It is the step that gets you hired. Here is the five-stage roadmap 1,000+ PMs have followed, and where that step belongs.

You have probably seen a few AI PM roadmaps by now.

They all look alike. Learn the basics. Build something. Think like a senior PM. Study real companies.

But almost all of them get one step wrong.

And it is the step that gets you hired.

Over 1,000 PMs have followed the roadmap in this post, including PMs at Google, Microsoft and Coinbase. Here is the plan. The five stages first, with one thing to do in each. Then the step most roadmaps put in the wrong place. And how to fix it, starting this week.

Let us start with the stage most people rush through.

Stage 1: Know when AI is the right answer

Before you learn a single model, learn when not to use one.

A lot of AI products should not be AI at all. They are a rule, a form or a search box with a chatbot on top.

Here is how I would think about it. Ask three questions before anything else.

1. Is the task fuzzy or exact? Summarising a customer complaint is fuzzy. Calculating a refund is exact. Models are good at fuzzy. Plain code is better at exact.
2. What does a wrong answer cost? A bad playlist name costs nothing. A wrong refund costs money and trust.
3. Does the model have what it needs? If the answer lives in your order database, the model cannot guess it. It has to look it up.

Take a bank app. “What is my balance?” does not need AI. It needs a database query. “Why was my card declined?” is different. The facts are exact, but explaining them in plain words to a stressed customer is fuzzy. That part can be AI.

This is the judgment companies pay AI PMs for. Anyone can add a chatbot. Fewer people can tell you when not to.

That is the thinking part. Stage 2 is where you start building, and it is simpler than most people make it.

Stage 2: Get very good at prompting, then build a prototype

Prompting is the AI PM’s version of writing a spec. You are telling a system what a good output looks like.

Here is a drill that works. Pick one boring task from your own job. Release notes, user research summaries or weekly status updates. Write a prompt for it. Run it on 10 real examples. Keep rewriting the prompt until 8 out of 10 outputs are good enough to use without edits.

Then build a rough prototype in Claude Code. A small tool you can click and show someone. It does not need to be ready for real users. It needs to work on your laptop.

A prototype changes every conversation you have. In a product review, people argue about a slide. They react to something they can click.

Most people stop here. They can prompt, and they have built a toy. Stage 3 is where you learn why toys break.

Stage 3: Build one RAG app, one agent and one eval

You will keep seeing these three terms in AI PM job posts. In plain words:

• RAG lets the model look things up in your documents before it answers.
• An agent lets the model take steps and use tools, not just reply.
• An eval is a test that tells you whether the answers are good.

Do not watch another video on them. Build one of each.

• A RAG bot that answers questions from a set of help docs.
• An agent that does a three-step task. Read a customer email, look up the order, draft a reply.
• An eval with 20 questions and known good answers. Run your RAG bot on it and score the results.

The eval is the one most people skip. It is also the one that matters most. Without it, you cannot tell whether a change made the product better or worse. You are guessing.

You cannot make good product calls about systems you have never seen fail. You see them fail by building them.

Now you can build. Stage 4 is where the real jump happens. From building things to deciding things.

Stage 4: Think like a senior AI PM

A senior AI PM owns the whole thing. The spec, the experience, the price and the proof that it worked.

Let us break this down.

The spec. For an AI feature, the spec is not only what it does. It is what a good answer and a bad answer look like, with real examples. If you cannot write that down, your engineers cannot test it.

The experience. What does the user see when the AI is unsure? What happens when it is wrong? Good AI UX plans for the wrong answer, not only the right one.

The price. This is the part most PMs miss. Every answer your AI gives costs real money.

Think of it this way. Say one answer costs ₹0.50 in model calls. A typical user asks 40 questions a month. That is ₹20 per user per month, before servers, support or anything else. If your plan is ₹99 a month, this one feature eats about a fifth of your revenue. Now you have a pricing decision, not only a feature.

The proof. You shipped it. Did it work? Run an A/B test. Half your users get the feature, half do not. Then check whether it moved the metric you care about. Not whether people liked it. Whether the number moved.

This is where an AI feature stops being a demo and becomes a business.

One stage left. Then the step most roadmaps put in the wrong place.

Stage 5: Study how real companies ship AI

Pick three AI features you use every week. Mix consumer apps and business tools. For each one, write four lines.

1. What problem does it solve?
2. Why did they use AI here, and not a simple rule?
3. Roughly what does each answer cost them?
4. How would they know it is working?

Do this for a few weeks and patterns start to show. You see why some features feel useful, and others feel bolted on.

Those are the five stages. If you have seen AI PM roadmaps before, most of this probably looks familiar.

Here is what is missing.

The step most roadmaps get wrong

Interview prep.

Almost every roadmap puts it at the end. Stage 6. Learn everything first, then get ready for interviews.

That order is backwards. Three reasons.

Interviews test a different skill. You have to explain your thinking out loud, under time pressure, about a product you did not build. Knowing the material is not the same as explaining it to a stranger in 45 minutes. That takes practice, and practice takes time.

Waiting means late feedback. If you start mock interviews only after months of learning, that is when you find your gaps. By then, they are expensive to fix.

Explaining is how learning sticks. When you answer “should this be AI?” out loud, you find out whether you understood Stage 1 or only read it.

The interviewer is not testing whether you finished a roadmap. They are testing whether you can think like an AI PM in front of them.

Interview prep does not belong at the end. It belongs on day one.

Check out our self-paced AI PM course
(Including AI PM Interview Preparation ) 4.9/5 · 1120+ enrollments ·
Click for Discounts, see testimonials and course details

The fix: make interview prep the base layer

Picture the roadmap as a staircase. Five steps going up.

Most roadmaps add interview prep as a sixth step at the top. Move it to the bottom instead. Make it a base layer that runs under every stage.

Now go back to the stages you just read. Each one has an interview question hiding in it.

• Stage 1 trains AI product sense. Try: “Your bank wants an AI chatbot for customer support. Should it be AI? What would you build first?”
• Stage 2 trains you to talk about your own work. Try: “Walk me through something you prototyped. What did you change after testing it?”
• Stage 3 trains AI root cause analysis. Try: “Your support bot’s resolution rate fell from 70% to 55% last week. Find out why.” Was it the documents, the model, the prompt or the users?
• Stage 4 trains AI ROI. Try: “Each answer costs ₹0.50. Is this feature worth building? How would you price it?”
• Stage 5 trains AI strategy. Try: “A much bigger company launches your feature for free. What do you do?”

The rule is simple. Whatever stage you are on this week, practise its interview question out loud this week. Record yourself on your phone. Listen back once.

What week one looks like

If you started today, your first week would look like this.

• Learn: Stage 1. Pick three AI features and run each one through the three questions.
• Practise: Answer “Should this be AI?” out loud for one product you use every day.
• Check: Play back your recording. Did you give your answer and your reason in the first 30 seconds?

That is the whole pattern. Learn a stage. Practise its interview question. Check yourself. Then move up a step.

I have put the full week-by-week plan in one place. Each week pairs one roadmap stage with one interview drill, from day one to your first full mock interview.

12-week program details can be found here -
[Alumni Shipping at Google, Microsoft, Uber] — 1 out of 3 is getting placed

About Author

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

Here Is What Hiring Managers Actually Check

Imagine you are an Engineer/Analyst/Designer at a fintech app in Bengaluru.

You apply for an Associate PM role. Then another. Then ten more.

Every rejection says the same thing, politely. We are looking for candidates with prior product experience.

You open the job description again. 2+ years of product management experience. For a role that is supposed to be the first step into product management.

It feels like a trap. And the way most people try to get out of it makes the trap worse.

But there is a way out that works.

People use it every year to get into PM without a single day of the PM title.

By the end of this article, you will know what it is, and why the hiring manager wants it more than the years.

Where PMs actually come from

Start with a fact that sounds obvious once you hear it. Nobody is born with PM experience.

Some PMs did not start as PMs at all. They moved in from other roles:

• Engineers who kept asking why a feature was being built
• Analysts who found where users were dropping off
• Designers who argued for the user in every review
• Consultants who were already breaking down business problems for clients

So what got them their first PM role, if not PM experience?

They had done PM work before they had the PM title.

They picked a real problem, made a call on what to do about it and owned what happened next.

In the interview, they were not describing what a PM does. They were describing what they had done.

Now you are probably thinking, fine, but my role does not let me make those calls. I find the problem, I build the dashboard, and someone else decides.

Hold on to that thought. We will come back to it. First, there is a mistake almost everyone makes at exactly this point.

Wait, do not buy another certificate yet

This is the part where most people open a new tab and enrol in a course.

You have probably done it already. A product management certificate.

An AI course. A free playlist on YouTube with 40 hours of video. You finished half of it, and now it sits on your LinkedIn.

The rejections still come.

Here is why. A certificate proves you learned something.

The hiring manager is not worried about whether you can learn. They are worried about whether you can decide.

Ten more certificates do not answer that worry. They answer a question nobody in the interview room is asking.

The twist: nobody really wanted the years

Think about why the experience line is there in the first place.

A PM is a risky hire.

A bad PM picks the wrong problems, ships the wrong features and burns months of an engineering team’s time.

The hiring manager cannot see inside your head, so they use years as a shortcut. Years mean someone else already trusted you with product calls.

So the experience line is not really asking for years. It is a risk filter.

And a risk filter can be cleared with proof, not only with time. What the hiring manager wants to see is this:

1. You found a real user problem on your own
2. You decided what to build and what not to build
3. You put it in front of real people
4. You checked if it worked, and changed it when it did not

Show those four things and the years start to matter a lot less. Not zero. But a lot less.

Which brings us back to the thought you were holding on to. Your current role does not let you make those calls. So where does this proof come from?

You build the proof yourself

Outside your job. On your own time.

Go back to you, the data analyst at the fintech app. For months, your dashboards have shown the same thing.

A lot of new users drop off at the KYC step. They do not know which documents they need, or why their upload got rejected.

You flagged it in two meetings. Nothing happened. It was not your call to make.

So you make it your call. You build a small AI assistant that tells first-time users which documents they need for KYC, and explains in simple Hindi or English why an upload usually fails.

A few years ago, this needed an engineer. Today you can build it over a few weekends with tools like Claude and Lovable, without writing much code yourself.

Then you do the part that actually counts:

• You share it with 50 people. Friends, family, a few small shop owners near your house
• You watch where they get stuck
• You notice most of them skip the explanation. They just want a checklist
• You cut the explanation, ship the checklist, and more people finish the flow

That last change is the whole point. You did not follow a tutorial. You looked at real users, made a call and moved one number.

That is PM experience. It just did not come with the title.

The interview question you can finally answer

Now walk back into the interview room.

In all those earlier rejections, one question kept sinking you. Tell me about a product decision you made and how you knew it worked.

You used to answer with a case study you had read. The interviewer could tell.

This time you tell them about the KYC assistant. The drop-off you saw in your dashboards. The 50 users. The explanation nobody read. The checklist that worked.

Something changes. The interviewer stops reading your resume and starts asking follow-ups. Why a checklist? What would you build next? How would this work for a few lakh users a month?

That is the moment the experience line stops mattering. You are not the candidate with zero PM experience anymore. You are the candidate who has done the job without the title, and that is definitely what the filter was looking for all along.

So why is not everyone doing this?

If building your own proof works, why are so many people still stuck at the experience line?

Because doing it alone breaks in three places:

• You pick a problem that is too big, and never ship anything
• You ship something, but never get real users to try it
• You get users, but cannot turn the story into an answer that holds up in an interview

None of these is about talent. They are about not having anyone tell you, early, that your problem is too big or that you are measuring the wrong thing.

On your own, you find that out after three months. With someone who has done it, you find out in week one.

Where to start?

This is exactly how we run the AI PM Builder Cohort at Technomanagers.

You do not collect another certificate.

You pick a real problem, build a real AI product with tools like Claude, put it in front of real users and learn to talk about it in PM interviews. There is dedicated interview prep inside the cohort, so what you build turns into answers that land.

Our placement rate is 30–35%, and our alumni now work at companies like Google, Microsoft and Oracle.

Want to build all of this live, with a capstone project and interview prep? Check the AI PM Builder Cohort here

You do not need the title to start doing the work. You just need to start.

AI Product Management 2027 — Winner’s Playbook

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

Stop doing what worked in 2026

Do this to win in 2027 (and stop doing what worked in 2026)

In this article, we will discuss:

1. The AI PM Trap of 2027

2. Why Most PMs Will Fail in 2027

3. What a Real Winner Looks Like in 2027

4. The New PM Interviews (AI Interviews 2.0)

5. The 2027 Roadmap to Become an AI PM

It is late at night, and you are scrolling LinkedIn.

One PM has posted a demo of an AI agent she built over the weekend. The next post is about another round of layoffs.

The one after that is a job description that says AI native PM; must have shipped agents in production.

You close the app. You open it again and find Microsoft, Amazon, and Google all planning layoffs.

If you are a Product Manager going into 2027, you are probably feeling this.

Last year you were told to learn AI. You did. You learnt prompting, you built a prototype, you even shipped a small AI feature.

And still it feels like the bar has moved again.

It has. Let me show you where it moved.

The AI PM Trap of 2027

In 2026, the trap was simple. People watched a few YouTube videos on vibe coding and called themselves AI PMs.

In 2027, the trap is more dangerous because it looks like real work.

You built a prototype in Lovable or Claude Code. It worked in the demo. Your manager was impressed. So you think you have made it.

Here is the truth.

A demo that works once is NOT equal to a product that works for a million users (Demo is easy, production is hard)

Prototyping is now table stakes. Every PM can do it. Your intern can do it. So it no longer makes you stand out.

Quick self-check. Give yourself 1 point for every question you can answer with confidence today.

• Can you write evals for your AI feature and tell which ones should be deterministic and which ones need an LLM as judge?
• Can you explain why your agent failed on step 4 of a 7-step task, and not just that the final answer was wrong?
• Can you tell what it costs you every time a user asks your AI one question?
• Can you decide what your agent is allowed to do on its own, and where a human must approve?
• Can you explain what context your model sees, and why giving it more context made it worse?
• Can you price an AI product that does the work, not just helps with the work?

If you scored 5 or 6, you are ahead of most PMs in the market.

If you scored 2 or less, do not panic. But do not ignore it either. This article is for you.

Why Most PMs Will Fail in 2027

The majority of Product Managers will learn AI tools. Very few will learn how AI products actually work.

And in 2027, that difference becomes very expensive.

Three things have changed.

1. First, AI now does the work most PMs were proud of. It writes the PRD, the user stories, the competitor research and even the first prototype. If your value was in writing documents, that value is gone.
2. Second, companies have moved from AI that talks to AI that acts. A chatbot that gives a wrong answer is embarrassing. An agent that books the wrong flight, refunds the wrong customer or deletes the wrong file costs real money. Companies need PMs who can control that risk, not PMs who can make a nice demo.
3. Third, tools change every three months. The PM who only learnt tools has to start again every quarter. The PM who learnt the system underneath (data, models, context, evals, cost) just picks up the new tool and moves on.

2026 exposed the PMs who never learnt AI.

2027 will expose the ones who only learnt the surface of it. I know this is brutal, but this is the truth, and I am definitely sure you don’t want to be the one.

What a Real Winner Looks Like in 2027

The bar has moved one level up. Here is the simplest way to see it.

• 2026 winner built a RAG chatbot. 2027 winner can prove it answers correctly and knows what every answer costs.
• 2026 winner vibe coded a prototype. 2027 winner took that prototype to production with specs and evals.
• 2026 winner knew what an agent is. 2027 winner decided what the agent is allowed to do without asking a human.
• 2026 winner wrote good prompts. 2027 winner designs the full context the model sees.

So a true Product Manager in 2027 should be able to

1. Write evals that catch failures before users do ( Deterministic evals, LLM as Judge, and step-level evals for agents )
2. Design agents with clear permissions: what it can do alone and where a human must click approve ( Autonomy vs Control )
3. Engineer the context, decide which documents, memory and tools the model sees, because more context is not better context
4. Manage the unit economics, cost per task, when a small model is enough and when you need a frontier model
5. Monitor AI in production, because a feature that worked in January can quietly break in March after a model update
6. Price AI products by outcome, per task or per resolution, not only per seat
7. Know when NOT to use AI, and keep simple rule-based flows where the audit trail matters
8. Many More

The New PM Interviews (AI Interviews 2.0)

Interviews have changed again.

In 2026 they asked you questions about AI. In 2027 they ask you to use AI in front of them.

Some interview loops now give you Cursor, Lovable or Bolt and ask you to build a working prototype in about 45 minutes while the interviewer watches. Meta has reportedly added a Product Sense with AI round, where you use AI during the case itself.

The test is not whether you can prompt. The test is whether you can spot when AI gives you a confident wrong answer, and fix it.

Behavioural questions also went one level deeper. Tell me about a trade-off is now: Tell me about a time your AI feature got worse after launch. How did you find out, and what did you do?

They are testing AI Product Sense, Agent Evals, Reliability, Cost, AI Pricing, AI Safety and your judgment on where not to use AI.

Some examples are below

1. Your AI support agent resolves most tickets on its own, but customer satisfaction has dropped. What do you do? ( Evals and Metrics )
2. Write the evals for a travel booking agent that can book and cancel tickets on its own. ( Agent Evals )
3. How would you price an AI agent that does the work of a sales rep? Per seat, per task or per outcome? ( AI Pricing )
4. Model costs drop 10x next year. What changes in your product strategy? ( AI Strategy )
5. You are the PM of a paid AI model. Your biggest enterprise customer wants to move to an open-source model that is good enough and much cheaper. How do you respond? ( AI Strategy )
6. In this product, where would you NOT use AI, and why? ( AI Judgment )

Notice one thing. None of these can be answered with a framework you memorised. You need to have built or broken an AI system to answer them well.

For More Real AI PM Interview Questions (Click Here)

The 2027 Roadmap to Become an AI PM

You need a plan to move from a traditional PM to an AI PM. Here is the step-by-step guide. I have added a checkpoint to every step, so you know when you are actually done and can move on.

This is the same roadmap 800+ PMs have followed in our AI PM Course, including PMs who now ship AI at Google, Microsoft and Coinbase.

#1 Understand Basic AI and ML

→ Machine Learning, Artificial Intelligence, Deep Learning, Generative AI, Linear and Logistic Regression, Decision Trees, Training vs Inference

#2 Learn from Real AI Case Studies

→ How Google, Microsoft, Amazon, OpenAI, Netflix, Nvidia and Uber use AI in their products and pricing

#3 Understand ML Systems and Data Architecture

→ AI Product Lifecycle, AI Data Pipelines, the AI Flywheel, ML System Design

#4 Master Advanced Prompting and Context

→ How to write the Perfect Prompt, System Prompts, Temperature Control, Advanced Prompting ( CoT, ToT, etc. ), getting consistent outputs, deciding what context the model sees

#5 Vibe Coding and Prototyping

→ Beginner to Advanced Vibe coding, Model Parameters, the PM Reliability Framework for prototyping

#6 Understand RAG, Agents and Evals

→ RAG System Architecture, Chunking and Retrieval, RAG Metrics, Designing Agents, Agentic Workflows, Evals, Deterministic and Probabilistic Evals, LLM as Judge

#7 AI Interview Preparation

→ AI Product Sense, AI Success Metrics, AI Pricing, AI Evals, AI Product Strategy ( Real questions like success metrics of GPT 5.0 and evals for a travel booking agent )

#8 Ship One Thing and Show It

→ Build one small AI product end to end, write its evals, and post what broke and how you fixed it

Steps 1 to 8are what the course is built around.

Are You Ready for 2027?

The world is changing fast.

You can keep learning tools that change every three months and feel the stress every time a new one drops. Or you can learn the system underneath them, and become the PM who decides how AI gets built.

Go back to the self-check at the top. Whatever your score is today, you can take it to 6 by the time 2027 starts.

Most Detailed AI Product Management Course ( 50+ Videos, 25+ Case Studies, 20+ Real AI PM Interview Questions )

1000+ PMs have enrolled, including PMs who now ship AI at Google, Microsoft and Coinbase. Highest Rated Course: 4.9 / 5. Testimonials Here

For the New Year, we are giving an EXTRA 40% OFF on our AI PM Flagship Course for a very limited time: Coupon Code: NYE27

Want to build all of this live, with a capstone project and interview prep? Check the AI PM Builder Cohort here

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

AI Product Management 2027 — Winner’s Playbook was originally published in Agile Insider on Medium, where people are continuing the conversation by highlighting and responding to this story.

Imagine you are a product manager at a lending app in Bengaluru.

Five years in, good reviews, and three AI courses finished this year.

You can explain RAG in a meeting. You know what an agent is.

So you apply for an AI PM role and clear the first round.

Then the hiring manager asks one question. You launched an AI feature. How did you know it was good enough to ship?

You start talking about A/B tests. She waits. She was asking about evals, and you have never written one.

That is the gap most PMs carry into AI PM interviews. It is not a knowledge gap. It is a doing gap.

You have read about all of it. You have done very little of it. No certificate shows the difference, but an interview shows it in about two minutes.

So before your next interview, score yourself on 8 skills. It takes 10 minutes.

How to score yourself honestly

Rate each skill from 1 to 5:

• 1: I have never done this
• 2: I know the idea but have not done it
• 3: I have done it once, with help
• 4: I do it alone at work
• 5: I could teach it

One rule. Rate what you have done, not what you have read.

A 4 needs work you could show someone. If you only watched a video on it, that is a 2, even if you understood every word.

Below are the 8 skills. Each one has a single test question that tells you your real score, and the score you gave yourself as our lending app PM.

3 AI PM skills your current job already trained

These feel familiar, so PMs rate them too high. Answer the test question, not the topic.

Picking problems where AI fits.

Test: can you name an AI idea you dropped because the data, the cost or the error risk did not work? Knowing when not to use AI is the skill here.

You reshaped one idea once, with your tech lead’s help. Your score: 3.3.

AI metrics and experiments.

Test: for one AI feature, can you name a quality metric, an adoption metric, a business metric and a guardrail? And do you know when an offline eval is enough and when you need an A/B test?

Five years of dashboards help. Your score: 3.7.

Shipping and communicating AI work.

Test: can you explain an AI feature’s quality, risk and cost to a senior leader in 5 minutes? Do you have a plan for when the model is wrong in production, with a fallback and a person who can roll back?

This is where your PM years show. Your score: 4.3.

3 AI PM skills the AI job adds

This is where the doing gap lives, and where AI PM interviews go deepest.

Defining quality and writing evals.

Test: have you written an eval set with common cases, edge cases and never cases, then made a ship or don’t ship call from the results?

PMs who know AI often score lowest here, because evals are written, not read about. Your score: 1.7.

RAG and retrieval.

Test: when a RAG bot gives a wrong answer, can you tell if retrieval fetched the wrong chunk, or the model misread the right one?

That one call definitely decides what the team fixes next. Your score: 2.3.

Agents and tools.

Test: have you built or specified an agent that calls at least one tool, and decided where it needs a human check before it acts?

Your score: 2.0.

2 skills that decide if your AI feature survives the budget review

Many AI features work in the demo and die in a finance meeting. These two skills are why.

Cost and latency.

Test: can you work out cost per query from tokens and the price per million tokens, then name three ways to cut it, with the quality risk of each?

The maths is simple, and nobody does it. Say your loan assistant gets 2 lakh queries a day. A 50 paise difference per query is Rs 30 lakh a month. If you can’t show that number, someone in finance will decide your feature’s fate without you. Your score: 2.7.

AI strategy and pricing.

Test: can you say where the moat of an AI product is, beyond the model? Every competitor can call the same model API.

The moat is your data, your workflow or your distribution. And can you say how cost per user changes the price? Your score: 2.7.

Download this FREE AI PM Skill Checklist

Why your overall AI PM score lies

Add it up, and you score 2.8 out of 5. That reads as working level. Not bad.

Now look at where the points come from.

• Your three highest scores are the skills your old PM job trained: communicating, metrics, picking problems.
• Your three lowest are evals, agents and RAG. The building skills.

That is exactly the profile that struggles in AI PM interviews. Strong at talking about AI work, weak at doing it. The average hides it completely.

So ignore the overall score. Sort the 8 skills from lowest to highest and look only at the bottom two. Treat them separately from the rest.

For you, that is evals at 1.7 and agents at 2.0.

What to do this week about your top 2 gaps

Don’t try to fix all 8. Fix two properly, then score yourself again.

If evals is your gap:

• Pick one AI feature, yours or one you use daily, like search on a quick-commerce app.
• Write 20 test cases: common ones, edge cases, and never cases where a certain answer must not happen.
• For each, write what a good answer must include.
• Run them, mark pass or fail, and make the call: would you ship it?

That moves you from a 1 to a 3. More useful, it gives you a real answer to the interview question this article started with.

If agents are your gap:

• Build a one-tool agent with Claude. For example, one that reads a CSV of support tickets and drafts a weekly summary.
• Write down every step it took.
• Then decide one thing it must never do without asking you first.

That last step is the part interviewers care about. Anyone can make an agent run. The PM decides where it must stop.

Score yourself again after 4 to 6 weeks. If your bottom two moved up, you are spending your time on the right things. If they didn’t, you were reading again instead of building.

I have put this whole self-check into a free scoring sheet that works in Google Sheets or Excel. 24 statements across the 8 skills.

You rate each one. It works out your level for every skill, ranks your gaps, and gives you a next step that fits your level.

If you want to close your gaps with someone reviewing your work every week, that is what our AI PM Builder Cohort is built around.

Our next AI Product Manager & Builder Cohort is starting. For an early-bird discount, you can fill out the form here: AI PM Builder Cohort
[Alumni Shipping at Google, Microsoft, Uber]

And if this helped, subscribe so the next practical breakdown lands in your inbox.

About Author

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

AI Learning Resources

1. 12 Weeks LIVE AI PM Cohort (Alumni now ship AI at Google/Microsoft, etc)
2. AI PM Course [Rated 4.9/5 from 1000+ Product Managers from Google, Coinbase, Microsoft, Uber, etc]
3. AI Program Management Course (For Program Managers and TPMs)
4. AI Business Course (For General Management, Biz Ops & Strategy Roles)

Why Is Amazon Turning Away Customers Who Are Ready to Buy?

AI Agent Strategy Explained

Imagine you are a product manager at Amazon.

Meta has just launched Muse, a personal AI shopping agent. You tell it what you need, and it browses, compares and buys for you once you approve. This is agentic commerce, and people are actually using it.

Subscribe to get FREE AI PM resources in your welcome Email.

Then your team ships a popup that blocks Muse from shopping on Amazon.

Why would the most customer-obsessed company in the world turn away a customer who is ready to pay?

Amazon’s stated reasons are permission and safety. Meta never asked, Muse doesn’t identify itself, and it appears to store login details. Fair points, but they don’t explain the whole decision.

The real answer shows up when you sit in four chairs, one by one.

Chair 1: Amazon, and why AI agents hurt its ad business

It’s Monday. Your VP opens two dashboards.

Sales looks great. Muse orders convert fast, with almost no drop-off. Ads looks empty. Not one sponsored product was seen on those orders.

Guess which dashboard your VP keeps staring at.

Here’s the concept that explains it.

A shopping page is not a shelf; it’s a negotiation. Every slot is ranked for many goals at once: what you want, which brand paid, what has better margin, what stock needs to clear.

Teams call this multi-objective ranking. An AI agent ends that negotiation, because the page works for the store but the agent works for the buyer.

And for Amazon, that negotiation is the business. Its ad revenue is now roughly double the profit of its entire stores business. Ads are not side income anymore; they quietly fund the store.

So when an agent skips the page, it skips the part that pays the bills. Blocking Muse is the rational call.

But watch what Amazon does elsewhere. Its own agent, Buy for Me, shops on other brands’ websites, and brands had to email Amazon to opt out. It’s also piloting ads inside ChatGPT through its own ad platform.

So Amazon is not against AI agents. It’s against AI agents it can’t charge.

Chair 2: Walmart, and the AI shopping stat everyone misread

When Walmart sold inside ChatGPT, purchases made in the chat converted at about a third of the rate of shoppers who clicked through to Walmart.com.

Most people read that and said, see, AI shopping agents don’t work.

Sit in Walmart’s chair and ask the boring PM question first. Compared to what?

Both groups came from ChatGPT. So this was never agents versus humans. It was checkout inside a chatbot versus checkout on Walmart’s own site.

Walmart’s team blamed the product, not the idea. Instant Checkout bought one item at a time, so people worried about getting separate boxes for what should have been one order. That’s a broken cart.

So Walmart did the opposite of Amazon. It opened an Agent Gateway for approved agents like Muse, ChatGPT and Gemini, with spending limits that customers control.

The strategy is simple. Most US shoppers start on Amazon, so when Amazon blocks agents, those agents look elsewhere. For the challenger, the leader’s wall is free distribution.

Chair 3: Shopify, where AI agents are pure upside

This one is simple.

Shopify has no big ads business to protect. It earns when its merchants sell, mostly through payments. So an agent that brings a ready buyer to a small store is only good news.

That’s why Shopify partnered with Muse for checkout with Shop Pay. Shopify doesn’t fight for the page. It fights for the pay button inside the agent.

A note for those of us building in India. UPI is a public rail, and NPCI has already tested agent payments on it. When no one can own the wallet, the fight moves one step up, to who hears the customer first.

Chair 4: Meta, and the awkward truth about AI agents

Before you get angry at Amazon, ask one question.

Would Meta let someone else’s AI agent scroll Instagram for its users and skip every ad?

I doubt it. Instagram makes money because you keep scrolling past ads. Muse does to Amazon exactly what Meta would hate someone doing to Instagram.

Even Muse’s business model gives it away. Meta plans to keep it free for most people and take a fee from businesses on the sales it helps complete.

Amazon does the same with Buy for Me. Everyone loves AI agents on someone else’s website.

So Meta’s smart move is to take away Amazon’s best excuse. With Sierra, it launched the Personal Agent Protocol, backed by Walmart, Shopify, Stripe, Visa, Mastercard and others. Agents identify themselves and never need your password.

It’s still a draft. But this is strategy through standards: when you can’t win the negotiation, change the rules everyone plays by. Once agents are safe by design, a block looks like what it is, protecting ad money.

The sale or scroll test: one question that predicts any AI agent strategy

Put the four chairs side by side, and a pattern appears.

• Amazon earns big on the scroll. It built a wall.
• Walmart has ads too, but it’s the challenger. It built a door.
• Shopify earns on the sale. It opened the house.
• Meta earns on the scroll at home and wants the sale everywhere else.

So for any product, ask one question. Do you make money from the sale or from the scroll?

• Mostly the sale: agents are a new channel. Open up.
• Mostly the scroll: agents are a threat. You will build a wall.
• Both, or you are the challenger: build a door you control.

Tell me how a company makes money, and I’ll tell you its AI agent policy.

How advertising changes when an AI agent does the shopping

Ads won’t die. They will move from the store’s page into the agent’s answer.

That’s why Amazon is selling ads into ChatGPT. It wants to keep selling the slot, even when the slot lives in someone else’s agent. The real fight was never about who sells the diapers; it’s about who sells the sponsored slot above them.

But there’s a catch. An agent is only useful if you trust it works for you, and every paid result it shows spends a little of that trust.

So ads aimed at agents can’t sell a mood with a nice photo. They will have to win on the facts an agent checks: price, delivery date, return policy and reviews.

What product managers should do about AI shopping agents

Start by running the sale or scroll test on your own product. Then:

• Find the metrics that depend on a human scrolling, like ad impressions, carousel clicks and attach rate. Agents break these first, so start tracking how often agents shortlist you instead.
• Make your catalogue readable by machines. Price, stock, delivery date and returns should be clean data, not text inside a banner.
• Turn cross-sell into something an agent can ask for. It will never see your carousel, but it can ask what goes with this.
• Track agent orders separately, and always ask compared to what before you judge them.
• Pick wall, door or open house on purpose. Don’t let a default setting decide your strategy.

What happens next

• For now, Amazon will definitely keep the wall, and Muse grows on Walmart and Shopify stores.
• Then Amazon opens a paid door. Approved agents get in, and brands pay to rank inside it.
• The real loser is the store with neither Amazon’s delivery nor a door. Agents will simply skip it.

Reading a business model, questioning a metric, designing for an agent instead of a screen. That’s the AI PM job now, and it’s what we practise every week.

Our next AI Product Manager & Builder Cohort is starting. For an early-bird discount, you can fill out the form here: AI PM Builder Cohort
[Alumni Shipping at Google, Microsoft, Uber]

Which of these four chairs would you least want to sit in right now? Tell me in the comments.

About Author

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

AI Learning Resources

1. 12 Weeks LIVE AI PM Cohort (Alumni now ship AI at Google/Microsoft, etc)
2. AI PM Course [Rated 4.9/5 from 1000+ Product Managers from Google, Coinbase, Microsoft, Uber, etc]
3. AI Program Management Course (For Program Managers and TPMs)
4. AI Business Course (For General Management, Biz Ops & Strategy Roles)

Why Apple Is Selling a Bigger Screen When AI Needs Fewer

Why is Apple selling a bigger screen at the exact moment AI is trying to make screens matter less?

Because the iPhone Duo is not really a foldable bet. It is Apple’s answer to AI agents. And it solves only half the problem.

First, the metric Apple is protecting. Then the four levers the Duo pulls. Then the risks the Duo cannot touch — finally, my bet on where phones and apps go next.

Apple’s North Star Metric

Imagine you are a product manager at Apple in 2026. iPhone revenue grew 22% to $54.3 billion in the June 2026 quarter, and Apple has more than 2.5 billion active devices.

iPhone Ecosystem Revenue = Hardware Revenue + Services Revenue

• Hardware Revenue = iPhones Sold x Average Selling Price
• Services Revenue = Installed Base x Money Per Device That Flows Through Apple

Money flows through Apple when you pay in the App Store, subscribe through Apple, or search in Safari, because Google pays Apple to be the default search engine.

Hardware is the bigger number. Services is the profit engine, with gross margins around 75% and $30.7 billion of revenue in the June quarter alone.

The Duo pulls the hardware levers. AI threatens the services term. That gap is the whole story.

Lever 1: Average Selling Price

Most premium buyers who want an iPhone already have one. So growth has to come from each buyer paying more, not from more buyers.

The Duo starts at $1,999 in the US and Rs 2,99,900 in India, the highest starting price of any iPhone. IDC forecasts 10 million Duos in the first 12 months, bringing in about $27 billion. That is roughly $2,700 per phone, well above the starting price, because buyers lean towards the bigger storage models.

Fewer phones, more money. A new shape also gives people a reason to upgrade, which a slightly better camera no longer does.

Lever 2: The Top of the Market

Foldables are small. Omdia counts 18.2 million shipped in 2025, under 2% of all smartphones.

Apple still needed one, because foldables are where the most status driven buyers go, and Apple had nothing there. Samsung and Huawei owned the segment. In the first half of 2026, Huawei alone had 57% of book-style foldables, per Omdia, and most of its strength is in China, Apple’s most contested market.

Losing that buyer is not a units problem. It is a brand problem. The person who buys the most expensive phone decides what premium means for everyone below them.

Lever 3: Timing

Apple waited until hinges, displays and demand were proven. Vision Pro shows why. It launched before a market existed and struggled.

Foldables already have a market, so entering late carries little risk. But growth had slowed to 6% in 2025, per Omdia. The market exists, yet nobody had given mainstream buyers a strong enough reason. That is the gap Apple’s software work aims at.

These three levers explain the Duo as a product. They do not explain why the new CEO opened the launch talking about AI, or why Apple rebuilt iOS for the Duo instead of just stretching it. That is lever 4.

Lever 4: Keeping the iPhone as the AI Hub

John Ternus opened his first launch as CEO by calling the iPhone the ideal intelligent personal hub for AI. That is the real bet: AI will not replace the phone; it will make the phone more essential.

To test that bet, split any personal AI into four layers.

• Model: the AI that thinks.
• Context: what it knows about you, like messages, photos, calendar and location.
• Action: its ability to actually do things inside apps and services.
• Device: where you see and use it.

Now look at who owns which layer.

• Apple rents the model. Google’s Gemini powers the new Siri, reportedly for about $1 billion a year. Apple owns the device and most of the context, and is fighting for action through App Intents, which let Siri work inside apps. Apple says Siri AI already works with over 300,000 apps.
• Meta owns the model and the action. Its Muse agent, launched a day before Apple’s event, runs on a dedicated cloud computer for each user and keeps working when you are not. But it borrows the device. You talk to it in WhatsApp, on your iPhone.
• OpenAI owns the model and is building its own device with Jony Ive’s team.
• Google owns the model and the Android device, and it also powers Siri. It wins either way.

Renting the model is Apple’s smartest AI decision in years. Models are becoming a commodity, and Apple kept the customer without winning the model race.

Apple is also widening its context. The new Apple Watch can listen through the day and summarise your conversations, a feature called Siri Recap. It is opt-in, and Apple says no recordings are stored. But the wearer agrees and the people they talk to do not, so this is where Apple’s privacy brand gets tested.

The Duo makes the device layer even stronger. But the device was never Apple’s weak layer. Action is.

The Risks: Where Apple Can Lose the Action Layer

Cloud agents work when your phone does not

The most useful agent will work while you sleep. A phone runs on a battery and sits in your pocket most of the day. That is why Meta put Muse in the cloud, while Apple’s hub lives on the device.

Apple’s approach needs permission; agents do not

For Siri to act in an app, the developer must build for it. Now the same developer is also asked to redesign for the Duo.

Developers do this only out of self-interest. Netflix refused to build a Vision Pro app in 2024 and showed up for the Duo, because the iPhone matters to its business and Vision Pro did not.

An agent that uses websites the way a person does needs nobody’s permission. Such agents are fragile and legally contested today, and Amazon has sued Perplexity over its Comet browser agent. So Apple’s approach is more reliable now. But reliability improves with time, while coverage is structural.

Apple uses AI to improve apps, not replace them

Every Apple Intelligence example improves an existing app, like updating your calendar and reminders when the school emails. The bigger shift may be an agent that just handles the task, so you never open the reminders app at all.

The money at risk is not where most people look

Agents will not hurt Apple’s cut on cab rides or grocery orders. Apple never took a commission on physical goods and services. The real exposure sits in three places.

• Search. Google paid Apple about $20 billion in 2022 to be Safari’s default, as revealed in the US antitrust case. Every question asked to an AI is a search Safari never sees.
• Subscriptions. Apple’s commission lives on digital goods, and US court rulings and EU rules already push Apple to allow links to outside payments. Agents that sign users up outside the app squeeze this further.
• Leverage. Developers come to Apple because the iPhone is the only route to the customer. If agents become a second route, Apple loses the pull that got Netflix to show up.

A foldable touches none of these.

What This Means for Phones and Apps

Phones shift from doing to approving. Agents cut the taps in a task, but you still compare options, read summaries and approve payments. A big screen helps you review, and Face ID, Touch ID and the secure chip become the proof that it is really you asking.

So the phone stays, but as the device that checks and confirms, not the place where the work happens.

Foldables stay small in volume. Even at 45 million units in 2030, Omdia expects them to be under 3% of all phones. They will matter for profit and brand, not units.

Apps split into two kinds.

• Experience apps stay: games, video, social and creative tools. People open them to spend time there. Apple showcased Netflix, Zoom and Slack for the Duo, so the Duo is a bet on this kind of app.
• Task apps get absorbed: booking, ordering, reminders and forms. People open them only because they have to. Agents will do these tasks, and the apps become back ends that agents call.

When the agent picks the cab, the cab app’s brand matters less. The new fight is not App Store ranking. It is being the option the agent picks.

My Bet

Short term, the Duo is definitely a smart move. It raises price when users are near a ceiling, closes Apple’s only gap at the top of the market, and carries little risk.

Long term, it strengthens the layer Apple already owns and leaves the one it can lose untouched.

If I had to bet, Apple ships a Siri that runs ongoing tasks in the cloud within three years, and lets it act on the open web beyond App Intents. Not because Apple wants to, but because users will expect what Muse already does.

Apple does not need the best model. It needs Siri to be the agent people trust with their tasks, not the entry point that hands those tasks to someone else’s agent.

If you want to go from reading strategy like this to building AI products yourself, that is what we do in the AI PM Builder Cohort.

Our next AI Product Manager & Builder Cohort is starting. For an early-bird discount, you can fill out the form here: AI PM Builder Cohort

[Alumni Shipping at Google, Microsoft, Uber]

About Author

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

AI Learning Resources

1. 12 Weeks LIVE AI PM Cohort (Alumni now ship AI at Google/Microsoft, etc)
2. Quick AI PM Course (20 Hours of Content) [Rated 4.9/5 from 1000+ Product Managers from Google, Coinbase, Microsoft, etc]
3. AI Program Management Course (For Program Managers and TPMs)
4. AI Business Course (For General Management, Biz Ops & Strategy Roles)

Most People Stop at the First One

This is applicable if you are looking for Product Management roles, Program Management Roles, or Business/Ops/Strategy Roles

Imagine you are a product manager at a fintech app that gives small personal loans.

It is Monday morning. Big meeting with the CEO.

Your director is going through your weekly update in front of everyone. She tells the CEO that one of your features is ready.

Then your phone buzzes. It is the data science lead. That feature is not ready. It broke on Thursday night.

Your director has already moved on. The whole room now believes something that is not true. And it came from your update.

Here is the worst part. You did not even write that update. Your AI did. It writes it on its own every Friday. Last week, you were proud of that.

You use the same ChatGPT and Claude as everyone else. So what went wrong?

Let us go back two weeks.

Your app has one big problem. To get a loan, people upload their bank statement and type in their salary. Then they wait up to 24 hours for someone to check it. This is step 4 of the application, and 41% of people give up right here. Some of them needed that money the same week, for a hospital bill or a school fee.

Your plan is simple. Use AI to read the bank statement and fill in the salary in a few seconds. No waiting. You write the PRD. You also run the whole project with four teams.

Over the next two weeks, seven things will go wrong. Each one teaches you one AI skill. Most people learn only the first one, and then say AI is overrated.

PM, program manager, TPM, data scientist or engineer, the skills are the same for you. Each section ends with what it means for your role.

Skill 1: Prompting

This is what you see: Monday evening. Your VP wants the PRD by Friday. You open Claude and type one line.

Write a PRD for an AI feature that reads bank statements in our loan app.

Ten seconds later, four neat pages. Personas, goals, metrics. You feel relief. Then you actually read it.

The personas are Busy Professional and First-Time Borrower. The main goal is to improve user experience. It does not know about step 4. It does not know about the 41%. It could be a PRD for any lending app in India. Your VP would see through it in one minute.

This is what you learn: the AI did exactly what you asked. You asked for a PRD. You never told it which one. A good prompt has four parts. Instruction: what you want. Context: what it cannot guess. Constraints: length, format and tone. Example: what good looks like.

This is what you do: same request, now in four parts.

Instruction: Draft the problem, goals and solution sections of a PRD.

Context: We are a personal loan app. Last month 1,20,000 users started an application and 38,000 finished. At step 4, users upload a bank statement, type their salary and wait up to 24 hours for ops to verify. 41% of users who reach step 4 leave. The new feature: a model reads the statement and fills in the income.

Constraints: Max 2 pages. Every goal must have a number. No personas. Add a section on what happens when the model is not sure. List open questions at the end.

Example: Attached is our UPI autopay PRD. My VP approved it in one round. Match its structure and depth.

Now the first line of the draft is about step 4. The goal reads: cut step 4 drop-off from 41% to 25%. There is a section on what happens when the model is not sure.

Of the four parts, the example does the heaviest lifting. It carries your company’s standard without you having to explain it. So keep your two or three best approved documents in one folder. You will use them again and again.

A bad answer is usually a bad question.

For your role: a program manager attaches last quarter’s best status report as the example. A data scientist attaches the eval report the team liked.

Skill 2: Make AI disagree with you

This is what you see: Tuesday, you get a big idea. If the model is confident about the income, skip the ops check fully. No 24-hour wait at all. You ask Claude if this is a good idea.

It says yes. Five benefits. Faster approvals, lower ops cost, happier users. You walk into Thursday’s review feeling sharp.

The risk head asks two questions.

1. How often does the model read a transfer from a friend as salary?
2. What is the error rate on scanned PDFs from smaller banks?

The room goes quiet. You look at your laptop as if the answer might be there. It is not. The idea goes back.

This is what you learn: AI agrees with you by default. This is called sycophancy. Models are trained partly on human ratings, and people tend to rate answers that agree with them higher. Anthropic’s 2023 research found this pattern in five leading AI assistants. So when you ask whether your idea is good, you are asking something that already leans towards yes.

This is what you do: you save a reviewer instruction in your project, so it applies to every chat.

You are my reviewer, not my supporter.
1. Before you agree with anything, name the one assumption I have not tested.
2. Argue the strongest case against my plan first. Do not soften it.
3. If I push back, change your view only if I give new evidence, not because I objected.
4. If you cannot find a real flaw, say so. Do not invent one.
5. End with the one question my toughest stakeholder will ask.

That evening, you run the same idea through it. The first line back: you are assuming a high confidence score means the income is correct, and you have no data on how often a confident answer is wrong.

It is the risk head’s question, almost word for word. Two days too late this time. Never again.

One caution. The pushback is useful, but any fact inside it still needs checking. If it quotes a number, ask where that number came from.

If your AI never disagrees with you, it is not reviewing your work.

For your role: a program manager runs the launch plan through it before the steering committee. An engineer runs the design doc through it before the architecture review.

Our next AI Product Manager & Builder Cohort is starting. For an early-bird discount, you can fill out the form here:

AI PM Builder Cohort

[Alumni Shipping at Google, Microsoft, Uber]

Skill 3: Context engineering

This is what you see: Thursday night, 10 PM. You have opened eleven chats this week. In every one, you pasted the same paragraph. Loan app. Step 4. 41% drop. 24-hour ops check. My VP hates long documents. Risk signs off on anything with income data.

On the days you were tired and skipped it, the answers went generic again. You are explaining your own product to a machine for the eleventh time.

This is what you learn: the fix in skill 1 was never really about the prompt. It was about the context. Most of what AI needs is in your head, not in the chat. This is why people now call it context engineering. Write it once in a plain text file. Most AI tools let you attach a file like this to a project, so every chat inside it starts with it.

This is what you do: you do not write the file yourself. You let AI interview you.

Interview me to build a context file about my work.
Ask one question at a time, 30 questions in total.
Cover my product, users, metrics, stakeholders, how my manager likes documents, and what I never want to see.
If my answer is vague, ask me for a real example.
At the end, put everything into context.md.

Twenty minutes later, you have this.

# context.md

## Product
Personal loans from Rs 20,000 to Rs 3 lakh. Android first.
Funnel: 1,20,000 starts a month, 38,000 completed. Worst step: step 4, income proof.
In progress: a model that reads bank statements and fills in income.

## Users
Salaried, 24 to 35, mostly tier 2 cities. Many on low-end phones and patchy data.

## Stakeholders
VP Product: wants one recommendation, not options. One page max.
Risk head: signs off on anything touching income or credit data.
Data science lead: reports accuracy on a fixed test set of 2,000 statements.

## What I reject
No personas in PRDs.
No goal without a number.
No bullet-point emails to my VP. She reads prose.

Look at the last section. It does more work than everything above it. Saying you like clear writing tells the model nothing. Saying no goal without a number changes every draft it writes.

Explain yourself once. Not every morning.

For your role: a program manager adds team owners, escalation paths and the meeting calendar. A TPM adds the system names and which team owns which service.

Skill 4: Connect AI to the apps you already use

This is what you see: Friday, 7 PM. The office is almost empty. Your friends are already at dinner, and the WhatsApp group will not stop buzzing.

You are still on the weekly program update. Jira for the onboarding and data science tickets. Three Slack channels for what Risk and payments said. The launch tracker sheet in Drive. Copy everything into Claude. Copy the summary into an email. Two and a half hours so far, and half of it was copy and paste.

This is what you learn: your context file covers what is in your head. It does not cover what is in your apps. Connectors let AI read from Jira, Slack, Drive, Gmail and your calendar directly, and in some cases act inside them. Most of them run on MCP, which we covered in the last article. Once AI can reach your tools, you stop feeding it data. You give it a goal.

This is what you do: you switch on the Jira, Slack and Google Drive connectors. Then you write the goal, not the steps.

Using Jira, Slack and the Launch Tracker sheet in Drive:
1. List every ticket in LOAN-ONB and LOAN-DS that changed status this week.
2. Find messages in #loan-risk and #loan-payments that mention a delay or a blocker.
3. Write the weekly update: what shipped, what slipped and why, top 3 risks, and asks from leadership.

The draft comes back in eight minutes. You read it, fix two lines and hit send. You still make it to dinner.

One rule here. Start with read-only access. AI connected to Slack can see every channel you can see. Let it send messages or move tickets only after you have watched it read correctly for a few weeks.

If you are copying and pasting into AI, you are still doing the AI’s job.

For your role: an engineer connects GitHub and the error logs. A data scientist connects the data warehouse, so AI runs the query instead of you pasting a CSV.

Skill 5: Automate your work without code

This is what you see: Saturday morning. Your director replies to the update, on the thread with all four team leads. This format is different from last week. Risks come first, and the asks are missing. Please keep it the same every week. I compare them side by side.

It is a small thing. It still stings because everyone saw it. And you know why it happened. You typed the prompt again from memory, a little differently this time.

This is what you learn: a task you repeat should not be a prompt you retype. It should be a skill. A skill is a saved file with the steps, the format and the rules, which the AI loads whenever that task comes up. Once a task is a skill, you can schedule it to run on its own. Neither needs any code.

Pick the task carefully. It should repeat, follow the same steps every time, and mostly not need your taste. The weekly update fits. The PRD does not.

This is what you do: on Monday, you ask Claude to turn Friday’s chat into a skill. You get a file like this.

# weekly-program-update/SKILL.md

When to use: every Friday, for the loan application program update.

Steps:
1. Pull this week's status changes from Jira projects LOAN-ONB and LOAN-DS.
2. Pull messages that mention delay, blocker or slip from #loan-risk and #loan-payments.
3. Read milestone dates from the Launch Tracker sheet.

Format, always in this order:
1. Shipped this week
2. Slipped, with the reason and the new date
3. Top 3 risks
4. Asks from leadership

Rules:
Max 300 words. Prose, no tables.
Mark the risk section DRAFT for me to edit.

Then you schedule it. Every Friday at 3 PM, run weekly-program-update and send the draft to me.

The next Friday, the draft is waiting at 3 PM. You edit only the risks. You leave the office at 6 for the first time in months. You even tell a friend that your weekly update now writes itself.

Automate the gathering. Keep the judgment. Only you know which delay will worry your VP.

For your role: a data scientist turns the weekly model metrics report into a skill. A TPM turns the release notes into one.

Skill 6: Loop engineering

This is what you see: this is the Monday from the start of this article. Your director reads out that the scanned PDF parser has shipped. Your phone buzzes. The ticket was reopened on Thursday night, after the parser failed on statements from two smaller banks.

Why did your skill miss it? It saw the ticket move to Done on Tuesday and listed it as shipped. It never checked where the ticket stands now. A skill runs each step once and trusts the first answer. Nothing in it checks the work.

This is what you learn: a skill does the task once. A loop does the task, checks it, and tries again until it passes or hits a limit. In June 2026, Boris Cherny, who created Claude Code, said on stage that he no longer prompts Claude. He writes loops that do the prompting, and writing those loops is now his job. He was talking about code, but the shape works for any task you repeat. Do the work. Check it. Retry once. Stop and hand over to a human.

This is what you do: you add a check block to the same skill, in plain English.

Check before sending:
1. For every item under Shipped, open the ticket again and confirm its status is Done right now.
2. Match every date under Slipped against the Launch Tracker sheet.
3. If anything does not match, fix it and run the check once more.
4. If it still does not match, do not send to the group. Mark the item NEEDS YOUR CHECK and send the draft only to me.

The next Friday, the check catches two tickets that were reopened on Thursday. The update flags them before anyone reads them out loud.

Notice who wrote the check. Only you know which mistake hurts most. For a weekly update, it is a wrong status in front of leadership. So that is the first thing the check looks at.

Automation without a check only helps you make mistakes faster.

For your role: an engineer’s loop is write the code, run the tests, fix, run again until they pass. A data scientist’s loop is run the eval set every night and flag any drop above 2%.

Skill 7: Learn anything with AI, on your own problem

This is what you see: the same week, data science sends a note. On their test set of 2,000 statements, the model has 96% precision and 82% coverage at a confidence threshold of 0.9. They want to lower it to 0.8. Coverage goes up to 90%. Precision drops to 93%. They need your sign-off by Wednesday.

In the meeting, you nod. Everyone nods. You are not sure anyone in the room really understood. And if you sign and it goes wrong, it is your name on it.

This is what you learn: the new default is to ask AI first, before a colleague or a search engine. But a definition is not what you need. A search for precision gives you a textbook example about cats and dogs, and you still will not know whether to sign. You need the idea explained on your own numbers, ending in your own decision. Your AI already has your context file, so it can do exactly that.

This is what you do:

Data science wants to move our income model's threshold from 0.9 to 0.8. Their note is attached.
1. Explain precision and coverage using our step 4 funnel, not a textbook example.
2. Tell me what the change does to wrong incomes and to statements sent to ops.
3. Give me 3 questions to ask the team before I sign off.
4. Then quiz me with 3 questions to check I understood.
5. For every claim, point to the line in their note you are using.

The answer is simple. At 0.9, out of 2,000 statements, the model fills in 1,640 and gets about 66 of them wrong. At 0.8, it fills in 1,800 and gets about 126 wrong. So 160 more users skip the 24-hour wait, and 60 more wrong incomes go into credit decisions.

Now it is not a vocabulary problem. It is a product decision, and you can make it.

You walk into Wednesday with three questions. What does risk do when an income is wrong? Does the test set include scanned statements from small banks, the same ones that broke the parser? How many of those 160 users would have left while waiting for ops?

The data science lead pauses and says that is a good question. Twice.

Do not learn the definition. Learn the decision.

For your role: a program manager uses this to understand a vendor contract before signing it. An engineer uses it to learn a new framework on the team’s own codebase.

Two weeks later

Two weeks ago, you typed one line into Claude and nearly sent a PRD that could belong to any company.

Today, the same Claude knows your product. It argues with you. It reads your Jira. It writes your Friday update and checks its own work before anyone else sees it.

You did not get a better tool. You got better at using the one you had.

This is exactly how we teach in the AI PM Builder Cohort. One real problem, built step by step, until it works.

About Author

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

AI Learning Resources

1. 12 Weeks LIVE AI PM Cohort (Alumni now ship AI at Google/Microsoft, etc)
2. Self-Paced AI PM Cohort Recordings (65 Hours of Content) | Use AIBUILDER20 for 20% Discount [Alumni Shipping at Google, Microsoft, Uber]
3. Quick AI PM Course (20 Hours of Content) [Rated 4.9/5 from 1000+ Product Managers from Google, Coinbase, Microsoft, etc]
4. AI Program Management Course (For Program Managers and TPMs)
5. AI Business Course (For General Management, Biz Ops & Strategy Roles)

Explained via a Real Case Study

Imagine you are a product manager at a food delivery company.

You own the customer support experience.

Monday morning, you open the dashboard.

Yesterday, your support team got 4,000 refund requests. One support executive takes about 6 minutes per request.

So that is 400 person-hours every single day. Your manager wants this to be under one minute per request using AI. Finance wants to make sure we don’t refund money we don’t owe.

You have two weeks. In these two weeks, you will come across eight terms.

You will not learn them from some glossary.

You will learn them because each one is either a problem you hit or the fix for it. So read in order. Each term solves the problem the previous one created.

1. Agents

This is what you see: day one, you ship a chatbot.

A customer writes that two items are missing from the order. The model writes a very polite reply. You watch the ticket. The chatbot cannot check the order. It cannot see the delivery photo. It cannot issue the refund. A human is still doing all of that. Handling time goes from 6 minutes to 5. You are nowhere near one minute.

This is what you learn: an agent is a model that can take actions, not just write text. You give it tools. It calls a tool, reads the result and decides what to do next. It keeps going till the job is done.

This is what you do: day two, you give the model four tools.

Read the ticket. Look up the order. Fetch the delivery photo. Issue a refund.

Now, for the same missing items ticket, it looks up the order, sees two items marked undelivered by the rider, refunds Rs 180 for those two items and writes the reply. Nobody from your team touched it.

See, a chatbot answers. An agent finishes the task. That is the whole difference.

Our next AI Product Manager & Builder Cohort is starting. For an early-bird discount, you can fill out the form here: AI PM Builder Cohort

[Alumni Shipping at Google, Microsoft, Uber]

2. Harness

This is what you see: day three, finance pings you.

The agent has refunded a full Rs 2,400 order. You open the ticket. The complaint was one cold dish. The customer asked for a full refund; the model agreed, and nothing stopped it. You go to check the logs. There are no logs.

This is what you learn: the model was not the problem. The problem was everything around the model. That everything is called the harness.

Which tools the agent gets. What the system prompt says. What limits apply. What gets logged. When a human steps in. Most of an agent’s quality lives here and not in the model. This is why you are hearing this word everywhere in 2026.

This is what you do: four changes to the harness. Any refund above Rs 500 needs human approval. Refund is capped at the value of the items the customer actually complained about. Every tool call is logged with the reason. The written refund policy goes into the system prompt.

Same model. Different harness. The Rs 2,400 mistake cannot happen again.

3. MCP (Model Context Protocol)

This is what you see: your agent needs order data, rider data and the payments system. Your engineers write custom code to connect each one. Each connection takes around a week. Day five, the rider team changes their API without telling anyone.

Your agent is down for a full day. Now you have three integrations and each one is something that can break.

MCP is a standard way for a model to connect to tools and data. Anthropic released it in late 2024, and today most major AI platforms support it. A team exposes its system once as an MCP server. After that, any agent in the company can use it, no new code needed.

You ask the payments team to build one MCP server with two tools, lookup_order and issue_refund. Your agent uses it.

Two months later, the marketing team’s agent uses the same server for order lookups. Nobody writes that connection again.

Your question to engineering changes. You stop asking whether we can build an integration. You start asking whether we have an MCP server for this.

4. Skills

The refund agent is working. Day six, your manager asks you to also handle the where is my order tickets. That is 3,000 more a day. Your system prompt is already 1,800 words of refund policy. You add the tracking rules. Now the prompt is 3,200 words. And refund accuracy drops, because the model is reading two policies for every ticket and mixing them up.

A skill is a folder of instructions, examples and scripts that the agent loads only when a task needs it. The main prompt stays short. The detail lives in the skill.

You split the prompt into files.

• skills/refunds/SKILL.md has the refund policy, 12 worked examples and the edge cases
• skills/order-status/SKILL.md has the tracking rules and the rider delay thresholds
• The main prompt is now 200 words. It only says which skill to open for which ticket type

A refund ticket loads only the refund skill. Accuracy goes back up. When policy changes, you edit one file.

5. Loop engineering

The agent is now running 7,000 times a day. You cannot read 7,000 outputs. Day nine, you pull a sample of 200 tickets.

In 6 of them the agent called lookup_order 20 times one after another because the order ID was wrong, then gave up and refunded anyway. You realise nobody ever decided when the agent should stop, when it should retry and when it should ask for help. The agent decided on its own, and badly.

That decision is loop engineering. The term was picked up in June 2026 after an essay by Addy Osmani and posts by Peter Steinberger and Boris Cherny all said the same thing. Stop prompting the agent one turn at a time.

Design the loop that runs it. A loop has four parts. Do the work. Check the work. Retry if the check fails. Stop at a limit and hand over.

This is what you do: you write the loop for one ticket.

1. The agent reads the ticket and drafts a decision with reasons
2. A second, cheaper model checks the draft against the refund policy
3. If the check fails, the agent retries once with the checker’s note
4. If it fails again, or crosses 8 tool calls, the ticket goes to a human

The 20 lookup problem is gone. You stop thinking about the prompt. You start thinking about the loop.

6. Computer use

Around 8% of tickets involve a third-party rider partner. Their delivery photos sit inside a web dashboard which has no API. Your agent cannot fetch them. So those 560 tickets a day still go to a human, and they are the slowest ones because the human has to log in and search manually.

Computer use lets a model operate a screen the way a person does. It takes a screenshot, finds the right button, clicks, types and reads whatever comes up. It works on any website or app, even the ones built with no AI in mind.

You add a computer-use step only for those 8% of tickets. The agent logs into the partner dashboard, searches the order ID, opens the photo and brings it back into the normal loop. It takes around 40 seconds instead of the 2 seconds an API call would take. It fails maybe one time in twenty, and those go to a human.

You use it because the alternative is 560 human tickets a day. Computer use is slow and less reliable than an API. It is the right choice only when there is no API.

7. GEO / AEO

Week 2, you notice a new type of ticket. Customers are asking why the refund takes 7 days. Your policy is 48 hours. You ask three of them where they got 7 days from. All three say they asked ChatGPT or Google AI mode before writing in. You try it yourself. The AI answer says 7 days. It is picking that up from some 2023 blog post, because your own help page has the policy buried in paragraph six.

GEO is Generative Engine Optimisation. AEO is Answer Engine Optimisation. In practice, both mean the same thing. SEO was about ranking in a list of links. GEO is about being the source the AI answer reads and quotes. If the AI cannot find a clear answer on your page, it will pick one up from somewhere else.

You rewrite the help page. The first line states the policy in one sentence. Each heading is a question a customer would actually type. The refund timeline is a small table, not a paragraph.

Three weeks later, the AI answer says 48 hours. The 7-day tickets stop.

8. AGI / Superintelligence

Day 13, a new model gets released. It is cheaper and scores higher on every benchmark than the one inside your agent. Your agent is not even two weeks old and the model inside it is already not the best option. And this will happen again in a few months.

AGI means artificial general intelligence, an AI that can do most cognitive work at the level of a capable human, across any field. Superintelligence means an AI that is better than the best humans at that work. Neither exists today. Every major lab says it is building towards them. Because they are racing, the model inside your agent will keep getting cheaper and better, on a schedule you do not control.

You take one design decision. Everything in sections 2 to 6 is built such that the model is a swappable part. The harness, the MCP servers, the skills and the loop do not depend on which model is running inside. When a better model comes, you change one line of config and rerun your 200 test tickets.

The model is the part that keeps changing. The system around it is what you own.

About Author

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

AI Learning Resources

1. 12 Weeks LIVE AI PM Cohort (Alumni now ship AI at Google/Microsoft, etc)
2. Self-Paced AI PM Cohort Recordings (65 Hours of Content) | Use AIBUILDER20 for 20% Discount [Alumni Shipping at Google, Microsoft, Uber]
3. Quick AI PM Course (20 Hours of Content) [ Rated 4.9/5 from 1000+ Product Managers from Google, Coinbase, Microsoft, etc]

8 New AI Terms You Should Know to Be Ahead of 99% of People was originally published in Agile Insider on Medium, where people are continuing the conversation by highlighting and responding to this story.

Our next AI Product Manager & Builder Cohort is starting — for an early-bird discount, you can fill out the form here — AI PM Builder Cohort

Imagine you want to buy running shoes.

Earlier, you would probably open Amazon, search for shoes, compare a few products, read reviews, and make a purchase.

Now, you might start with ChatGPT.

You ask it to recommend running shoes under $150. It compares your options, explains the differences, and helps you narrow down your choices.

You haven’t opened Amazon once.

This creates an interesting problem for Amazon. The shopper still intends to buy something, but Amazon may no longer be the place where the shopping journey begins.

Amazon’s response is interesting: instead of trying to control the entire AI shopping experience, it is helping advertisers reach shoppers inside ChatGPT.

Through its advertising integration with ChatGPT, selected US advertisers can extend their Amazon Ads campaigns into ChatGPT’s conversational advertising experience.

Why would Amazon choose this strategy? Why not open its product catalogue to ChatGPT, pay for referrals, or try to own checkout inside the AI experience?

To understand the decision, we need to look at the incentives of both companies.

Start with the business problem, not the deal

Let’s separate two things that are easy to confuse.

Product discovery is where a customer researches products, compares alternatives, and decides what to buy.

Transaction is where the customer actually purchases the product.

Historically, Amazon has been strong at both. Customers discover products on Amazon, click on listings, and complete their purchases there.

AI assistants create the possibility of separating these activities.

A customer might discover a product on ChatGPT, click an advertisement or recommendation, and complete the purchase on Amazon. Alternatively, they might go directly to a brand’s website.

The critical question is: who controls each part of this journey?

If ChatGPT owns product discovery but Amazon owns the transaction, Amazon may still benefit from the purchase. However, it risks losing some of the advertising opportunities associated with product discovery.

This matters because Amazon has two related businesses.

• Retail: It earns through products sold on its marketplace.
• Advertising: Brands pay to get their products in front of potential customers.

These businesses reinforce each other. More shopping activity can create more advertising opportunities, while advertising can help products get discovered.

If shoppers increasingly start their journeys elsewhere, Amazon needs a way to participate in those journeys.

It does not necessarily need to own the first interaction. It needs to preserve its ability to create value from it.

Break the business problem into four levers

When analysing a strategy decision, don’t jump directly to solutions. First identify which part of the business is threatened and which parts remain strong.

For this case, I would analyse Amazon using four levers.

Lever 1: Reach

How many relevant shopping journeys can Amazon participate in?

When customers begin on Amazon, the company has multiple opportunities to influence their choices.

When customers begin on ChatGPT, Amazon might not be considered at all.

Advertising inside ChatGPT creates another opportunity to reach these customers.

Lever 2: Monetisation

How can Amazon and its advertisers make money from that attention?

On Amazon, brands can bid for visibility through advertising formats such as Sponsored Products.

ChatGPT introduces another place where shoppers can encounter a sponsored product while researching a purchase.

Amazon’s advertising infrastructure gives existing advertisers a way to extend campaigns into this new channel.

The important distinction is that Amazon is not creating the customer’s intent. It is helping advertisers reach that intent in a different place.

Lever 3: Measurement

Can the business establish whether the advertising actually generated value?

A brand needs to know whether its spending created additional purchases, not simply whether someone clicked an advertisement.

Amazon has an advantage when a transaction happens on its marketplace because it can observe the order directly.

However, measuring purchases across external websites and AI interfaces is more complicated.

The more difficult it becomes to connect advertising spend to business outcomes, the harder it is for advertisers to justify increasing their budgets.

Lever 4: Conversion

Once a shopper decides to buy, where does the purchase happen?

Amazon has an established retail experience, including product listings, checkout, fulfilment, and Prime benefits.

If a shopper clicks an advertisement and purchases on Amazon, the company can potentially benefit from the transaction as well as the advertising activity.

If the shopper purchases directly from a brand’s website, that retail benefit may not accrue to Amazon.

This is why the destination of the advertising click matters.

The strategic objective is not simply to win more clicks. It is to reach new shopping journeys while preserving the ability to generate profitable transactions.

For this analysis, that is the outcome I would optimise for. It is an analytical framework for this case, not a claim about Amazon’s officially declared company-wide North Star metric.

What options did Amazon have?

Once we identify the four levers, the available strategies become easier to compare.

Option 1: Keep AI assistants out

Amazon could continue to restrict access to its product catalogue and avoid participating in external AI shopping experiences.

The advantage is control. Amazon retains its own shopping experience, advertising placements, and transaction flow.

The problem is reach.

Customers can still use AI assistants to research products. Restricting access does not make those shopping journeys disappear. It simply limits Amazon’s ability to participate in them.

Blocking access may make sense where the risks exceed the benefits. But it is not, by itself, a strategy for capturing demand in a channel that continues to grow.

Option 2: Open its product catalogue to ChatGPT

Amazon could make product details, prices, availability, and other information accessible to ChatGPT.

This could allow the assistant to recommend Amazon products more directly.

The advantage is better product discovery. The trade-off is control.

ChatGPT could become the interface where customers compare products and make choices, while Amazon becomes the destination they visit only when ready to purchase.

Amazon would also need to consider how external product discovery affects its own sponsored listings and advertising business.

Opening the catalogue could create more traffic, but it would not automatically preserve the economics of Amazon’s existing shopping experience.

Option 3: Pay for referrals

Amazon could compensate an AI platform for customers referred to its marketplace.

This would make the economics dependent on successful referrals or purchases, depending on the agreement.

Consider an illustrative example. If a $130 purchase carried a 4% referral commission, the fee would be $5.20.

That may be reasonable if the referral creates incremental demand. It becomes less attractive if Amazon pays for a purchase that would have happened anyway.

There is another trade-off: referral payments do not necessarily preserve the advertising opportunities associated with Amazon’s own shopping experience.

The model can work, but Amazon would need to evaluate the cost of acquiring each incremental customer against the profit generated by that customer.

Option 4: Move the entire purchase into ChatGPT

Another option is to let the AI platform handle more of the purchase itself.

In principle, this creates a simpler experience. Customers can research, decide, and buy without leaving the conversation.

However, there is a difference between making a good recommendation and becoming a trusted transaction platform.

Research cited by EMARKETER found that 10% of AI-assisted digital shoppers usually completed purchases on the AI platform itself, compared with 69% who usually purchased through a retailer’s website or app.

This is a useful signal, although it should not be treated as proof that in-chat checkout cannot succeed in the future.

The strategic question is whether Amazon should give up control of a transaction experience it already knows how to operate before customers have clearly demonstrated a preference for buying through the AI platform.

Option 5: Sell advertising inside ChatGPT

This is the approach Amazon and OpenAI are testing.

Rather than immediately transferring product discovery or checkout, Amazon enables its advertisers to extend their campaigns into ChatGPT.

Amazon gains access to a new discovery environment through its advertising relationships. OpenAI gets another way to monetise the ChatGPT experience.

And if an advertisement sends the customer to Amazon, Amazon has the opportunity to retain the transaction as well.

This option does not eliminate every risk. Amazon does not control the ChatGPT interface, and not every advertising click will produce an Amazon sale.

But it allows Amazon to participate in a new channel without first rebuilding its entire shopping experience around that channel.

That is the central strategic trade-off: enter a new environment without giving away control of the parts of the business that already work.

Why the advertising model makes sense for both companies

The deal becomes clearer when we look at the incentives on each side.

What Amazon gets

First, it can reach customers who begin their research outside Amazon.

Second, existing advertisers get a way to extend their media campaigns into a conversational environment without treating it as an entirely separate advertising channel.

Third, Amazon can explore the economics of AI-driven product discovery before making a larger commitment to a different commerce model.

But there is an important condition.

Amazon benefits most when advertising in ChatGPT creates incremental value, especially when it leads to profitable transactions on Amazon.

If a brand pays to reach someone in ChatGPT, but the person buys directly from the brand, the advertising may still benefit the brand and the advertising channel. Amazon’s retail business, however, may not gain the same value.

So Amazon should not evaluate this initiative only by the volume of advertising spend. It should also understand where the resulting purchases occur and whether those purchases would otherwise have happened.

What OpenAI gets

ChatGPT has to balance monetisation with the experience that brings users back.

Its published advertising principles state that advertisements are clearly labelled and do not influence the answers ChatGPT provides. OpenAI also says conversations are not shared with advertisers, who receive aggregated information about ad performance.

These boundaries matter because users come to ChatGPT for help, not simply to browse advertisements.

OpenAI needs advertising revenue, but it cannot treat every conversation as an ordinary search-engine results page without risking the trust that makes the product useful.

Amazon brings an established advertising business and relationships with brands that already spend money on digital marketing.

That lowers the friction involved in getting advertisers to test a new channel.

The deal therefore fits both incentives: Amazon gains access to a new environment for its advertisers, and OpenAI gains a route to advertising demand without having to build every advertiser relationship from scratch.

What advertisers get

Advertisers get another opportunity to reach people while they are actively researching a decision.

A person asking for the best running shoes under $150 has expressed a more specific need than someone casually browsing a general website.

That does not guarantee the person will purchase. But it creates a potentially valuable context for advertising.

Advertisers still need to test whether the channel delivers incremental sales at an acceptable cost. Intent alone is not enough.

Follow the money with one simple example

Suppose Emily is shopping for running shoes that cost $130.

There are two hypothetical ways to monetise her journey.

Model A: Referral commission

If an affiliate agreement pays 4% on a completed order, a $130 purchase generates a $5.20 referral fee.

The payment depends on a qualifying transaction under that hypothetical agreement.

Model B: Advertising

Suppose a shoe brand pays $2 for a click on its advertisement. The brand pays for the click under a cost-per-click arrangement even if Emily does not eventually purchase the shoes.

Both amounts are illustrative; neither represents a disclosed term of the Amazon–OpenAI integration.

The economics are not directly comparable. One amount is tied to an order, while the other is tied to a click.

But the difference reveals something important about the business models.

In the advertising model, the brand funds the acquisition of attention. In a referral model, the business paying the commission funds that acquisition in exchange for a share of the resulting transaction economics.

Amazon can potentially earn advertising-related revenue through the media arrangement and retail profit if the purchase happens on Amazon. Those benefits depend on the actual campaign economics and destination of the sale.

The larger lesson is this: a company should understand not just how much revenue a new channel creates, but who pays for the customer acquisition and who captures the value after the click.

Who has the stronger strategic position?

My bet is that Amazon has the stronger immediate position, although the outcome is not guaranteed.

Here is why.

Amazon already has an established commerce business, a large advertiser base, and a way for brands to buy advertising. The integration lets advertisers test a new discovery channel without requiring Amazon to replace its own marketplace with ChatGPT.

OpenAI gains a monetisation route and access to advertisers through an established ecosystem. However, relying on external advertising partners could limit how much control it has over advertiser relationships in the long run.

The incentives are different:

• Amazon wants to participate in AI-driven discovery while protecting its retail economics.
• OpenAI wants to monetise discovery while preserving the user experience that makes people trust its answers.
• Advertisers want incremental customers, not merely another place to spend money.
• Customers want useful recommendations without feeling that the recommendations have been distorted by advertising.

These goals overlap, but they are not identical.

Amazon’s advantage weakens if the ads mainly direct shoppers to competitors’ websites, if advertisers do not see sufficient returns, or if ChatGPT becomes less dependent on Amazon’s advertising infrastructure.

OpenAI’s position strengthens if it builds its own direct advertiser relationships, develops effective measurement, and demonstrates that its advertising experience can scale without undermining trust.

The real outcome will depend on how the economics evolve, not on the existence of a partnership alone.

What should we watch to know whether this strategy is working?

A Product Manager analysing this launch should go beyond impressions and clicks.

I would monitor five things.

1. Incremental reach

How many relevant shoppers does ChatGPT bring into Amazon’s ecosystem who would not otherwise have entered it?

A new channel is valuable only if it creates additional opportunities or improves the economics of existing ones.

2. Incremental purchases

Do exposed shoppers make more profitable purchases than comparable shoppers who were not exposed to the advertising?

This is more informative than simply counting orders attributed to clicks. Some customers would have purchased from Amazon anyway.

A properly designed experiment or holdout group can help distinguish incremental sales from sales that would have happened regardless.

3. Advertising efficiency

Do advertisers get an acceptable return on their spending?

Track cost per incremental purchase, conversion rate, and return on advertising spend where the measurement setup supports them.

A high click-through rate is not enough if the traffic does not convert.

4. Purchase destination

Do shoppers go to Amazon product pages, or do they complete purchases on brand websites?

This helps establish whether the new channel strengthens Amazon’s retail business, primarily benefits its advertising operation, or creates value elsewhere.

5. User trust and experience

Do users continue to find ChatGPT helpful when sponsored placements appear?

If advertising makes recommendations seem less trustworthy, short-term revenue could come at the expense of long-term product usage.

The test is not simply whether ChatGPT ads get clicked. It is whether the advertising model can create durable value for advertisers and the platform without weakening the product experience.

Learn to break down AI and product strategy

In the AI Product Builder Cohort, we break down real product and AI strategy decisions using frameworks like these: business outcomes, key levers, strategic options, trade-offs, and experiments. Members also build practical AI products themselves.

About Author

Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.

AI Learning Resources

1. 12 Weeks LIVE AI PM Cohort (Alumni now ship AI at Google/Microsoft, etc)
2. Self-Paced AI PM Cohort Recordings (65 Hours of Content) | Use AIBUILDER20 for 20% Discount
3. Quick AI PM Course (20 Hours of Content)

PM interviews have changed

Why are experienced PMs getting stuck in PM interviews they would have cleared two years ago?

Because the interviews have changed.

The old rounds are still there.

Product sense, metrics, estimation and behavioural. But many PM interviews now add an AI layer on top.

That layer does not ask you what RAG is or what an agent is. It asks what you would do when one breaks.

That is the gap I talk about in the video above. Many PMs have learned AI. Far fewer have built with it. The new questions are written to tell the two apart.

Below are the four types of AI questions we see most. For each one, I show what the question looks like, what it takes to pass, and why building helps.

1. Take-home assignments

Some companies now send a take-home assignment. Here is a real one for a Product Manager, Platform Experience role. The problem was customer questions at scale.

• You get 4 to 6 hours. If you reach hour seven, you stop and send what you have.
• You build a small working prototype of the first thing you would build. Any tool is allowed, including AI tools. A rough script over the sample data, a clickable page or a working demo all count.
• You write a PDF of up to 4 pages or 8 slides. It covers how you framed the problem, what you chose to build and not build, and how you would know it is working.
• You have five days to send it.
• Then there is a 45-minute call where they push back on your choices.

The brief says they judge what the prototype shows, not how it is engineered. It also says a prototype that answers one question well and correctly refuses one it shouldn’t beats a polished mock of everything.

To pass, you need to do five things.

• Pick one narrow problem and make it work on the sample data.
• Make it refuse questions it cannot answer, instead of making something up.
• Write down what you left out, and why.
• Say how you would measure it, with a few test questions and one or two numbers.
• Defend every choice in the follow-up call.

This is where building helps. If you have built AI products before, you know how much fits in 6 hours, so you keep the scope small. You know an AI tool will answer anything with confidence, so you add the refusal on purpose. Someone who has only learned AI often spends the 6 hours on a polished mock. The brief tells you that it loses.

2. RAG questions

RAG is when a chatbot first fetches the right pieces of your documents and then answers from them. It is the AI topic we see asked the most. It rarely comes as a question about what RAG is. It comes like this.

Chunking

Our chatbot answers from a 200-page policy document. The answers are often half right. What would you change?

To pass, talk about how the document is cut into pieces, called chunks. Small chunks lose context. Big chunks bring in text that has nothing to do with the question. Split by headings and sections, not by a fixed length. Keep tables in one piece. Let chunks overlap a little, so a sentence cut at the edge still appears whole in one of them.

Reranking

The right chunk is found, but it comes back eighth, and the model only sees the top three. What do you do?

To pass, explain reranking. First, fetch more chunks, say 30, with fast search. Then a second model scores each chunk against the question again, and only the best 3 to 5 go to the model. Also mention the cost. Every answer gets a little slower.

Root cause

Our support chatbot started giving wrong answers last week. Find out why.

To pass, split the problem into two. Take 20 wrong answers and check if the right chunk was fetched. If it was not, the problem is retrieval, so look at chunking, search and reranking. If it was fetched and the answer is still wrong, the problem is generation, so look at the prompt and the model. Then ask what changed last week. It is usually the documents, the prompt or the model.

Coverage

You have a Harry Potter book, and Snape is mentioned 400 to 500 times. Sometimes by name, sometimes just as he, sometimes as the Half-Blood Prince. How does your chatbot use every mention when it answers a question about him?

To pass, first spot the problem. Normal search fetches only the top few chunks, and it misses the chunks where Snape is never named. Then fix the index. Work out which character every pronoun and nickname points to, and tag each chunk with the characters in it. Add a short line of context to each chunk before storing it. For questions about the whole book, answer from chapter summaries instead of hundreds of chunks.

This is where building helps. When you build RAG, you watch a table get cut in half, and the answer go wrong. You see the right chunk come back eighth. You change the chunk size and see the answers change. In the interview, you talk about what you saw, not a definition.

3. AI agent questions

An AI agent is a model that decides its own next step and uses tools, like search or email, to finish a task. Agent questions look like this.

Agent or workflow

Would you build this as an agent or as a fixed workflow?

To pass, say that a fixed workflow is better when the steps are known in advance. It is cheaper, faster and easier to test. Use an agent only when the next step depends on what the user says or what a tool returns. Good answers start with the simpler option.

Handoffs between agents

Your system passes a task from one agent to the next. How do you know the handoffs are working?

To pass, name the metrics. What share of tasks finish end to end. At which step tasks fail the most. Whether the next agent got everything it needed. How many tasks end up with a human. Time and cost per task.

Stuck agents

Your agent keeps calling the same tool again and again. What do you do?

To pass, start with the trace, which is the step-by-step log of what the agent did. Then set a step limit and a clear stop condition. Write better tool descriptions, so the agent knows when each tool is the right one.

Human approval

What should the agent never do without asking a person?

To pass, point to actions that cannot be undone. Refunds, payments, emails to customers and deleting data. Those wait for a human to approve.

This is where building helps. Once you build an agent, you see it loop, pick the wrong tool, or pass half the information to the next step. You learn to read traces. That is exactly what these questions ask about.

4. AI product sense and metrics questions

The classic product sense round now has an AI version.

Should this be AI at all?

Design an AI feature for a food delivery app.

To pass, start with the user problem, not the model. Say when a simple rule would do the job. Then cover what AI adds, what each answer costs to run, and what the user sees when the AI is wrong.

Launch decisions

Your model is right 92% of the time. Do you launch?

To pass, ask what the other 8% looks like. A bad song suggestion is fine. A wrong refund amount is not. Group the failures by type, decide what a wrong answer costs the user, and set the bar from that.

After launch

How would you know it is working?

To pass, use two kinds of checks. Before launch, build a test set of real questions with the right answers marked, including questions it should refuse. After launch, watch how users react. How often they accept the answer, give a thumbs down, or ask for a human.

This is where building helps. When you build, you write those test questions yourself. You watch a bot you thought was good fail more of them than you expected. After that, you stop treating one accuracy number as the answer.

How to prepare for the new PM interviews

Look at what every question above has in common. None of them asks for a definition. Each one asks what you would do when something breaks, and why. You can only answer that well if you have seen it break.

That is why we built the AI Product Builder Cohort at Technomanagers around one objective. To make you job-ready to build AI products.

• 12 weeks of real projects across RAG, AI agents, evals, prototyping and AI product strategy
• 10+ projects you build yourself, with mentor feedback
• One portfolio-ready AI product as your capstone
• 20 hours of AI PM interview prep, including mock interviews on questions like the ones above

Our 1st cohort ended with 10 job offers, 4 of them Senior AI Product Manager roles.

This cohort is not for everyone. If you want a certificate, it is the wrong program. It asks you to build every week. That is why there is an application.

We have started taking registrations for our 3rd Cohort.

Currently, we are running an Early Bird Discount for our 3rd Cohort — Fill the Form if you are interested

AI Learning Resources

1. 12 Weeks LIVE AI PM Cohort (Alumni now ship AI at Google/Microsoft, etc)
2. Self-Paced AI PM Cohort Recordings (65 Hours of Content) | Use AIBUILDER20 for a 20% Discount