. Matrix Choices

Principles Throughout The Journey:

There is no perfect org structure. Org design fundamentally is about flow of information and decision-making. In the perfect organization, all information flows to everyone in the company, and no time is spent on communication overhead. Once an organization is >1 person, this breaks down, and the complexity of communication scales exponentially as the company grows. So you design the organization to make sure the right information gets to the right people, the right decisions move quickly, and the right amount of overhead is spent on internal communication. This means that there are tradeoffs — what information needs to move quickly, what decisions need speed vs. perfection, etc. The key is to be intentional with your tradeoffs and design the organization that makes sense for your strategy.

Build your bench. Throughout the journey, the biggest bottleneck is usually having the right person who can step into a key role — you need the right overlap of talent, cultural fit, ability to lead, ability to communicate with adjacent functions, etc. Building your internal bench (by hiring well, promoting well, and growing people internally) gives you many, many more degrees of freedom as you grow.

Treat org design as a living experiment. As they grow, many companies have a major reorg that they plan for months, and expect stability after that. That rarely works: a high-growth company changes frequently (and decisions around org design are often wrong). So instead, building a company culture that expects frequent micro-optimizations — and views the organization itself as a form of experimentation — is ideal.

Build operating cadences to compliment your org design. Org design is important as the company grows, but it’s ultimately just one lever on how information flows within the company. Be intentional with how you build operating cadences (team meetings, management meetings, internal aliases, etc.) to manage the information flow as well.

Different people thrive at different stages. There are some people who are exceptional at the early stages, but who strongly dislike the structure that comes with later stages. And there are people who will be great at later stages who can’t handle pre-product-market-fit ambiguity. A very small number of people will thrive throughout — but you’ll want to reassess at each stage who the right people are (and who wants to be involved for the next chapter of the journey).

Try to stay lean. The smaller you are, the more you can accomplish per capita — and the less effort needs to go into organizational design. I generally think a high-performing 100-person organization can accomplish 2–3x what a 10 person organization can, and a high-performing 1,000-person organization can accomplish 2–3x what a 100 person organization can. Hiring well, biasing towards automation, having aggressive goals for each person, and strong performance management go a long way in ensuring you don’t scale more than you have to.

0–10 People: The Anarchy Phase

At 0–10 people, org design is not (and should not be) a topic you spend much thought around — but there are a few key foundations you want to get right.

A few thoughts for this stage.

Keep the org flat. In general, at this size, everyone either reports to the CEO, or the CEO + one other leader (for example, an engineering leader). Organizationally, the most important thing at this size is that everyone has a clear charter, is getting feedback, and knows who to communicate with for key decisions. Any more organization design is overkill at this stage.

Set the right foundations for the future:

• Hire really well, and set a really good culture — these are the foundations for all future scaling
• Be intentional with early titles. I prefer not having hierarchical titles (eg. “Sr. Director of Marketing”) within an organization at any size because of the politics it causes, but it’s only possible to have this policy down the line if you set the precedent of not giving out big titles.

10–50 People: The Leader-Centric, Structured-Chaos Phase

During this stage, you’re still not spending a huge portion of your time thinking about the org design — but you’re beginning to have some foundations of org structure.

A few thoughts for this stage:

Build around people rather than theoretical structure. At this size, there are usually 2–4 superstars in the company who are involved in every major decision, and decision-making de facto involves pulling them in (and they generally know how to stay in sync with each other). So if they’re good leaders, organize the company around them. At LiveRamp, we had Product, Recruiting, and Marketing all report to the same person around this size — not because there’s some logical grouping of those functions, but because she was a superstar who could manage lots of things well. Embrace this for now, and build around the strengths of your team.

Be mindful of team sizes and layers. In general, a good leader can handle ~6–8 direct reports — that allows them to give enough attention to each one, know what’s going on on their team, and provide useful feedback. During this phase, that often means you need to introduce a second layer in your engineering team. You’ll want to be thoughtful about who your best leaders are (often not the same as your best coders), and also begin being thoughtful about building out your leadership bench by creating growth paths for folks on the team.

Focus on getting a few senior leaders. By the time you hit 50 people, you’ll probably want to have 2–3 core leadership roles filled out (over time, you’ll need to fill out heads of sales, engineering, product, people, finance as the most senior roles). These are often external hires, and are extremely important to get right culturally (and obviously important to get a great performer). A lot of your time in this phase should be spent on recruiting these key seats.

Begin designating a management team. While you want to avoid “The Management” being separate from everyone else, it’s helpful to have a small group of leaders meeting regularly who can surface topics across functions — so you will need to begin to designate a management team (usually, at this stage, it’s your direct reports, but not necessarily). Begin introducing this, but be mindful of minimizing the politics it can introduce.

Think about how decisions get made that span multiple functions. At LiveRamp and Datavant, I was explicit that Product could break a tie between functions — which made decision-making fast (and reduced decisions that required my input), and set a cultural tone for the company. That’s not necessarily the right answer for all companies — but there should be an explicit way ties get broken and decisions get made at this size.

Pay attention to how you organize around a customer. As you start to have a team of Customer Success, a team of Product Implementation, and a team of Salespeople, all of which touch a customer — be particularly thoughtful about how handoffs happen between these functions and how the account is owned.

50–250 People: The Functional Stage

Some thoughts on this phase….

Build out the leadership team. In this phase, you are going to fill out all your core leadership functions, and also promote+hire many mid-level leaders (at 250 people, you might have ~50 people in the company managing others — so lots of mid-level manager roles to fill). This is your most important job at this stage: bring in people who can help you scale effectively, and you’ll spend much of your time recruiting and training people internally.

Structure around key functions. Unlike at earlier stages, org structure now matters, and in general, you’ll want to organize things around the major functions (ie. sales, product, engineering, etc.). Key decision points include

• Customer Success — should it report to Sales, or be standalone? And does it own upsell or just renewal? I’ve historically had customer success report into sales, and had CSMs own upsell numbers (except on the biggest accounts) — but this varies from org to org.
• Delivery/Product Success — Many orgs put this in sales/customer success. I’ve historically put it in Product, to keep the Product team internalizing customer challenges (I deep dive on my thoughts on this function here).
• What reports to you? At this stage, you can start to make decisions about what reports to you — both based on your strengths and what’s needed for the org. Do you want marketing as a top-level function or reporting to sales? Do you want privacy as a top-level function? Do you want a COO or CFO to manage all the back-office functions?

Consider sub-team structures within sales and engineering. Once you cross 100 people, both sales & engineering likely have multiple levels of management. You’ll want to think through how you organize subteams. For Sales, there are generally territories — which might be geographic, or might be type of customer (for example, Enterprise vs. SMB). For engineering, you’ll want to have some teams that are organized by product, and some that are system-wide in nature (eg. DevOps team, Infra team, etc.). You’ll likely play around quite a bit with these definitions as you cross ~100 people or so.

Consider whether Product is organized by segment of market or by engineering team. As you start getting a team of product managers, there is an advantage to keeping PMs directly mapped with engineering teams, and there is also an advantage to having PMs deeply embedded in a given market. At Datavant, we ultimately split the team into “Platform PMs” (embedded with engineering teams) and “Solution PMs” (embedded with sales teams and deep in specific markets.

Be thoughtful on how you show up to a customer. As you scale, there will be a moment when you have several PMs representing several products, a customer success team, a sales rep, a CSM, and someone accountable for a new product all trying to talk to the same customer. Making sure you have internal clarity on who quarterbacks the customer’s holistic experience is extremely important.

Get a great head of HR. Starting at this stage, you’ll be spending as much time thinking about the org (recruiting, retention, and org structure) as you spend thinking about any other function. Getting a strategic head of HR who can be a thought partner at this stage is invaluable.

Get yourself leverage. At this stage, there are thousands of pulls on your time, and it’s critical to get yourself leverage. The most important way to get leverage is to get strong leaders in each function. At this stage, I also added an EA and a Special Projects team (people who I could work with on things outside the standard org structure).

Control headcount growth by function. Every leader believes they can accomplish more with more resources. At this stage, you generally don’t have intense budgeting and controllership processes, but teams can rapidly scale without you really paying attention. Make sure you have a strong point-of-view on what teams are bloated and control where headcount goes within the company. Similarly, by this stage, you’ll also have leaders who are great at hiring and leaders who are bad at hiring; you’ll want to help with quality control on the functions that don’t hire as well.

Build the right operating cadences. This is a moment of growth where a lot of process needs to be created to scale effectively and communicate well internally. Be intentional with what operating cadences you create to complement your org structure).

250+ People: Functional vs. Divisional vs. Matrix Choices

At this scale, there are three main approaches to organizing the company:

1. A functional org structure organizes around the major functions in the company — i.e. there is a single product organization, a single sales organization, and a single engineering organization. The pros of this model are that it helps ensure a cohesive product experience across parts of the business and decisions that take into account holistic information, but the con is that this model tends to be much slower to adapt to specific sub-markets and give less entrepreneurial autonomy.
2. A divisional org structure organizes around major types of customers. The leaders in this model are general managers, and they generally have their own sales teams, their own delivery teams, their own product teams, and often their own engineering teams. For example, Datavant is organized with a GM of Life Sciences, a GM of Providers, and a GM of Payers, because the markets are quite different; similarly, ride sharing apps organized around GMs of local markets, and at a much larger scale, there’s a GM of the AWS business within Amazon who is accountable for AWS’s success. The advantage of this model is that it’s nimble to serve specific types of customers and allows teams to be entrepreneurial in servicing the market; cons are that decisions can be disconnected across different business units, product experience can become disconnected, and employee experience can become disconnected (an engineer in one part of the org might be less good than engineers in another part of the org and even use a different tech stack).. Note that most divisional org structures will have a set of “shared service” functions that do not split divisionally (finance, HR, and corporate IT are usually shared services; engineering sometimes is a central function as well).
3. Matrix org structures. To mitigate the downsides of each of these structures, the vast majority of companies end up is creating matrices: you have GMs of each line of business, and ALSO functional leaders. There are many nuances to making this successful, but in general, employees report to the functional leader (who makes hiring/firing decisions and sets compensation consistently across the team, and decides on allocation across divisions), and their priorities are managed by GMs. The upside on this is that you can have both intense market focus AND consistency across the organization, the downside is that — unless you have great alignment and decision-making capabilities — progress can grind to a halt as there are many cooks in the kitchen.

The reality is, most companies end up with some hybrid: they might run the core business as a matrix and a new product as a separate division, or they might alternate between more divisional and more functional. The keys at this stage are:

Be intentional on the functional vs. divisional vs. matrix decision. Even if you change the structure later, have a clear point of view. Decide if you are prioritizing nimbleness or consistency, and what decisions require each.

Decide when to change. It’s fine to switch between models as you grow. And even if you’re consistent with one model, your business units will change over time. When HP was a hot high-growth company, they had a mantra of splitting a business unit into two automatically whenever it exceeded a certain size threshold. So defining the optimal territories and business units will constantly evolve as the business grows.

Decide what is an exception to the model. You might decide to acquire a company, and keep them in a separate org structure. Or make your Japan business completely autonomous from the rest of the company. Or have a new product that moves at a different pace. This is fine, and do it intentionally, and explain it internally.

Figure out where your zero-to-one people go. Many of the people who thrived in the early stages will not thrive in the stages where there are more decision-makers and alignment matters more. Decide if there’s still a home for them in the company, and be intentional about creating more entrepreneurial parts of the company that can be home for zero-to-one teams.

Decide where consistency matters. Do you care strongly about consistency of compensation across teams? Talent caliber? UI? Pricing? Be intentional in designing the org structure and processes to match your thoughts on where consistency matters, and where you want to allow local decision-making.

***

This playbook was intended to help you think through your org structure as you scale. But remember: there is no perfect org design, and your org structure should evolve as you grow, and you should constantly experiment with it.