Optimizing Workflow Processes

Context graphs are quickly becoming one of the most talked about ideas in enterprise AI, with investors Jaya Gupta and Ashu Garg calling them a “trillion-dollar opportunity.” The reason? If we want agents to do real work, they don’t just need better models. They need a reliable model of how work gets done inside an organization. A context graph connects your people, content, and systems to the time‑ordered traces of actions between them, so agents can see real processes, not just static data. At Glean, we’ve found that building a useful context graph comes down to three core elements: • Observe real work, not just final states. Capture fine‑grained activity across the tools where work happens (edits, comments, messages, status changes, meetings) instead of relying only on the current record in a single system. • Put structure on top of that activity. Use knowledge graphs (projects, customers, products, teams) and personal graphs (what each person is working on) to turn noisy events into coherent tasks and end‑to‑end processes. • Continuously learn from humans and agents. Treat every successful resolution (whether done by a person or an agent) as another trace in the graph, so the system’s “playbooks” improve over time. Before we shipped this, we tested it on ourselves. With employees’ opt‑in, we analyzed real work sequences for flows such as AE mid‑market deal cycles, SE proofs‑of‑concept, on‑call incident response, and PM feature launches to understand what effective paths actually look like in practice. Our team has published a deep dive on how we build these context graphs at Glean, from deep connectors and knowledge graphs to agentic feedback loops. https://lnkd.in/gBXRmvM5

How can businesses go beyond using AI for incremental efficiency gains to create transformative impact? I write from the World Economic Forum (WEF) in Davos, Switzerland, where I’ve been speaking with many CEOs about how to use AI for growth. A recurring theme is that running many experimental, bottom-up AI projects — letting a thousand flowers bloom — has failed to lead to significant payoffs. Instead, bigger gains require workflow redesign: taking a broader, perhaps top-down view of the multiple steps in a process and changing how they work together from end to end. Consider a bank issuing loans. The workflow consists of several discrete stages: Marketing -> Application -> Preliminary Approval -> Final Review -> Execution Suppose each step used to be manual. Preliminary Approval used to require an hour-long human review, but a new agentic system can do this automatically in 10 minutes. Swapping human review for AI review — but keeping everything else the same — gives a minor efficiency gain but isn’t transformative. Here’s what would be transformative: Instead of applicants waiting a week for a human to review their application, they can get a decision in 10 minutes. When that happens, the loan becomes a more compelling product, and that better customer experience allows lenders to attract more applications and ultimately issue more loans. However, making this change requires taking a broader business or product perspective, not just a technology perspective. Further, it changes the workflow of loan processing. Switching to offering a “10-minute loan” product would require changing how it is marketed. Applications would need to be digitized and routed more efficiently, and final review and execution would need to be redesigned to handle a larger volume. Even though AI is applied only to one step, Preliminary Approval, we end up implementing not just a point solution but a broader workflow redesign that transforms the product offering. At AI Aspire (an advisory firm I co-lead), here’s what we see: Bottom-up innovation matters because the people closest to problems often see solutions first. But scaling such ideas to create transformative impact often requires seeing how AI can transform entire workflows end to end, not just individual steps, and this is where top-down strategic direction and innovation can help. This year's WEF meeting, as in previous years, has been an energizing event. Among technologists, frequent topics of discussion include Agentic AI (when I coined this term, I was not expecting to see it plastered on billboards and buildings!), Sovereign AI (how nations can control their own access to AI), Talent (the challenging job market for recent graduates, and how to upskill nations), and data-center infrastructure (how to address bottlenecks in energy, talent, GPU chips, and memory). I will address some of these topics in future posts. [Original text: https://lnkd.in/gbiRs2mi ]

“Stop calling everything an agent” Folks are slapping 'agent' on anything that touches AI. It’s confusing. I jumped on Peter Yang’s podcast to unpack why that’s misguided, and why the middle ground (agentic workflows) is where real ROI lives. Workflows = determinism and reliability Agents = judgment and flexibility Stitch them together and you get the safest, most powerful form of AI orchestration. Give your agent just enough tools and context to do one job exceptionally well, and orchestrate the rest with workflows. Two examples of this I’m using personally: 𝗘𝘅𝗲𝗰𝗖𝗼𝗻𝗻𝗲𝗰𝘁 (Workflow / Determinism)  👉 A teammate submits an Interfaces form to request exec engagement, the AI drafts an on-brand message, posts to Slack, and tracks it in Tables. This one runs hundreds of times a month. Zero chaos, total clarity 𝗜𝗻𝗯𝗼𝘅 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 (Agent / Inference) 👉 When a new email arrives, the agent reads it, checks HubSpot, uses reasoning to categorize it, and sorts it in Slack. It flags customers, intelligently routes the right things to my EA Cortney, and clears all of the noise. It reasons like a human assistant, saving us hours each week with full transparency logs Cortney’s Inbox Management agent is actually available as a template, you can steal it here: https://lnkd.in/gR8UjTCN Full episode coming soon. Thanks for having me, Peter.

Working with multiple LLM providers, prompt engineering, and complex data flows requires thoughtful organization. A proper structure helps teams: - Maintain clean separation between configuration and code - Implement consistent error handling and rate limiting - Enable rapid experimentation while preserving reproducibility - Facilitate collaboration across ML engineers and developers The modular approach shown here separates model clients, prompt engineering, utils, and handlers while maintaining a coherent flow. This organization has saved many people countless hours in debugging and onboarding. Key Components That Drive Success Beyond folders, the real innovation lies in how components interact: - Centralized configuration through YAML - Dedicated prompt engineering module with templating and few-shot capabilities - Properly sandboxed model clients with standardized interfaces - Comprehensive caching, logging, and rate limiting Whether you're building RAG applications, fine-tuning foundation models, or creating agent-based systems, this structure provides a solid foundation to build upon. What project structure approaches have you found effective for your generative AI projects? I'd love to hear your experiences.

"Why cognitive load (not clean code) is what really matters in coding" What truly matters in software development isn't following trendy practices - it's minimizing mental effort for other developers. I've witnessed numerous projects where brilliant developers created sophisticated architectures using cutting-edge patterns and microservices. Yet when new team members attempted modifications, they struggled for weeks just to grasp how components interconnected. This cognitive burden drastically reduced productivity and increased defects. Ironically, many of these complexity-inducing patterns were implemented pursuing "clean code." The essential goal should be reducing unnecessary mental strain. This might mean: - Fewer, deeper modules instead of many shallow ones - Keeping related logic together rather than fragmenting it - Choosing straightforward solutions over clever ones The best code isn't the most elegant - it's what future developers (including yourself) can quickly comprehend. When making architectural decisions or reviewing code, ask: "How much mental effort will others need to understand this?" Focus on minimizing cognitive load to create truly maintainable systems, not just theoretically clean ones. Remember, code is read far more often than written. #programming #softwareengineering #tech

𝐓𝐡𝐞 𝐛𝐞𝐬𝐭 𝐝𝐞𝐜𝐢𝐬𝐢𝐨𝐧𝐬 𝐚𝐫𝐞 𝐦𝐚𝐝𝐞 𝐰𝐢𝐭𝐡 𝐜𝐥𝐚𝐫𝐢𝐭𝐲, 𝐧𝐨𝐭 𝐠𝐮𝐞𝐬𝐬𝐰𝐨𝐫𝐤. Digital twins take the guesswork out of decision-making by creating a virtual model of your operations that reflects reality in stunning detail. From improving design to reducing downtime, they transform the unknown into actionable intelligence. To simplify the broad range of potential digital twin applications, a classification approach I like to use is called the “𝟓 𝐏𝐬“. This model is easy to remember and covers nearly all use cases of industrial digital twins: • 𝐏𝐚𝐫𝐭 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐓𝐰𝐢𝐧: Digital representation of individual components or parts typically to understand the physical, mechanical, and electrical characteristics of the part. This allows companies to monitor, analyze, and predict the performance and health of that particular part, optimizing maintenance schedules and extending its lifecycle. • 𝐏𝐫𝐨𝐝𝐮𝐜𝐭 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐓𝐰𝐢𝐧: Digital representation of the interoperability of components or parts as they work together as part of a product. This enables companies to simulate and test product behavior under various conditions, improving design, ensuring quality, and speeding up the time to market. • 𝐏𝐥𝐚𝐧𝐭 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐓𝐰𝐢𝐧: Digital representation of a plant, facility, or system to understand how assets work together at an operational level. This allows businesses to enhance operational efficiency, reduce downtimes, and optimize production processes through real-time insights and predictive analytics. • 𝐏𝐫𝐨𝐜𝐞𝐬𝐬 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐓𝐰𝐢𝐧: Digital representation of a specific process or workflow within a system or a facility. This helps companies refine and optimize processes, identify inefficiencies, and ensure smoother and more cost-effective operations. • 𝐏𝐞𝐫𝐬𝐨𝐧 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐓𝐰𝐢𝐧: Digital representation of a person to capture their movements, habits, interactions, skills, knowledge, and preferences. This helps companies gain insights into workflow patterns, fatigue patterns, and safety concerns ensuring increased productivity and a reduction in workplace-related injuries. 𝐇𝐨𝐰 𝐝𝐨𝐞𝐬 𝐚 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐓𝐡𝐫𝐞𝐚𝐝 𝐫𝐞𝐥𝐚𝐭𝐞 𝐭𝐨 𝐚 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐓𝐰𝐢𝐧? A digital thread is a continuous flow of data and information that integrates processes, systems, and devices throughout the product lifecycle. It serves as the foundation for a digital twin, which is a virtual representation of a physical product or system, leveraging data from the digital thread to simulate, predict, and optimize its performance. For high-resolution image and to read full version: https://lnkd.in/ezmPkSag ******************************************* • Visit www.jeffwinterinsights.com for access to all my content and to stay current on Industry 4.0 and other cool tech trends • Ring the 🔔 for notifications!

🔮 How To Prioritize UX Work (Framework) (https://lnkd.in/eGQrPm2N), a very practical guide on how to choose and estimate the right level of research and UX work needed for a successful outcome of a project — along with the process to follow and UX estimates to set. Kindly shared by Jeremy Bird. 🤔 Planning is typically done for the delivery phase only. 🤔 Design, research, discovery, ideation are not planned. 🤔 Effort, estimates, roadmaps, capacity are rare for UX work. 🚫 Not every project needs the same level of research/design. ✅ Goal: set realistic expectations for UX work in a timeframe. Jeremy suggests to estimate research and design efforts separately, and across different dimensions: we assess research by mapping Risks and Problem Clarity. And we estimate design effort needed by mapping Risk and Level of Complexity: 🔮 Clarity: Low ↔ High New, unknown problems usually come with a lot of assumptions and very low clarity. Well-known problems with shared understanding in the team and some extensive research have higher degree of clarity. 🔥 Risk: Low ↔ High Some projects are relatively easy to roll back and they don't really affect business-critical workflows (low risk). Others are much more difficult to reverse and operate within users' key journeys (high risk). 🚀 Complexity: Low ↔ High Self-contained projects in well-understood workflows are typically straightforward (low complexity). Some projects that involve many systems, external dependencies, stakeholders scattered across teams with little existing knowledge (high complexity). ✅ We start by defining a problem to solve + business impact. ✅ Then, we shape desired user outcome and success criteria. ✅ Next, we assess design effort and research effort levels. ✅ Run a kickoff meeting to prioritize and decide the scope. ✅ Designers break down UX work, estimate it, add to Jira. Personally, I always find it remarkably difficult to estimate the effort for research or design work. Even after so many years, with 20–30% buffer, I’m often underestimating the little nuances, blockers, constraints and bottlenecks hidden away somewhere between complex dependencies and external stakeholders. One thing is certain though: considering risk early is a very, very effective way to guide UX work in the right direction. High risk always requires some level of research and discovery. And early prioritization helps UX teams focus their effort where they add most value — saving time on resources for projects that deliver value to users and businesses. Finally: I can highly recommend to consider John Cutler's Effort vs. Value curves (https://lnkd.in/evrKJUEy) for prioritization work as well. Much of the work isn’t completed once it's delivered. More often than not, it will significantly add to maintenance costs over time. We better account for it early. #ux #design

On Monday, I fired 2 agents! So 3 weeks ago, I built 4 agents and I did everything "right”: - I created detailed .md files - I gave them context - I wrote thorough BRDs - I tested them repeatedly - I refined prompts - I ran multiple rounds of validation before pushing them into production The on Monday, I fired 2 of the 4 agents. A 50% failure rate 😒and that’s not unusual. Most people just don't talk about it because it doesn't sound as impressive as posting screenshots of agents that supposedly run entire businesses. The two agents I fired weren't broken. They were just extremely unreliable. 🔥 One couldn't consistently find calendar events that absolutely existed. 🔥 Another failed to rerun a process if the machine was powered on but the scheduled task wasn't actively running. Neither failure was catastrophic because I chose use cases I could recover from and manually intervene, BUT, they did destroy trust. Here are some of the most common reasons I've seen agents fail: 💎Missing calendar entries that clearly exist 💎Failing when a webpage layout changes slightly 💎Forgetting to rerun a workflow after an interruption 💎Stopping when a login session expires 💎Misinterpreting exceptions as completed tasks 💎Losing context between steps in a process 💎Getting trapped in loops and repeating the same action 💎Declaring success without actually validating the outcome. The reality is that agents don't fail because they're "dumb” I think they fail because enterprise work is messy and humans are surprisingly good at handling ambiguity, exceptions, missing information, bad data, and broken processes. That’s why NOT EVERYTHING CAN BE AUTOMATED, although we’re trying too! The biggest lesson from building agents isn't learning how to deploy them, it’s actually appreciating and learning how many invisible decisions humans make every single day. This is why I keep saying: Automation isn't the hard part, Reliability is! I suspect most of us are quietly firing more digital employees than we're hiring.

I stopped obsessing over publishing… and that’s when my research career took off. (Here’s what I discovered) Four years ago, I believed publishing was the only path to academic success. My inbox? Empty. My collaborations? Stagnant. My impact? Limited to footnotes in other people’s papers. Then, I did something radical: I shared my work outside journals. A blog post about the LCA work I did for a partner. A LinkedIn post breaking down techno-economic assessment and process design methods. A webinar sharing my research outputs. Crickets. For weeks. Until a founder DM’ed: "Liked the recording of your webinar. Can you do something like this for us to verify our TEA?" A week later, an academic mentor slid into my DMs: "Saw your recent work on carbon capture. Can we co-write a research bid?" I wasn't sure what to do. This wasn’t “real” academic work. I’d been pre-conditioned to share my work only in scholarly journals and conferences. But suddenly, my research was solving problems, not merely gathering dust. So I leaned in. I built a simple system:  1. Every paper made available as PDF with posts on problem, method, outputs  2. Conference slides became PDFs shared with key takeaways  3. Complex science converted into trade magazines and blog posts The response? A FTSE 250 energy company invited me to perform a market study for their new direct air capture business strategy. Learned societies invited me as a keynote speaker for their events. And yes, citations keep coming, but now tied to real-world impact. Here’s what academia won’t tell you: Visibility isn’t vanity. It’s the bridge between your work and the problems it can solve. You don’t need 100 papers to make a difference. You need the right people to see your work. Now? I teach researchers to build this bridge. Because your career shouldn’t hinge on how many journals you’ve cracked. It should hinge on how many minds you’ve changed. Start there. Let the papers follow. P.S. What is the most significant challenge that pre ents you getting your research seen by others? #science #scientist #research #publishing #phd #postdoctoral #professor #academia #highereducation

Cohere 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝗱 𝗮 𝗻𝗲𝘄 𝗴𝘂𝗶𝗱𝗲 𝗼𝗻 𝗵𝗼𝘄 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲-𝗿𝗲𝗮𝗱𝘆 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗶𝗻 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗲𝗱 𝗶𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗲𝘀: ⬇️ This is an excellent read for teams deploying agents in finance, healthcare, and other high-stakes environments — where security, tooling, and failure modes actually matter. 𝗛𝗲𝗿𝗲 𝗮𝗿𝗲 6 𝗸𝗲𝘆 𝗶𝗻𝘀𝗶𝗴𝗵𝘁𝘀: ⬇️ 1. 𝗧𝗼𝗼𝗹𝗶𝗻𝗴 𝗶𝘀 𝘁𝗵𝗲 𝗻𝗲𝘄 𝗿𝗶𝘀𝗸 𝘀𝘂𝗿𝗳𝗮𝗰𝗲 → Every tool an agent touches is a potential failure point. Cohere suggests defining each tool "like an API product" — with strict specs, validation layers, and fallback behavior. 2. 𝗟𝗟𝗠 𝘁𝗲𝗺𝗽𝗲𝗿𝗮𝘁𝘂𝗿𝗲 𝘀𝗲𝘁𝘁𝗶𝗻𝗴𝘀 𝗺𝗮𝘁𝘁𝗲𝗿 𝗺𝗼𝗿𝗲 𝘁𝗵𝗮𝗻 𝘆𝗼𝘂 𝘁𝗵𝗶𝗻𝗸 → They recommend a sweet spot between 0 and 0.3 for agent reasoning — to reduce randomness while still allowing adaptive behavior. It’s not just a UX tweak; it affects the agent’s ability to reason. 3. 𝗛𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗼𝗻𝘀 𝗮𝗿𝗲𝗻’𝘁 𝗷𝘂𝘀𝘁 𝗮𝗯𝗼𝘂𝘁 𝘄𝗿𝗼𝗻𝗴 𝗮𝗻𝘀𝘄𝗲𝗿𝘀 — 𝘁𝗵𝗲𝘆’𝗿𝗲 𝗮𝗯𝗼𝘂𝘁 𝘄𝗵𝗲𝗻 𝘁𝗼 𝗲𝘀𝗰𝗮𝗹𝗮𝘁𝗲 → Confidence thresholds and structured output (like JSON) are now standard. But the real importance is routing low-confidence outputs to human review before they hit production (human in the loop). 4. 𝗙𝗼𝗿𝗴𝗲𝘁 𝗽𝗿𝗼𝗺𝗽𝘁 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 — 𝘀𝘁𝗮𝗿𝘁 𝘄𝗶𝘁𝗵 𝘁𝗼𝗼𝗹 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 → The report flips the script: engineering precise, well-scoped tools is now more important than crafting clever prompts. Bad tool definitions = brittle agents. 5. 𝗠𝘂𝗹𝘁𝗶-𝘀𝘁𝗲𝗽 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀 𝗻𝗲𝗲𝗱 𝗹𝗮𝘆𝗲𝗿𝗲𝗱 𝗳𝗮𝗹𝗹𝗯𝗮𝗰𝗸 𝗹𝗼𝗴𝗶𝗰 → If one data source fails, agents should try backups, query logs, or escalate — *not* break. This kind of chained contingency planning is what separates MVPs from prod-ready systems. 6. 𝗠𝗼𝘀𝘁 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗶𝘀𝘀𝘂𝗲𝘀 𝗮𝗿𝗲𝗻’𝘁 𝗺𝗼𝗱𝗲𝗹-𝗿𝗲𝗹𝗮𝘁𝗲𝗱 — 𝘁𝗵𝗲𝘆’𝗿𝗲 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻 𝗳𝗮𝗶𝗹𝘂𝗿𝗲𝘀 → The real scaling bottlenecks are queue management, caching, circuit breakers, and tool timeouts. Without this ops layer, your agent won’t survive real-world load. More in the comments and report below. Enjoy. 𝗪𝗮𝗻𝘁 𝗺𝗼𝗿𝗲 𝗯𝗿𝗲𝗮𝗸𝗱𝗼𝘄𝗻𝘀 𝗹𝗶𝗸𝗲 𝘁𝗵𝗶𝘀? Subscribe to Human in the Loop — my new weekly deep dive on AI agents, real-world tools, and strategic insights: https://lnkd.in/dbf74Y9E