The same message lands in my inbox about once a week now, in slightly different words each time.

Someone shipped more this year than in any year of their career. Reports that took two days take two hours. They cleared the backlog, covered a vacant role, made the team look good. Then the review came back: meeting expectations. No promotion.

And the conclusion they draw, almost every time, is that AI quietly removed the rung before they could reach it.

I understand why that lands. It is also, based on what the data actually shows and what I watch happen inside real organisations, backwards. Your promotion did not disappear. The evidence you used to earn it did. Those are very different problems, and only one of them is yours to fix.

Two numbers, and almost nobody quotes the second one

The frightening number is real, so let’s start there. Stanford’s Digital Economy Lab, working with ADP payroll records covering millions of US workers, found that workers aged 22 to 25 in the occupations most exposed to AI are running about 19% below where their employment levels should be. That gap has widened steadily since August 2025, and it is happening through slowed hiring rather than layoffs, which is why it took so long for anyone to notice.

The mechanism behind it is not mysterious. In ZipRecruiter’s 2026 AI Employer Report, surveyed in June across more than 1,000 hiring managers, 38% said they had moved basic data processing away from entry-level workers and onto AI, and 31% said they had raised the experience requirements for entry-level roles as a direct result. The bottom rung did not vanish. It got moved up out of reach.

Now the number that rarely appears in the same article. PwC’s 2026 Global AI Jobs Barometer found that entry-level roles in the occupations most exposed to AI grew 35% since 2019, while entry-level roles in the least exposed occupations declined 10%. The same report puts the wage premium for AI skills at 62%, up from 57% the year before. And Randstad, analysing more than 35 million job postings globally, reported in June that professionals with demonstrated AI fluency are securing promotions roughly three and a half times faster than their peers.

These datasets are not measuring the same thing, and I want to be careful about that. One counts people on payroll, the others count openings, requirements and postings. You cannot cleanly stack them. But they disagree in a direction that matters: the same wave thinning one group is visibly accelerating another, inside the same occupations, at the same time.

Something is sorting people. It is not sorting them by how much work they produce.

What a promotion was actually buying

Here is the thing that gets lost when we talk about promotions as rewards.

A promotion is not payment for last year’s work. It is a bet on your judgment at a larger scope. Your manager is trying to answer one question: if I hand this person more surface area, more budget, more people, more risk, will their decisions hold up when I am not in the room?

Judgment is almost impossible to observe directly. So organisations built proxies for it, and we all learned to produce them. A clean, well-structured analysis. A deck that anticipated the follow-up question. The person who turned things around fast without dropping quality. None of those things were judgment. They were the residue judgment left behind, and for thirty years they were reliable enough to promote on.

The companies themselves are now saying the criteria have changed. HiBob’s AI Skills 2026 Report, a July survey of 1,200 decision-makers, found that 67% of organisations already link AI proficiency to promotions, 50% tie it to performance ratings, and 28% link it directly to pay.

Sit with that for a second. Two thirds of companies have written AI proficiency into their promotion criteria. Almost none of them have a defensible way to measure it, because the obvious measure, better output faster, is exactly the measure AI broke.

The proxy collapsed

Two years ago, a tight synthesis of a messy 40-page document at 11pm told your director something real about you. It signalled that you could hold complexity, decide what mattered, and stand behind a cut. That is why it moved your career.

Today, that same artifact tells your director that you have a Claude subscription.

I call this the proxy collapse, and it is the single most useful frame I have for what people are experiencing. The signals we spent our careers learning to emit are still being emitted. They have simply stopped carrying information, because the cost of producing them fell to near zero and did so for everyone at once. When a signal becomes free, it stops sorting anybody.

Your promotion did not disappear. The evidence you used to earn it did.

At the same time, the number of seats above you is being deliberately reduced. Amazon told its leadership teams to increase the ratio of individual contributors to managers by at least 15%, and plenty of large employers have followed the same logic since. Fewer management slots, each covering more ground. So the bar is not only being measured differently, it is being measured for a scarcer and larger job.

This explains the specific shape of the frustration I keep hearing, which is not “I am being replaced.” It is “I did more and it counted for less.” Both halves of that sentence are true, and they are not a contradiction. More output, less signal.

Four datasets, one answer

Once you see promotion as a judgment bet rather than an output reward, the newer data stops looking contradictory and starts looking obvious. Four independent sources, using four different methods, are pointing at the same replacement criteria.

PwC found that entry-level roles in the most AI-exposed occupations are now seven times more likely to demand skills that used to appear only in senior job descriptions: judgment, creativity and leadership. Read that as a hiring manager rather than as a candidate. Those were promotion criteria. They are becoming entry criteria.

Subscribe now

ZipRecruiter’s employers said the same thing when asked what matters more now than it used to: critical thinking, 65%. Judgment and decision-making, 59%. Creativity, 58%. Randstad’s posting analysis found demand for emotional intelligence up 173% and creativity up 168% over the past year, while traditional customer service roles contracted 33% and AI-integrated versions of those same roles grew 565%.

And the Stanford work points at the mechanism from the other side. The declines are concentrated in roles where AI substitutes for the human task. Where AI complements the work, employment is flat or rising. Which side of that line your role sits on is usually reported as fate, something decided by your job title. In my experience building AI products and teaching this to professionals, it is much closer to a design decision. Two people with identical titles can spend the year very differently: one racing the model at its own game, one designing systems the model runs inside. By December those two people are not in the same labour market.

How to make your judgment visible before your next review

This is the part you can act on, and none of it requires a new tool.

Stop shipping tasks. Start owning one system. Pick a single recurring workflow in your patch that costs your team real hours, design an AI workflow that carries it end to end, and put your name on it. One named system with a before and after number does more for a promotion case than a year of individually excellent documents, because a system is evidence of design judgment and a document is now evidence of nothing.

Keep a decision log, not an output log. Most people preparing for review write down what they produced. Write down what you decided instead: the recommendation you made, the option you rejected and why, the point where you overruled the model because you knew something it did not. That last category is the most valuable page in your file. It is the direct record of the thing your manager cannot observe and is trying to bet on.

Build an AI portfolio a stranger could evaluate. Three artifacts is enough. The system you designed, with the metric it moved. A decision you got right that the data did not obviously support. A workflow you deliberately left human, with your reasoning for where the checkpoint belongs. In a market where 67% of companies say AI proficiency affects promotion and almost nobody can define it, being the person who can show it is an unreasonable advantage.

The part I would want you to take away

AI is not taking promotions and it is not handing them out either. Both framings put you in the passenger seat of your own career.

What it did was destroy a measurement instrument that most of us never noticed we were relying on, and organisations have written new criteria they do not yet know how to assess. That gap is uncomfortable, and it is also the entire opportunity. For the next few review cycles, the people who make their judgment legible will be visible in a way that is almost impossible to be visible in a normal year, because everybody else is still busy producing signals that no longer signal anything.

The work was never what got you promoted. It was just the last reliable place your judgment showed up. Find the new one.

Get Free Invite to my 3 Live Sessions for Becoming the Go-To -AI Leader of your workplace and show your AI superpower while learning how to go from zero to hero using AI.

Stop figuring things out on your own and follow this step-by-step free course.

Reserve your Free Seat Here

PS. If you want to build that system with structure instead of guessing at it, that is most of what we do inside the AI SuperPractitioner Certification.

Note 1

Two years ago, turning a messy 40-page document into a tight synthesis by 11pm told your director something real about you.

It signalled you could hold complexity, decide what mattered, and stand behind a cut.

Today the same artifact tells your director that you have a Claude subscription.

The signal is still being sent. It just stopped carrying information, because it became free for everyone at the same time. That is the whole story behind “I did three times the work this year and got meets expectations.”

Note 2

67% of organisations now link AI proficiency to promotions. 50% link it to performance ratings. (HiBob, July 2026, 1,200 decision-makers.)

Now ask the harder question: how are they measuring it?

The obvious measure, better output faster, is exactly the measure AI broke. Two thirds of companies have a criterion they cannot yet assess.

That gap is the opening. Anyone who can show their judgment directly, rather than implying it through polished output, is going to be very visible for about two review cycles.

Note 3

Four different datasets, four different methods, one answer about what gets rewarded now:

PwC: entry-level roles in the most AI-exposed occupations are 7x more likely to require judgment, creativity and leadership. ZipRecruiter: employers rate critical thinking (65%), judgment (59%) and creativity (58%) as more important than before. Randstad: demand for emotional intelligence up 173%, creativity up 168%. Stanford: employment holds where AI complements the work, falls where it substitutes.

Those first three used to be promotion criteria. They are becoming entry criteria. The ladder did not get shorter at the top. It got steeper at the bottom.

Leave a comment

I used it twice. Then I quietly went back to my old spreadsheet.

Nobody asked why. I didn’t volunteer it. On paper, the tool worked fine, it did what it said it would do. But I didn’t trust it enough to stop checking my own version behind it. Which meant I wasn’t saving an hour. I was spending extra time doing the work twice.

I think about that a lot now, because I hit the same wall when I started building my own AI workflows.

Subscribe now

The problem: you're optimizing for "it worked once"

When you start building with AI instead of just chatting with it: getting an impressive output once is easy. Getting something you'd actually stake your name on, every single time, without re-checking it, is the real skill.

Its easy to stop at the first win. Claude drafts a great customer email, or summarizes a messy call transcript perfectly, and you think: this is it, I’ve got my AI workflow. Then three uses later it gets something subtly wrong, and you quietly go back to doing it manually, exactly like I did with that tool.

If you never figure out why it broke, you will stop trusting it and chalk it up to

“AI isn’t reliable enough yet.”

That’s the wrong conclusion. The tool wasn’t broken. The workflow around it was never built to be trustworthy in the first place.

The shift: You're building a system you can rely on.

This is the difference between using AI and building with it. Using AI is asking a question and hoping for a good answer. Building with AI means setting up the conditions so a good answer is the reliable outcome — not the lucky one.

Here’s what that actually requires, and none of it involves code:

• You can see the reasoning, not just the output. If Claude gives you a number or a conclusion with no “why” attached, you’re not checking it — you’re gambling on it.

  Action - Ask it to show its work every time.

• It tells you when it’s guessing. A workflow you can trust is one where you’ve explicitly instructed it: flag anything you’re not confident about, instead of presenting a guess as fact.

  Action - Instruct it to flag anything its not confident about

• It fails loud, not silent. If you feed it a transcript with a chunk missing, does it tell you, or does it quietly fill the gap with something plausible? You have to test for this on purpose.

  Action - Test if it stops when it fails or makes ‘wrong’ assumptions to keep going

• You stay the checkpoint on anything that goes out the door. Claude drafts. You send. That’s not a limitation — that’s the actual design.

  Action - Build Human-in-the-loop workflows

• It’s consistent. Same input, same structure of output, every time. If your workflow gives you a wildly different format every session, it’s not a workflow — it’s a slot machine.

  Action - Instruct it for consistent output

What this looked like when I rebuilt it

The tool I abandoned years ago failed on almost every point above. No reasoning shown. No uncertainty flagged, ever. No visible failure state — wrong answers looked exactly as confident as right ones.

When I started building my own workflows in Claude, I applied the opposite standard. For a weekly reporting summary I used to build by hand, I set it up so Claude always cites which rows or data points it pulled from, flags anything ambiguous instead of smoothing over it, and outputs the same structure every single time — so I check the flags, not the whole document. That one change took it from “cute demo” to something I actually use every week without redoing my old process behind it.

Same underlying tool. Completely different trust level. The difference wasn’t the AI getting smarter. It was me building the workflow smarter.

Leave a comment

Try this on the next thing you build

Before you consider a Claude workflow “done,” ask yourself one question: would I stop double-checking this, or am I still secretly redoing it myself?

If the honest answer is “I’m still redoing it,” you haven’t built a workflow yet — you’ve built a demo. The fix usually isn’t a better prompt. It’s adding one of the five checks above.

This is the exact gap between people who chat with AI occasionally and people who become the person their team turns to for it — and it’s the core of what I’m building into the course I’m putting together for exactly this audience: non-technical people who want to build AI systems they can actually rely on, not just impressive one-off outputs. If you want a head start, my free Lightning Lessons run Saturdays at 9am PST on Maven — link in my profile.

What’s an AI output you stopped trusting — and did you ever figure out why?

Subscribe to AI First Academy

AI Product Leader building AI-native products and helping professionals thrive in the AI-first era. From non-technical PM to AI leader — sharing insights on AI, agentic systems, product growth, and career leverage.

Subscribe now

My Slack had a message waiting for me this morning.

It listed my meetings, flagged the three emails I actually needed to respond to, told me a document had been shared that needed sign-off before my 9am call, and recommended the one thing I should focus on first.

I didn’t write it. I didn’t schedule it. I didn’t even ask for it.

Claude did. While I was asleep.

Subscribe now

Here’s the thing about AI Agents — the name makes them sound like something built by a team of engineers in a lab somewhere. Something complicated. Something not for you.

That framing is wrong. And today, 4,000 non-technical professionals proved it.

The difference most people miss

Most people use AI the same way they use Google.

Type something in. Get something out. Move on.

That’s not wrong — but it’s leaving most of the value on the table. Because AI can do something much more powerful than answer your questions. It can pursue your goals.

Here’s the simplest way I know to explain the difference:

Automation is: “Here’s the essay. Post it on Substack at 8am Wednesday.” The machine does what you tell it. No thinking involved.

AI Automation is: “Write an essay on this topic. Post it at 8am Wednesday.” Claude writes it. But you’re still defining every step.

AI Agent is: “Goal: grow my Substack audience. Topic: X.” Claude plans the approach. Researches what to write. Drafts the essay. Shows it to you for approval. Posts it at the best time based on your audience data.

You didn’t manage the steps. You set the destination.

That’s the shift. And once you see it, you can’t unsee it.

What an AI Agent actually has

An AI Agent isn’t magic. It’s a structured system with four components working together:

• Memory — it knows your context, your role, your preferences

• Planning — it breaks your goal into steps and reasons through them

• Tools — it can access your calendar, email, documents, apps

• Action — it executes, checks in where needed, and keeps going

Put those four things together and you don’t have a chatbot anymore. You have a colleague who works through the night and never needs a coffee break.

How we built one in 45 minutes

In today’s session, we used Claude Cowork to build a Morning Brief from scratch.

Cowork is Claude’s interface for agentic work — the place where you stop prompting and start delegating. It has five zones: a Goal field, Connectors for your tools, Instructions for your rules, a Schedule for when it runs, and an Output for where the result lands.

Here’s what we built:

The goal: Every weekday before 7am, read my Gmail, Google Calendar, and Google Drive. Produce a structured one-page Morning Brief and post it to Slack.

The instructions: Here’s the format — meetings with conflicts flagged, top 3 emails that need a reply, documents shared in the last 24 hours, and one focus recommendation based on everything above.

The schedule: Every weekday. 6am. Runs without me.

The output: Slack. Waiting when I wake up.

Total build time: under 20 minutes. No code. No IT ticket. No waiting.

The first time it ran, the room went quiet. Not because it was technically impressive. Because it was immediately, obviously useful.

What actually changes

The Morning Brief isn’t just a productivity hack. It’s a different relationship with your working day.

You stop starting from zero every morning — scanning emails, rebuilding context, trying to remember what mattered. You start from a briefed position. Like you already had a 10-minute conversation with a brilliant assistant before you even opened your laptop.

That’s what an AI Agent gives you. Not speed. Clarity.

If you’ve come this far, leave a comment and let me know how it went

Leave a comment

What’s next

In Lesson 3 we go further.

MCP — Model Context Protocol — and multi-agent workflows. This is where Claude stops working alone and starts coordinating with other agents across your tools and systems. The closest thing to having an AI team running in the background.

If you missed today’s session, it’ll be available as a replay shortly.

And if you’re ready to go beyond the free series — the AI Super Practitioner Certification is a 4-week programe where we build your full AI Operating System from the ground up. Live sessions, hands-on builds, and a system designed around your actual role.

No code. No slides. Just systems you ship and scale.

— Keyuri

You’re in the middle of a solid Claude session. You’ve uploaded your files, given context, asked your questions. The answers have been sharp.

Then somewhere around the 45-minute mark, something shifts.

Claude starts repeating itself. It forgets something you told it three messages ago. The quality drops noticeably. You blame your prompt. You rephrase. You get frustrated.

This isn’t a prompt problem.

You hit Claude’s memory limit. And nobody told you it existed.

Subscribe now

First: What Is a Token?

A token is the unit Claude uses to read and process everything — your words, your files, its own responses. You don’t see tokens. But they’re being counted in the background every single second of your session.

Here’s the number that matters: 1,000 tokens = roughly 750 words, or about 1.5 pages of text.

On Claude Pro ($20/month), you’re working with a context window of up to 200,000 tokens per conversation. That’s the equivalent of a 300-page book sitting in Claude’s working memory.

Sounds like a lot. It disappears faster than you think.

PS: This newsletter mostly grows from your shares. It’s free & helps me stay laser focused.

Share

Next: What is Context?

Imagine Claude is a brilliant chef. Your context window is the countertop in their kitchen.

Every ingredient they need to cook your meal has to sit on that counter — your instructions, the files you uploaded, the conversation so far, Claude’s own responses. The counter is big, but it’s not infinite.

Here’s the critical part: the counter never gets cleared mid-meal. Everything you’ve discussed, every file Claude has read, every response it has given — it all stays on the counter, taking up space.

As the counter fills up, the chef starts pushing the oldest ingredients off the edge to make room for new ones. They don’t tell you they’ve done this. They just quietly stop having access to what fell off.

That’s when you notice Claude “forgetting” things. It hasn’t malfunctioned. It just ran out of counter space.

Then: Context Rot — The Thing Nobody Warns You About

There’s a subtler version of this problem that’s even more common.

You don’t always hit the hard limit. Instead, as your conversation gets longer, Claude starts to lose the thread — answers get vaguer, it misses nuance it caught earlier, it starts hedging where it used to be precise.

This is context rot. The window isn’t full yet, but it’s cluttered enough that the most important instructions — the ones you gave at the start — are now buried under thousands of tokens of back-and-forth. Claude can technically still see them, but they’re no longer getting the attention they deserve.

Think of it like a whiteboard that’s been written on so many times you can barely read the original notes underneath.

Context rot is why a session that started brilliantly feels mediocre by the end. The model didn’t change. The signal-to-noise ratio did.

Now: Lets Fix it. 5 Habits That Keep Claude Sharp

1. Convert Files Before You Upload Them

PDFs are one of the biggest token drains most people never think about. When Claude reads a PDF, it processes not just the words but all the invisible formatting — headers, spacing, layout code, metadata. That overhead consumes tokens without adding any value to your task.

💡 The Fix: convert your PDF to a plain text or .md file before uploading.

The simplest way to do this without any technical tools:

1. Open your PDF

2. Select all the text (Ctrl+A on Windows, Cmd+A on Mac)

3. Copy it

4. Open a free tool like Notion, Google Docs, or even Notepad

5. Paste the text in

6. Copy that clean text and paste it directly into your Claude message — no file upload needed

If you’d rather keep it as a file, use CloudConvert.com — it’s free, works in your browser, no account required. Upload your PDF, convert to .txt or .md, download, upload to Claude.

Same information. A fraction of the token cost.

2. Keep Your Prompts Short — Let Claude Ask the Questions

Most people write long prompts trying to pre-answer every possible question Claude might have. This burns tokens before the real work starts — and often still misses what Claude actually needs.

💡 The Fix : Flip the approach.
Use this structure: I want to do [task] to achieve [success criteria]. Ask me questions before you start.

That’s it. 30 words or fewer.

Claude will ask you exactly what it needs to know. You answer only those questions. No tokens wasted on information Claude never needed. No missing context because you guessed wrong about what to include.

This also produces sharper output — because Claude’s questions reveal what actually matters for the task, and your answers are targeted rather than rambling.

Thanks for reading! This post is public so feel free to share it.

Share

3. Edit Your Last Prompt, instead of sending a Follow-Up.

This is the most underused token-saving habit in Claude Desktop.

Every follow-up message adds to the conversation history — which adds to the token count. If you sent something unclear or forgot a detail, the instinct is to just send another message correcting it.

💡 The Fix : Don’t. Edit the original instead.

In Claude Desktop: hover over your previous message → click the pencil icon → make your edit → resend.

Claude treats it as if the original message always said that. The correction never gets added to the pile. Your context stays lean.

4. When the Session Degrades — Start a New Chat

This is the simplest fix and the one most people resist.

When you notice Claude getting vaguer, repeating itself, or losing the thread — that’s your signal. The counter is cluttered. The right move is to open a new chat, paste in only the context that’s relevant to your current task, and start fresh with a clean, full 200,000-token window.

💡 The Fix : One task, one chat. When the job is done or the quality drops, close it out and start over.

The new chat IS the fix. Everything else on this list helps you delay needing it. But when the time comes, don’t fight it.

If the idea of starting over makes you nervous — this is for you.

The reason most people avoid starting a new chat is the same reason they keep a cluttered desk: the cost of setting up feels higher than the cost of working in the mess. Re-explaining your role, your context, your preferences — it takes time and it breaks the flow.

Claude Projects removes that cost entirely.

A Project is a persistent space in Claude Desktop where you store your standing instructions once. Every new chat you open inside that Project automatically inherits everything you’ve saved — your role, your company context, your preferences, your reference files — without you typing a single word.

How to set it up (once):

1. In Claude Desktop, click “Projects” in the left sidebar

   

2. Create a Project for your role or recurring workflow — “CS Leadership,” “Weekly Reporting,” “My Role”
   

   

   

3. Click “Project Instructions” and write your core context — who you are, what you’re working on, how you like Claude to respond

   

4. Upload any files that apply across your work — frameworks, templates, org charts

That’s it. Now every new chat inside that Project opens with Claude already fully briefed. You get a clean window and a model that already knows you.

Fresh chat. Zero repetition. No excuse not to start over.

If you’ve come this far, leave a comment and let me know how it went

Leave a comment

5. Match the Model to the Task

Not every task needs the most powerful model. Using Opus for a quick question is like taking a helicopter to cross the road — it works, but it burns through your usage allowance faster than necessary, and you’ll hit limits sooner in the day.

Here’s a simple rule of thumb:

TaskModel to reach for

Quick question, fast lookup - Haiku

Drafting emails, summarising notes - Sonnet

Analyzing files, writing reports - Sonnet or Opus

Complex reasoning, high-stakes decisions- Opus

Switch models from the dropdown at the bottom of the chat window before you send your first message. Takes two seconds. Saves your Opus quota for when it actually matters.

The One Thing to Remember

Claude’s context window is your most valuable resource in any session. Every token you spend on unnecessary file formatting, follow-up corrections, repeated instructions, and wrong-sized models is a token that could have gone toward sharper, more accurate output on your actual task.

The people who get consistently excellent results from Claude aren’t using a different tool. They’re just more intentional about what goes on the counter — and they know when to wipe it clean and start fresh.

That’s not a technical skill. It’s a working habit. And now you have it.

— Keyuri

Let me guess why you opened this.

Your company has mandated AI. Everyone’s expected to use it. And somewhere in the back of your mind there’s a quiet, nagging feeling that everyone else already knows what they’re doing — and you’re the only one still figuring it out.

Here’s what I want you to know before we go any further:

You are not behind. Almost everyone is a beginner right now. The gap between where you are and where you want to be is not intelligence or technical skill. It's one thing.

In this post I’m going to walk you through the exact framework I teach in my AI Super Practitioner Certification — a structured, repeatable approach to mapping your role against AI capabilities and finding exactly where to start. No coding required. No jargon. Just a clear method that works for business leaders, operators, and managers at every level.

By the end you’ll have a framework, a step-by-step audit process, and a free downloadable template to do this yourself in under 15 minutes.

The Mindset Shift That Changes Everything

Most professionals approach AI like a gym membership in January. They take a course, try it twice, get a result that isn’t quite right, and quietly stop. Three months later they feel even more behind than when they started.

That’s not a skill problem. That’s a practice problem.

Think about how anyone builds a real physical skill — a runner training for their first 5K, an athlete developing match-level fitness. They don’t get there by showing up once. They don’t get there by watching. They get there by building a daily practice, however small, and repeating it until it becomes instinct.

AI fluency works exactly the same way

What does building the AI muscle look like in practice? It’s faster and easier than you think:

You write a report. Instead of starting from a blank page, you open Claude first. You get a confusing email. You ask Claude to help you think through your response. You have a meeting coming up. You ask Claude to help you prepare better questions.

None of those require a computer science degree. They require a habit. And habits are built by starting with the right first action — which is exactly what the rest of this post is designed to help you find.

Are you ready to build your AI muscle?

Leave a comment

The Framework: Jobs to Be Done

Most AI guides ask: “What can AI do?” That’s the wrong starting point. It leads to a long list of features you’ll never use.

A better question: “What are you actually hired to get done — and where is the friction that’s costing you time and energy?”

This reframe comes from the Jobs-to-Be-Done framework, originally developed by Harvard Business School professor Clayton Christensen. JTBD argues that people don’t hire products for their features — they hire them to accomplish an outcome. Applied to your career and AI, it works like this:

For example: if you’re an Operations Manager, your functional job isn’t “manage operations” — it’s “ensure the business runs predictably so leadership can focus on growth.” Your micro-jobs include things like weekly reporting, supplier coordination, onboarding new hires, and running stand-ups. Your friction points are the parts of those micro-jobs that drain your time every week — the ones you dread, rush, or do on autopilot.

Friction is the signal. That’s where we start the audit.

Subscribe now

The AI Role Audit: A Step-by-Step Guide

The AI Role Audit is a structured exercise where you use Claude to map your role against AI capabilities — and discover exactly where to start getting your time back. It takes 10–15 minutes, requires no technical background, and produces a personalized action plan you can act on the same day.

Here’s how to run it.

Step 1 - Set the stage

Before you share anything, give Claude a role and a goal. This frames every response that follows.

Copy this prompt

“I’m going to share my job description with you. Act as a career strategist who understands AI capabilities. Help me find where AI can save me time, amplify my output, or give me a competitive edge. Ready?”

Step 2 - Add your job description

Paste your actual job description — or if you don’t have one handy, type your top 5–7 responsibilities in plain language. It doesn’t need to be polished. A rough list works just as well.

Step 3 - Run the audit

Ask Claude to sort your role by AI opportunity across three buckets.

Copy this prompt

“Based on this job description, identify which tasks could be meaningfully assisted by AI today. Group them: (1) High-impact quick wins, (2) Tasks requiring human judgment with AI support, (3) Tasks AI can handle almost entirely.”

Step 4 - Go deeper on one area

Pick the most interesting or surprising result — the one that made you think “wait, really?” — and ask Claude to make it immediately actionable.

Copy this prompt

“For [specific task], give me a concrete example of how I could use Claude today to do this faster or better. Include the exact prompt I would use.”

Step 5 - Build your AI Opportunity Map

Finish by asking Claude to package everything into a clear, actionable summary.

Copy this prompt

“Give me my Top 3 AI Quick Wins. Format: Task | Why it matters | Prompt to use.”

Choosing the Right AI Relationship for Each Task

Once you have your list of micro-jobs, the next step is deciding how to involve AI in each one. Not every task belongs in the same relationship. The difference between effective AI use and frustrating AI use often comes down to choosing the wrong archetype.

AI Role Audit Download

Why Non-Technical Professionals Have an Edge

Here’s something that surprises most people when I say it: the professionals with the most to gain from AI are not engineers. They’re the business leaders, operators, and managers who know what good looks like in their field.

Think about what AI actually needs to produce great output. It needs context. It needs domain expertise. It needs someone who can recognise quality, ask the right questions, and apply judgment to the result. Those are not technical skills. They’re the exact skills that experienced non-technical professionals have spent years developing.

Your domain expertise + Claude = output that a junior hire couldn’t match.

The professionals who will get left behind aren’t the ones who don’t know how to code. They’re the ones who never build the habit of using AI intentionally. That’s the only gap that actually matters — and it’s entirely within your control to close.

The skill AI can't replicate is knowing what good looks like in your industry. Context, judgment, relationships, trust — these are business superpowers. And they're yours

If you’ve come this far, leave a comment and let me know how it went

Leave a comment

Your one action from this post: Run your AI Role Audit this week. Use the five prompts above directly in Claude, or download the full guided template below. It takes 15 minutes, and it will show you more about where AI can help you than any course ever could.

If this post was useful, share it with one colleague who’s been telling you they feel behind on AI. They need to read the opening three lines.

And if you want to go deeper — the AI Super Practitioner Certification covers this framework and much more, with live sessions, hands-on demos, and a full curriculum built for non-technical professionals who want to lead in an AI-first workplace.

— Keyuri

Discover your AI superpower and how to go from zero to hero using AI. Join my July cohort of AI SuperPractitioner Certification. Stop figuring things out on your own and follow this step-by-step course.

Most people open Claude, type a question, and wait.

They get an answer. Sometimes a good one. They close the tab and move on.

Then they open Claude the next day and start from zero again. Same context. Same explanation of who they are, what they do, what they need. Every single time.

This is the most common mistake non-technical professionals make with AI — and it’s completely fixable in about 20 minutes.

This post is about what to do before you start using Claude for real work. The setup that most people skip. The start before the start.

Why Most People Use AI Wrong

Here’s the honest truth: Claude, used casually, is a very expensive search engine.

You put something in, you get something out, you forget it happened. There’s no memory. No context. No sense of who you are or how you work. You are, to Claude, a stranger every single time.

That’s not Claude’s fault. That’s a setup problem.

The professionals who get the most out of AI aren’t necessarily the smartest prompters. They’re the ones who built a system so that Claude already knows them — before the conversation starts. They did the setup work once, and now every single task starts from a better place.

That’s what we’re building today.

What You’re Building

By the end of this post, you’ll have:

A folder system — one home base on your computer where Claude lives and works, with a clear structure it can navigate.

Three context files — short documents that tell Claude who you are, how you write, and what rules to follow. It reads these before every task.

A Cowork workflow — a live demonstration that the whole system works, using a real task from your actual job.

Let’s go.

Step 1 — Build Your Claude-Work Folder

Open your Desktop. Create a new folder. Call it Claude-Work.

Inside it, create four subfolders:

• About Me — who you are, how you work, your writing style

• Projects — one subfolder per live project or client

• Templates — your best past work, saved as starting structures

• Outputs — where Claude saves everything it creates

That’s it. That’s the whole structure.

Why does this matter? Because Cowork — Claude’s desktop tool for file-based tasks — needs a defined space to operate in. You give it access to Claude-Work, and it can read, write, and organise within it. You stay in control. Nothing happens outside this folder without you knowing.

Think of Claude-Work as your AI’s desk. Right now, most people have their AI working on the floor with no desk, no filing system, and no idea where anything lives.

Step 2 — Write Your Three Context Files

This is the most important part of the setup. Most people never do this. It’s also the thing that separates people who feel like AI is useful from people who feel like AI is magic.

Open Notepad (Windows) or TextEdit (Mac — set it to plain text first via Format → Make Plain Text). You’re going to create three files and save them into your About Me folder.

File 1 — About-me.txt

This file answers one question: who are you and what do you do?

Not your life story. Not your LinkedIn summary. The working version. What your actual day looks like, who you work with, what tools you use, what you’re trying to get better at.

Here’s an example of what a good one looks like:

My name is Sarah Chen. I’m a Senior Account Manager at a marketing agency. I manage 6 client accounts across retail and healthcare. My days are mostly client calls, writing briefs, reviewing campaign performance, and chasing my team for updates. I work closely with Maya (Creative Director) and Tom (Data Analyst). I use Google Workspace, Slack, HubSpot, and Asana. I’m not technical — if something requires coding, tell me.

Notice what makes this useful: it’s specific. It has real names, real tools, real context. The more specific your about-me file, the more relevant Claude’s output becomes.

Save it as: about-me.txt in your About Me folder.

File 2 — My-voice.txt

This file teaches Claude how you write.

Here’s why this matters: when Claude drafts an email or a report for you, it defaults to a generic professional tone. Technically correct. Completely not you. You end up rewriting half of it anyway.

My-voice.txt fixes that.

Write down how you’d describe your own style. Direct? Warm? Formal with clients, casual with your team? Do you hate corporate jargon? Do you write short emails or long ones? Do you use bullet points or paragraphs?

Then — and this is the most powerful part — paste in two real examples of your own writing. A client email. A Slack message. Anything that sounds like you. Claude learns your voice faster from one real example than from ten descriptions of it.

Also add a list of words or phrases it should never use in your name. “Leverage.” “Circle back.” “As per my last email.” Whatever makes you cringe — put it here.

Save it as: my-voice.txt in your About Me folder.

File 3 — My-rules.txt

This file sets the guardrails.

It tells Claude what to do before it starts any task, what to do if it gets stuck, and what it must never do under any circumstances.

A simple version looks like this:

Always read my About Me and My Voice files before doing anything. Ask me a clarifying question before you begin. Show me a plan before you execute — wait for my approval before starting. Never delete a file. Never send anything on my behalf. If you’re unsure, stop and ask.

Three non-negotiables that belong in every my-rules file:

1. Ask before starting. You’d rather answer a question than redo the work.

2. Show a plan first. Surprises in AI output are rarely good surprises.

3. Never delete. This one is non-negotiable. Always.

Save it as: my-rules.txt in your About Me folder.

Heres a downloadable template for all 3 files to get you started. Simply download and make them your own.

Download Templates

Step 3 — Set Your Cowork Instructions

Open Claude Desktop. Go to Settings → Cowork → Edit Instructions.

Paste something like this:

I’m [Your Name], [Your Role]. Read my files in About Me before every task. Ask me questions before you start executing anything. Always show me a plan first. Never delete a file without my explicit approval.

This runs automatically on every Cowork task. You set it once. It never forgets.

Step 4 — Run Your First Real Workflow

Now let’s test the whole thing.

Go to your Projects folder inside Claude-Work. Create a subfolder — call it something like Client-Meeting-June. Drop in 3–5 files from a real meeting: rough notes, a PDF, an email, a bullet list you typed on your phone. Whatever you have. The messier the better.

Open Cowork in the Claude desktop app. Give it access to your Claude-Work folder. Then type:

Read my About Me files, then read everything in Projects/Client-Meeting-June. Write me a clean action report with three sections: key decisions made, open questions, and next steps with who owns each one. Save it as action-report.docx in the Outputs folder.

Press enter. Step away.

When you come back, open your Outputs folder. Your action report is there.

That’s the system working. Claude read your context, applied your rules, processed your files, and delivered a structured output — without you explaining who you are, formatting anything, or starting from scratch.

If you’ve come this far, leave a comment and let me know how it went.

Leave a comment

The Takeaway

Here’s what you actually built today:

You built a briefing system.
That runs before every task.
No need to repeat yourself.
It makes Claude’s output sound like you, follow your rules.

Most people using AI right now are doing it the hard way but you are set up to not do that anymore.

The start before the start is the thing that makes everything after it better.

Keyuri!

Discover your AI superpower and how to go from zero to hero using AI. Join my July cohort of AI SuperPractitioner Certification. Stop figuring things out on your own and follow this step-by-step course.

Weekly posts on Building AI Products and Mastering AI for your daily life to increase productivity and improve quality of life. From an AI Product Leader & Educator who has up skilled over 3000+ professionals in AI.

Andrej Karpathy is one of the people who actually built the technology behind modern AI. He co-founded OpenAI, led AI at Tesla, and now spends a lot of time thinking about what AI means for how humans learn and work.

He said something recently that I keep coming back to:

“You can outsource the execution. You can’t outsource the understanding.”

It sounds obvious. But I don’t think most professionals using AI right now have actually sat with what it means.

The trap nobody’s talking about

Here’s what’s happening inside most companies right now: someone gets access to Claude or ChatGPT, uses it to write faster, summarize longer, respond quicker, they call it AI adoption.

The output looks polished. The speed is real. And then six months later, that same person realizes they can’t explain why any of their AI-assisted work was good. They just know it came back clean.

That’s a judgment gap that’s quietly widening. They are loosing taste. The taste that makes them good at their job.

The people who will actually matter in AI-transformed orgs aren’t the ones who learned to prompt faster. They’re the ones who kept their judgment sharp while letting AI handle the execution. Those are two different skills. And most AI training right now is only teaching one of them.

What “outsourcing understanding” actually looks like

Let me make this concrete, because I think it lands differently when you see it in your own work.

Imagine you’re a Customer Success Manager. You’ve been using AI to summarize customer calls, draft renewal emails, and prep for QBRs. The output is good. Your manager is happy. You feel productive.

Now imagine you’re in a room with your VP and a customer who just decided not to renew and your VP turns to you and says:

“Walk me through the last 90 days of this account. What signals did we miss?”

You cant say “I’m not sure, I had AI summarize the calls,” but that’s the moment the gap becomes visible. The problem wasn’t that you used AI to summarize. The problem was using AI to replace the synthesis, the part where you connect the dots, notice the pattern, form the judgment. That’s the part that can’t be outsourced without cost.

AI should make your thinking sharper, not optional

This is where Karpathy’s line becomes actionable, not just philosophical.

The best use of AI isn’t to get output faster. It’s to get to the right question faster and then bring your own thinking to it harder.

In practice, that means building prompts that force friction, not just fluency. Here’s what that looks like:

INSTEAD OF THIS

“Summarize this call transcript.”

TRY THIS

“Summarize this call transcript. Then list the three things the customer said that I should be most concerned about and tell me what follow-up question I should be asking that I probably haven’t thought of yet.”

The first prompt outsources the work. The second one uses AI to push your thinking somewhere it wasn’t going on its own.

Same tool. Very different relationship to your own judgment.

Pick one AI task you do regularly — summarizing, drafting, reviewing, prepping. Something you’ve put on autopilot.

This week, add one of these lines to your prompt:

→ “What am I not asking that I should be?”

→ “What’s the strongest argument against this?”

→ “What would I need to believe for this to be wrong?”

Notice what comes back. Notice whether it makes you think harder, or just gives you more polished output to nod at.

That noticing, that active judgment about what AI is actually giving you is the skill that separates people who are useful in an AI-first org from people who are just fast.

Speed without judgment is just faster mediocrity.

If this resonates, the next step is building that judgment into a repeatable system, beyond just one prompt, but a way of working where AI handles the execution and you stay sharp on the thinking.

💡If you enjoyed this content, share it with a friend.

Share AI First Academy

Learn to use AI to 10X your productivity and go from AI Curious to AI SuperPractitioner with my live cohort-based course: AI SuperPractitioner Certification. We start in July!

Every week, someone asks me the same question:

“I want to build in AI… but how do I know what’s actually worth building?”

It’s a good question.

When you can build anything, what should you actually build?

The best AI products are not born from model capability but from operational pain.

Start with:

• Where does work slow down?

• Where does money get stuck?

• Where do humans spend time validating, coordinating, reviewing, or chasing information?

• Where does the business bleed operational efficiency?

That’s where the real opportunities are.

Over time, I’ve developed a simple framework to identify high-value AI product opportunities before building anything. The same framework can be applied to figuring out what personal productivity agents you should build.

Here’s the framework.

1. Identify the “Friction Gates”

Every business has invisible traffic jams.

Moments where work pauses because humans need to:

• verify something

• approve something

• reconcile information

• coordinate across teams

• manually review documents

• transfer context

I call these Friction Gates.

And AI products become valuable when they reduce the waiting time around those gates.

Let’s use construction as an example.

In construction projects, one of the biggest friction points happens between:

• the Prime Professional (architects/design teams)

• subcontractors

• project managers

• compliance teams

Thousands of documents move back and forth:

• RFIs

• submittals

• permits

• insurance certificates

• change orders

• contracts

• inspection reports

Most delays happen because humans are acting as API layers between fragmented systems.

That’s the opportunity.

The AI PM mindset is:

“Where are humans repeatedly validating information at scale?”

That’s where AI thrives.

Especially when the workflow involves:

• high-volume document review

• repetitive approvals

• structured decision-making

• cross-system reconciliation

• status tracking

• exception handling

The best AI products often remove coordination overhead, not labor itself.

That distinction matters.

2. Define the Autonomy Level

One mistake I see often:

Teams jump straight into building fully autonomous AI agents before understanding whether the workflow actually needs autonomy.

Not every problem needs an AI employee. Some problems just need better assistance.

I like to categorize AI opportunities into two buckets:

Supportive AI

The AI helps humans move faster.

Examples:

• summarizing 500-page subcontractor agreements

• extracting clauses from contracts

• prioritizing RFIs

• generating project status reports

• identifying missing compliance documents

The human still owns the final decision.

These are often the easiest AI products to launch because:

• liability is lower

• trust barriers are lower

• adoption is faster

• workflows require less orchestration

Then comes the second category.

Agentic AI

The AI takes action.

Examples:

• monitoring insurance certificates

• blocking subcontractor access if coverage expires

• automatically routing approvals

• triggering escalation workflows

• coordinating deliveries

• generating and filing documentation autonomously

This is where AI starts behaving less like software and more like an operational worker.

But here’s the important part:

Agentic AI is not inherently “better.”

Higher autonomy means:

• higher trust requirements

• more observability needs

• governance complexity

• failure management

• operational accountability

A lot of teams are trying to build Level 5 autonomy products when the market is still learning how to trust Level 2 assistance.

That’s why some of the best AI opportunities right now are:

• semi-autonomous workflows

• human-in-the-loop systems

• recommendation engines with action capabilities

• copilots with constrained execution

The sweet spot is often:

“High operational leverage with bounded risk.”

3. Map the Liability Chain

This is the step most builders skip and it’s why many AI products fail after the demo phase.

Before building any AI workflow, ask:

“If this agent makes a mistake… who pays, whats the risk?”

That answer changes everything.

For example:

An AI agent verifying structural rebar placement?
Extremely high risk.

A mistake there can create:

• safety issues

• legal exposure

• insurance problems

• regulatory violations

Now compare that to:

An AI agent tracking appliance deliveries across 200 apartment units.

If it fails:

• someone makes a phone call

• delivery gets rescheduled

• operations slow temporarily

Very different liability profile.

This is why many successful enterprise AI products initially focus on:

• operational coordination

• administrative automation

• workflow acceleration

• compliance tracking

• communication orchestration

Not core mission-critical decision replacement.

The best AI PMs understand:

AI adoption follows trust gradients.

And trust builds faster in low-liability environments.

That’s where product-market fit often begins.

4. Validate ROI Through “Smart Milestones”

This is my favorite filter.

A use case becomes dramatically more valuable when the AI output directly connects to a financial event.

I call these Smart Milestones.

Meaning:
The AI doesn’t just produce information.

It produces verifiable operational proof that unlocks money movement.

That’s where AI products become incredibly sticky.

Example:

Imagine an AI agent using computer vision to verify floor leveling in a construction project.

Instead of:

• waiting for manual inspection

• generating paperwork

• scheduling approvals

• processing payment requests

The agent creates:

• proof of completion

• verification logs

• timestamped evidence

• compliance validation

And automatically triggers the next payment draw.

Now the value isn’t:
“AI helped summarize data.”

The value is:

• payment cycles drop from 30 days to 24 hours

• subcontractors get paid faster

• project velocity increases

• disputes reduce

• cash flow improves

That is measurable ROI.

And measurable ROI is what transforms an AI experiment into a real business.

When evaluating AI product ideas, ask:

• Can this workflow produce verifiable proof?

• Can it accelerate a financial process?

• Can it reduce operational lag?

• Can it reduce coordination costs?

• Can it shorten time-to-revenue?

• Can it prevent leakage?

The strongest AI products usually sit directly next to:

• money movement

• compliance enforcement

• operational bottlenecks

• decision latency

That’s where businesses are most willing to pay.

The Real Secret

Most people think successful AI products come from breakthrough models.

In reality, they come from deeply understanding operational systems and user problems.

The future AI winners will be those that work closely with users.

They’ll be:

• workflow experts

• systems thinkers

• trust architects

• operational economists

The AI PM advantage is not:
“knowing the latest tech.”

It’s being able to see invisible inefficiencies inside industries and redesigning the workflow around intelligence.

That’s the shift.

And once you start seeing businesses through:

• friction gates

• autonomy levels

• liability chains

• smart milestones

…you’ll never run out of AI product ideas again.

Happy Building!
Keyuri

Discover your AI superpower and how to go from zero to hero using AI. Join my July cohort of AI SuperPractitioner Certification. Stop figuring things out on your own and follow this step-by-step course.

By Keyuri Anand

Weekly posts on Building AI Products and Mastering AI for your daily life to increase productivity and improve quality of life. From an AI Product Leader & Educator who has up skilled over 3000+ professionals in AI.

But here’s the uncomfortable truth: most AI product teams are still optimizing the wrong metrics.

The success of an AI product isn't about whether the model is accurate. It’s about whether the product creates value—consistently, measurably, and at scale.

And that all starts with one question: Have you picked the right core metric?

Why AI Needs a Different Success Lens

Traditional SaaS products often succeed by driving engagement. More clicks, more sessions, more time spent in-app. These make sense when your product is a tool.

But AI changes the game. AI products don’t just help users do work—they often do the work for them.

Which means we need to move from engagement metrics to outcome metrics.

If your AI system successfully completes a job with zero user clicks, that's not a failure of engagement—it’s a success of automation.

The Core Metric: What It Is and Why It Matters

A core metric is the one thing that tells you whether your AI product is creating real value.

It’s your north star.

It drives prioritization, stakeholder buy-in, UX decisions, and even model improvements. It’s not about tracking everything—it’s about aligning everyone around what matters most.

If your team can’t articulate your AI product’s core metric in one sentence, you haven’t set one.

How to Choose the Right Core Metric for an AI Product

Here’s a simple framework to identify a strong core metric:

1. Tie it to a Business Outcome
   Ask: What job is this AI supposed to do? Then define success in business terms. Think:

   • Time saved

   • Cost reduced

   • Accuracy improved

   • Decisions made

   • Revenue unlocked
     Not: number of chats started or models deployed.

2. Make It User-Centric, Not Model-Centric
   AI performance (e.g., 92% accuracy) matters—but only in service of a user goal. If your users can’t understand or feel the value, your metric isn’t strong enough.

3. Ensure It’s Measurable and Actionable
   Can you track it end-to-end? Can you influence it with product changes? Your metric must be both visible and responsive to iteration.

4. Prefer Lagging Indicators with Business Impact
   Leading indicators like clicks or logins are tempting—but for AI, lagging indicators like time to resolution or cost per invoice tell the real story.

Examples of Strong Core Metrics for AI Products

Notice how none of these involve “clicks,” “DAUs,” or even “model accuracy” as the core metric.

Those may still be supporting metrics—but they’re not what defines success.

Common Pitfalls to Avoid

❌ Chasing Vanity Metrics
You can grow weekly active users without creating value. Focus on what delivers outcomes.

❌ Overweighting Model Metrics
Your model could be 99% accurate—and still irrelevant if it doesn’t solve the right user problem.

❌ Ignoring Time-Based Metrics
Speed is often the key value prop of AI. Always ask: how much faster are we making this job?

❌ Having Too Many North Stars
If you have five “core” metrics, you have none. Pick one. The others can be secondary or diagnostics.

Make It Visible, Make It Cultural

Once you’ve chosen your core metric:

• Put it in your dashboards.

• Review it in every sprint retro.

• Share it with your execs.

• Bake it into your launch criteria.

Your team should feel it. Every feature should connect back to moving this number in the right direction.

Final Reflection

AI products aren’t judged by how often people use them—they’re judged by what they get done.

When you set a single, business-aligned, user-focused core metric, you align your AI product with what really matters: impact.

And when your metric measures impact, your team builds for outcomes—not just interaction.

How They Work

Traditional AI relies on predefined rules and algorithms to analyze data and make decisions. Think of it as a skilled analyst: it processes inputs and produces outputs based on established patterns.

• Example: A spam filter in your email uses traditional AI to detect keywords and patterns associated with unwanted emails. Its model is trained to categorize emails but cannot create new content or deviate from its training.

Generative AI, on the other hand, creates new content based on patterns it has learned. Powered by advancements like deep learning and large language models (e.g., GPT or DALL-E), it generates text, images, music, or even code from scratch.

• Example: It’s like an artist trained on thousands of masterpieces, creating its own original work while adhering to learned styles and themes.

Key Differences in a Nutshell

• Purpose: Traditional AI excels at tasks with well-defined rules, while Generative AI focuses on creating novel outputs.

• Learning: Traditional AI relies on explicit programming, whereas Generative AI learns patterns from massive datasets.

• Output: Traditional AI provides solutions or classifications, while Generative AI produces original content.

• Complexity: Generative AI models are often more complex and require significant computational resources.

• Control: Traditional AI is easier to control and interpret, while Generative AI can sometimes produce unexpected or unpredictable results.

• Bias: Both types of AI systems can be susceptible to biases present in their training data.

Why It Matters - PM Strategies

Understanding the distinction between these AI paradigms is key to setting the right expectations for your product:

1. Define the Problem: Traditional AI excels in solving well-defined, rule-based problems, such as fraud detection, demand forecasting, and inventory optimization.

2. Unleash Creativity: Generative AI is perfect for innovation and user engagement. It can help create conversational agents, automate content creation, or personalize marketing campaigns at scale.

3. Data-Driven Decisions: While both AI types require quality data, generative AI often needs vast and diverse datasets to perform effectively. Consider data availability and quality during product development.

4. Ethical Guardrails: Generative AI introduces new challenges, like managing misinformation and biases in generated content. Incorporating AI governance and monitoring is vital for responsible deployment.

1. Identify the Problem Scope:

   • Is your problem rule-based and well-defined? → Go for Traditional AI.

   • Does your problem require creativity or generating new content? → Consider Generative AI.

2. Evaluate Data Availability:

   • Do you have structured, labeled datasets? → Traditional AI is likely sufficient.

   • Do you have access to large, diverse datasets for training? → Opt for Generative AI.

3. Assess User Needs:

   • Are you aiming to optimize or predict? → Use Traditional AI.

   • Are you trying to engage or innovate with users? → Choose Generative AI.

4. Analyze Ethical Considerations:

   • Ensure that data governance and ethical guidelines are in place for Generative AI to mitigate risks of misinformation or bias.

5. Prototype and Test:

   • Build small-scale models for each type to see which best aligns with your goals.

6. Iterate and Improve:

   • Use feedback to refine the chosen AI type, ensuring it meets business objectives and user expectations.

Real-World Applications: What You Can Build

Traditional AI:

• Predictive Maintenance: Detect equipment issues before failures occur, saving costs.

• Recommendation Systems: Suggest products or content based on user behavior (e.g., Netflix, Amazon).

• Fraud Detection: Identify suspicious transactions in banking and finance.

Generative AI:

• Chatbots and Conversational Agents: Engage users with human-like interactions.

• Content Creation: Generate marketing copy, visuals, or personalized campaigns.

• Creative Tools: Build apps that create art, music, or personalized experiences.
  

  Generative AI and Traditional AI both offer immense value but serve distinct purposes. As a product manager, understanding their differences ensures you’re choosing the right tool for the job, whether it’s optimizing existing workflows or driving innovation. By following the step-by-step guide, you can align AI capabilities with your business goals, unlocking the true potential of artificial intelligence in your product.

  Thanks for reading! This post is public so feel free to share it.

  Share

AI > ML > DL > Gen AI

Four acronyms. A thousand conversations. And yet, I’ve seen countless professionals (even in tech) confuse them or use them interchangeably.

If you’re an AI Product Manager, you can’t afford to be part of the confusion. So, let’s break it down.

Subscribe

AI: The Parent Term

Artificial Intelligence (AI) is the broadest concept. It’s about making machines mimic human intelligence, solving problems, learning, or making decisions.

Think: chatbots or recommendation engines. It’s not always cutting-edge, but it’s foundational.

ML: The Engine of Modern AI

Machine Learning (ML) is AI’s workhorse. Instead of pre-programming every behaviour (rule based systems where you pre define all conditions <if, then, else>), ML lets machines learn from data.

Think: Netflix recommending a show based on your watch history.

ML is like teaching a child - correcting mistakes, reinforcing patterns.

DL: ML on Steroids

Deep Learning (DL) is a subset of ML. It uses neural networks that mimic the human brain’s structure.

This is why we have self-driving cars and near-human speech and image recognition. But it’s resource-intensive.

Think: expensive GPUs and massive data sets.

Generative AI: The Newcomer Stealing the Spotlight

Generative AI? Like the name suggests, It generates. It doesn’t just analyze data but it builds new content- text, images, code, music, 3D models and so on.

Think: ChatGPT, DALL·E

It’s why “AI” has gone mainstream.

But that’s not where this stops. Gen AI is just one slice of the broader AI pie and all the innovation that’s currently happening.

Why This Matters

As an AI Product Manager, you need to demystify these terms for your team, stakeholders, and customers.

Not knowing the difference could mean misaligned strategies or unrealistic product roadmaps.

A clear AI lexicon isn’t just jargon, it’s your credibility.

Subscribe

Final Thought

AI isn’t magic. It’s tools, algorithms, and systems designed to enhance what we do.

Understanding the nuances and how this works, that’s where you set yourself apart as a Product Manager.

P.S. If you found this helpful, repost for others in your network. Let’s spread clarity.

Share

As AI continues to transform industries, it’s clear that product managers play a crucial role in bridging the gap between AI’s potential and real-world impact. This blog is my way of navigating that journey and inviting you along for the ride.

Let me set the stage: I’m not an AI wizard or a machine learning savant. I’m a product manager, someone who has spent years building products by connecting technical intricacies with user needs. My success has always come from asking the right questions, finding actionable insights and hands on building. Being curious and learning relentlessly have been my growth levers.

What to Expect?

AI PM Playbook is a series of blogs, structured around three key pillars listed below, and growing in complexity from Foundations > Deep Dive > Mastery.

• AI/ML Concepts: Breaking down the complex AI and ML Technical details to make it understandable. Learn ‘how’ it works.

• AI Product Management Strategies: Tools, frameworks, and processes to lead AI-driven products. Learn ‘why’ it matters.

• Real-World Applications: Insights from experimenting with AI tools and workflows. Learn ‘what’ you can build with it.

Published blogs on each of these topics will be linked to the corresponding titles below. This page will remain pinned for easy reference. Use this playbook like you would use the instruction manual for building Legos.
Iterative and Fun!

The Playbook!

1. FOUNDATIONS

1.1 AI/ML Concepts: Breaking Down the Technical Foundations

• Generative AI vs. Traditional AI: Key differences and applications.

• The AI Lexicon: AI, ML, DL, and Gen AI demystified

• How Generative AI Works: Mechanisms behind AI outputs.

• Computational Thinking: Deterministic vs. Non-Deterministic methods.

• LLMs & Transformers: Tokenization, Attention, and their significance.

• Introduction to RAG and vector databases.

• Neural Networks & NLP basics: Understanding embeddings, stemming, and lemmatization.

• AI Agents vs. AI Assistants: Functional differences and use cases.

• Prompt Engineering 101: Zero-shot, few-shot, and prompt-taxonomy basics.

1.2 AI Product Management Strategies

• The Role of Data in AI: Why it’s the fuel of AI products.

• Supervised vs. Unsupervised Learning: Decoding workflows and metrics.

• Overfitting in Models: Avoidance strategies for reliability.

• Building Intuition for Model Evaluation: Metrics, feature selection, and NLP nuances.

1.3 Real-World Applications

• AI Ethics and Governance: Tackling bias, copyright, and misinformation.

• Generative AI’s Ethical Challenges: A societal perspective.

• AI in Action: Simple tools and workflows for new AI PMs.

• Breaking Down AI Hype: Separating promise from reality.

2. DEEP DIVE

2.1 AI/ML Concepts: Evolving Knowledge

• Advanced Architectures: LLMs, transformers, and fine-tuning strategies.

• Advanced Prompting Techniques: CoT, ReAct, and Self-Reflection.

• Neural Networks Deep Dive: From RNNs to transformers.

2.2 AI Product Management Strategies

• AI Product Lifecycle: Unique cycles for building AI-driven solutions.

• Building a Data Moat: Data strategies to ensure a competitive edge.

• AI Personalization: Tailoring experiences for diverse segments.

• MLOps Practices: Ensuring model safety and observability.

2.3 Real-World Applications

• Technical Building Blocks: Orchestration, data labeling, and synthetic data workflows.

• Gen AI Product Development: Advanced workflows and models.

• Graph RAG and Personalization Strategies: Real-world examples.

3. MASTERY

3.1 AI/ML Concepts: Expert-Level Insights

• Model Optimization Techniques: Fine-tuning for performance and scalability.

• AI in Edge Computing: Deployment on edge devices.

• Scalability Challenges: Distributed computing and caching for AI systems.

3.2 AI Product Management Strategies

• Advanced Frameworks: The 10× PRD Framework for scalable AI products.

• Agent-Centric Workflows: Memory systems, multi-agent design, and orchestration.

• Cross-Model Orchestration: Combining models for advanced capabilities.

3.3 Real-World Applications

• Advanced RAG Systems: Designing sophisticated retrieval products.

• AI Pipeline Optimization: Balancing cost, reliability, and performance.

• Advanced Responsible AI: Governance, bias detection, and real-world ethics.

  Bookmark this page & Subscribe follow along

  
  
  

1️⃣ AI vs. ML vs. DL
AI is the what: systems that mimic human intelligence.
ML is the how: teaching machines to learn patterns from data.
Deep Learning (DL) is the turbocharged version: think neural networks solving complex problems like image recognition.

2️⃣ Supervised vs. Unsupervised Learning
Supervised: Machines learn with labeled data (e.g., emails tagged spam or not).
Unsupervised: Machines find patterns on their own (e.g., grouping customers by behavior).

3️⃣ Model Training Basics
Models are trained on historical data (e.g., training Netflix to recommend shows).
The more relevant the data, the smarter the model gets. Garbage in, garbage out!

4️⃣ Data is the Real MVP
Clean, high-quality data fuels AI success.
PMs play a key role in defining what data matters and how to measure outcomes.

5️⃣ Explainability Matters
AI decisions need to be understood by users and stakeholders. As a PM, you'll champion transparency and fairness.

👉 Why PMs Thrive in AI
Your experience prioritizing features, understanding users, and aligning stakeholders directly applies to building AI Products. You'll add a layer of collaboration with data scientists and ML engineers while keeping users at the center.

Best way to learn is to Build!

Start small - experiment with AI tools or build your own mini-projects. The AI world is complex but incredibly rewarding for PMs ready to dive in.

Thanks for Reading!

Subscribe now

Share

Leave a comment