How (& Why) to Build Your Own AI Tech Stack

L&D got <15% of the AI tech tools that product managers got. Here's how to build the rest yourself.

May 14, 2026

Hey folks 👋

Last month, Lenny Rachitsky published his fourth biannual State of the Product Job Market. The headline numbers are striking…

Product Manager (PM) hiring is at a three-year high.
AI-specific PM hiring has increased fastest - up 340% since 2024.
AI-first Product Managers now get paid on average $200K–$305K, with a documented 22% premium over traditional senior PM roles.
Some senior AI PM roles offer annual salaries of up to ~$1M.

Meanwhile - despite deep similarities in their role and workflows - we are not seeing anything like this level of AI disruption in the world of L&D. Between 2024 and 2026, L&D salaries flatlined. There is an emerging ~20% premium for AI-fluent L&D professionals, but only at the edges, and only in a minority of advanced organisations.

In this week’s blog post, I ask: why is this — and what would it take to change it?

Let’s dive in!

AI-Enhanced Product Management

The rapid growth of the PM role is powered by a rich, multi-layer AI tool stack built specifically for the product workflow. Without this stack, the growth of product management as a profession — and the salary premium for AI-enhanced employees — doesn’t exist. There’s no $1M AI PM if there’s no tooling to make that role 10× more productive than a traditional PM.

So what does this stack look like in practice? Think specialised AI tooling weaved throughout the PM workfloe, from discovery through to evaluation:

Dovetail for discovery — uses AI to cluster user interviews, support tickets, and survey responses into named themes, so PMs can synthesise hundreds of sources in under an hour.
Productboard for prioritisation — uses AI to aggregate feedback from Slack, sales calls, and support tools, then score features against impact and effort, so PMs can defend roadmaps with data.
ChatPRD for spec docs — built by Claire Vo, this tool uses AI trained on thousands of PRDs to turn rough notes into structured requirement documents, user stories, and acceptance criteria in minutes.
Cursor and v0 for production — use AI to generate working code and UI mockups from natural-language prompts, so the engineers and designers PMs work with ship faster.
Linear for execution — uses AI to watch commits, pull requests, and branch activity, then auto-update roadmap status, so PMs stop chasing engineers for progress reports.
Amplitude for analytics — uses AI to detect anomalies in user behaviour and answer natural-language questions like “why did activation drop in Europe last week?” in seconds instead of days.

Now, let’s look at L&D through the same lens.

AI-Enhanced L&D

If L&D had been disrupted the way Product Management has, you’d expect to see a parallel stack: AI tools that cluster stakeholder interviews to surface capability gaps. Tools that prioritise training requests against business impact. Tools that turn diagnosed performance problems into structured learning design briefs. Tools that measure whether what shipped actually moved the metric.

Instead, here’s what the AI-for-L&D market actually built:

L&D got <15% of the AI stack built for other industries, like Product Management

Let’s take a moment to compare the PM AI stack against the L&D equivalent:

Discovery — nothing purpose-built. No L&D equivalent of Dovetail. Most L&D teams cluster stakeholder interviews in spreadsheets or ChatGPT, by hand.
Strategy & decision support — nothing purpose-built. No agent watching the L&D function’s data continuously, surfacing decisions like “this onboarding course hasn’t moved the conversion metric in 90 days — recommend retire.” That agent doesn’t exist.
Specification — nothing purpose-built. No L&D equivalent of ChatPRD. Most designers use general-purpose Claude or ChatGPT with bespoke prompts.
Prioritisation — emerging. Some early-stage startup activity in the last year, but nothing with the depth of Productboard.
Execution — partial. L&D teams have been borrowing Jira, Asana, and Trello for years.
Production — saturated. Synthesia, Colossyan, Articulate AI, Easygenerator, Mindsmith — plus dozens more. The one layer where the AI tech stack is large and growing rapidly.
Analytics & behavioural data — emerging. No L&D equivalent of Amplitude. Watershed exists with limited scope; nothing that links learning data to business performance.

Product Managers got six layers of AI tooling; L&D got one. But the gap here isn’t only about numbers — it’s about kind.

PMs got tools that help them decide what to make, evaluate whether it's working, and prove the impact to the business. L&D got tools that help us make more courses and more content. One stack equips a function to think strategically, the other equips a function to produce faster. The difference between those two stacks is the difference between a role that gets elevated by AI and one that gets compressed and devalued.

Why Has L&D Been Left Behind?

A question I’ve been thinking about in the last few weeks is: why is the L&D AI tech stack so concentrated in one phase and so limited in its impact compares with other industries? So far, four key themes emerge and converge:

⏰ Time

PMs buy in days, L&D buys over months. A PM typically has the freedom to sign up to a tool like ChatPRD at $15/month and becomes a champion in a week. An LMS decision takes 6–18 months to work its way through IT security, HR procurement, finance, and legal. By the time an AI-for-L&D vendor clears that process, the category has moved on.

💰 Money

PMs have suitable budget,L&D doesn’t. PM sits next to engineering with CTO/CPO sponsorship and six-figure tool budgets. L&D sits inside HR with budgets capped at LMS-renewal-scale. There are more L&D professionals than PMs globally, but PMs control more dollars per head. Bigger budgets attract more venture investment, which drives more innovation, which builds more tools. The AI-for-PM market exists because PM buyers had the budgets to fund it.

📈 Rigour

The PM demands more technical value than L&D. PMs read GitHub repos and assess AI product design on technical merit. The AI-for-PM market rewards sophisticated tools because buyers can - and do - evaluate them rigorously before they’re bought. L&D isn’t expected to evaluate this way, and the AI-for-L&D market tends to reward what demos well in 20 seconds. Production tools demo well in 20 seconds, decision-support tools don’t.

📊 Data

PM tools require data, L&D tools don’t. PMs prove value and impact through activation, retention, NPS and revenue. L&D proves value through completion rates and smile sheets. Vendors serving L&D don’t build impact analytics because buyers don’t measure impact. Buyers don’t measure impact because the LMS doesn’t track it. The loop closes, and nothing changes.

The result: two near-identical workflows on two completely different trajectories.

“But can’t we just use the AI tools built for PMs etc?”

This is a very reasonable question that I get asked a lot. If Dovetail can cluster interviews, why can’t L&D use it for stakeholder discovery? If Linear is a great project management tool, why don’t L&D teams adopt it?

Honestly, some of these tools work fine for L&D tasks. Notion AI handles L&D specs well. Linear and Jira are already widely borrowed for some L&D project management. Dovetail could be adopted for capability discovery with some translation work. These are real options, and the “adopt the PM stack” move is genuinely available to L&D folks who want to act quickly.

But if my work with large corporates has taught me anything, it’s that three structural barriers make AI adoption a lot harder than it looks.

1.The buying motion is wrong. These tools are sold to product teams through product procurement channels, priced for product budgets, and demoed against product use cases. The L&D buyer rarely encounters them through the normal vendor evaluation process. When L&D does adopt a PM tool, it’s usually because someone on the team championed it personally — not because the tool was built to find us.

Productboard - a long-time, go-to tool for Product Managers which uses AI to aggregate feedback & score features against impact and effort.

2.The integration ecosystem doesn’t fit. Dovetail integrates with Productboard and Figma, not your LMS or HRIS. Linear talks to GitHub, not Workday. The insights and outputs these tools produce don’t flow into the rest of the L&D stack, which means every adopted PM tool creates an island of data that can’t speak to the systems that matter.

3.The vocabulary and workflow are PM-shaped. These tools prompt users to identify feature requests, user personas, PRDs. L&D thinks in capability gaps, performance problems, learning briefs. A skilled L&D professional can translate — but the translation itself is friction that purpose-built tools wouldn’t impose.

ChatPRD, a tool which uses AI to help Product Managers to generate design documentation.

Headline: the market built an AI tech stack with PMs and engineers in mind. This leaves us two options: we can either wait and hope for someone to repackage it, or build the missing pieces ourselves.

Imaging the Full Fat AI Tech Stack for L&D

The other question I’ve been exploring is what the L&D-side stack would look like if it got built. How many layers it would have, what each layer would actually do, and how it would change the role of the human in the loop.

Here’s my first hot take:

Discovery — clusters stakeholder interviews, performance data, manager observations, and learner feedback into capability gaps mapped to roles. Pulls from the HRIS, not Productboard.
Strategy and decision support — an agent watching learning data and business metrics continuously, flagging decisions like “this onboarding course hasn’t moved the conversion metric in 90 days — recommend retire.”
Specification — turns a diagnosed performance problem into a structured learning brief: target behaviour, AACTT spec, learning objectives, transfer plan, evaluation criteria. Trained on thousands of briefs, not thousands of PRDs.
Prioritisation — scores training requests against business impact, learner capacity, and capability priority — and flags requests that don’t pass the diagnostic threshold (“is this actually a training problem?”).
Execution — tracks SME availability, review cycles, accessibility checks, translation rounds, and launch readiness, not commits and pull requests.
Production — the one layer L&D already has. The fix here isn’t a new tool. It’s connecting the existing ones to the six layers around them.
Analytics — ingests data from the LMS, HRIS, CRM, and performance management system, then surfaces insights like “managers who completed the leadership programme have 14% lower direct-report attrition over six months.”

The interesting part here isn’t the tooling itself - it’s what the tooling does to our role.

In a function with this stack, L&D professionals stop producing content as their primary output. They diagnose performance problems, design interventions, defend prioritisation decisions, and prove impact to leadership. The work becomes upstream and strategic. The output becomes capability change, not course completion. The role gets rewritten in the way PM’s did — and the salary premium that follows AI PMs becomes available to AI L&D leads too.

That’s the version of L&D worth fighting for: the big question is, what does it take to get there?

Build Your Own AI Tech Stack 101

So how do you become the kind of L&D pro whose role gets elevated, rather than compressed, by AI?

The bad news: at the current pace, the AI tech stack L&D actually needs is three to five years away from being built. The good news: most of it can be built today, by you, with tools you already have access to — and the build gets easier every month.

Here's what I'd build as a beginner, in roughly the order I'd build it:

Build #1: A Discovery Assistant

What to build: An AI assistant that takes notes from your stakeholder conversations, asks the right diagnostic questions, and tells you what kind of capability gap you’re dealing with.
Recommended tools: ChatGPT (custom GPT), Claude (Project), or Microsoft Copilot (agent) — whichever your IT team allows. These are all ways of creating a “trained” version of an AI tool that has your materials and methodology baked in.
The AI input: Your stakeholder conversation notes or transcripts.
What the AI does: Asks the diagnostic questions you’d ask — “What does good look like? What are people doing instead? Is this a skill gap, knowledge gap, motivation gap, or process gap?” — then analyses the answers against the examples you’ve trained it on.
To train it: Upload your diagnostic frameworks, interview templates, and 5–10 example stakeholder transcripts with the capability gaps you identified in them. Then test it against a transcript it hasn’t seen. Keep adjusting until it diagnoses as well as you would on a normal day.
The output: A structured diagnosis of the real capability gap behind a stakeholder request.
The value: You walk into every stakeholder meeting with an assistant that’s “read” every diagnostic conversation you’ve ever had — and gives you a faster, more consistent starting point than starting from scratch.

Build #2: A Brief-Writer

What to build: An AI assistant that takes a diagnosis and turns it into a structured learning design brief written in your house style.
Recommended tools: Same setup as Build #1 — a custom GPT, Claude Project, or Copilot agent.
The AI input: The diagnosis from your Discovery Assistant (Build #1), or a set of analysis / discovery notes from a good old-fashioned manual analysis process.
What the AI does: Generates a structured brief covering target behaviour, learning objectives, transfer plan, and success measures — written in the style and structure your team already uses.
To train it: Upload your best existing learning design briefs, your AACTT template, and your evaluation criteria. The trick is training it on your methodology, not generic best practice — so the briefs it returns look like yours, not like everyone else’s.
The output: A first-draft brief in your / your team’s format.
The value: Every diagnosed problem arrives 90% briefed by the time you sit down to finalise it.

An example from my own self-built stack: a Claude Project that turns documents and notes into an optimal instructional script to upload to Colossyan

Build #3: A QA Reviewer

What to build: An AI assistant that critiques your deliverables before stakeholders do.

Recommended tools: Same setup again — custom GPT, Claude Project, or Copilot agent.
The AI input: Any deliverable you’d normally send to stakeholders — a brief, a business case, a programme design.
What the AI does: Reviews the deliverable against your quality criteria and flags weaknesses, gaps, and unclear logic.
To train it: Upload examples of strong vs. weak versions of each deliverable type, alongside the quality criteria you’d use to tell them apart. Examples could be: a defensible business case vs. an enthusiastic one, a sharp brief vs. a vague one.
The output: A structured critique of your work, before it leaves your desk.
The value: The cheapest, fastest QA upgrade L&D has ever had — and it catches the things you’d miss at 4pm on a Friday.

Build #4: An Analytics Layer

What to build: A way to connect what learners did (LMS completion data) to what actually changed in the business — so you can show whether your programmes worked.
Recommended tools: Claude or ChatGPT’s data analysis features (you can upload a spreadsheet and ask questions in plain English), or Microsoft Copilot in Excel if your data already lives there.
The AI input: Your LMS completion data, paired with one business outcome you can actually measure — sales performance, retention, manager assessment scores, call-resolution times.
What the AI does: Looks for patterns between learning activity and business outcomes, and helps you explain what you’re seeing.
The output: A rough but real picture of whether your programme moved the metric.
The value: You can finally answer the question every CFO is asking — did this actually work? Even a rough version beats the smile-sheet status quo by an order of magnitude.

Note: this one’s harder and slower than the others. Expect it to take longer to set up — but the payoff is the evidence layer L&D has been missing for decades.

Another example from my self-built stack: a GPT which processes usage data from an LMS data and turns it into actionable insights

Build #5: A Connected Stack

What to build: A pipeline that wires your four AI assistants together so a stakeholder request flows from intake to evaluation without you manually moving it between tools.

Recommended tools: Zapier, Make, or Microsoft Copilot Studio. These are “agent” or “automation” platforms — they let you connect different tools so that the output of one automatically becomes the input of the next. If you’re on a Microsoft stack, Copilot Studio is usually the path of least resistance because it integrates natively with Teams, SharePoint, and Outlook (which means less plumbing and a faster path to something your IT team will sanction).
The AI input: A new stakeholder request coming in via your intake form.
What the AI does: Routes the request through your pipeline automatically:
Operational handoffs: intake form → triage queue → first-pass assessment by Discovery Assistant → weekly LMS summaries flagging underperforming courses.
Architectural handoffs: Discovery Assistant → Brief-Writer → QA Reviewer → your existing production tools → Analytics Layer watches what happens after launch → feeds insights back into Discovery.
The output: A single connected workflow instead of five disconnected tools.
The value: A stakeholder request now flows from intake to diagnosis to brief to QA to production to measurement — without you manually moving it between tools. Operational overhead of running L&D drops by up to 50%.

That’s it. Five weekends of building and you’ve built the thing the market hasn’t built yet, with tools you already pay for.

Of course “five weekends“ is an oversimplification. For many of us, building with AI feels out of reach — partly because we don’t see ourselves as “tech people,” and partly because building tools wasn’t in the job description when we became learning designers. It’s a real shift, in both skill and in identity.

So here’s the honest version: you don’t need to become a developer. You need to become someone who designs systems for their own work — which is something good L&Ds have always done, just usually with worksheets and process diagrams instead of with prompts and agents. The skill underneath is the same: noticing where a workflow leaks, designing a better one, testing it, refining it. What’s new is the toolkit, not the skill.

A few things to help you to get started:

Start with one assistant. Build #1 — the Discovery Assistant — is the lowest-risk place to begin, because the worst case is “the diagnosis it gives me isn’t quite right and I redo it myself.” That’s a free experiment with no real downside.
Treat the first version as a draft, not a tool. Your first custom GPT won’t be good, but neither was your first course. The point is to learn what you’d want from this kind of assistant, by building a version that doesn’t quite do it. Then, learn and iterate.
Get used to writing instructions for AI the way you’d write instructions for a new hire. Specific, examples-led, edge cases named. If you’ve ever explained what you do in your job and how you do it to a colleague or friend, you’re better prepared for working with AI than you think.

The L&Ds who’ll capture the AI premium won’t be the ones who were already technical - they’ll be the ones who decided to become someone who builds — even though it felt unfamiliar at first. The skill is learnable. The identity shift is the harder part. Both are worth it.

Concluding Thoughts

This “build your own” approach to AI might seem daunting and ambitious, but it’s also very real. The data shows very clearly that the AI PMs commanding $300K–$1M salaries today aren’t the ones who waited for vendors to ship the perfect tool; they’re the ones who built their own custom GPTs, their own agents, their own AI-powered workflows.

As one Meta hiring manager put it, “We don’t hire people who talk about AI. We hire people who build with AI.” The same message has been given by leaders at LinkedIn, Anthropic and others. Building with AI is rapidly becoming a differentiating skill across the AI economy. TLDR: the people who can build with AI are in high demand, and more likely will thrive - rather than be compressed and replaced - in the AI era.

But the deeper reason to learn how to build with AI isn’t the salary premium - it’s what the work becomes when you are backed by a fully-formed AI stack.

With a dynamic, cross-workflow AI tech stack, the work changes. L&D moves upstream. Our output becomes capability change rather than content and our role becomes more strategic, not functional.

That’s the version of L&D worth fighting for - and the work starts by rolling your sleeves up and learning how to build with AI.

Good luck and happy building,

Phil 👋

PS: If you want to learn how to build with AI, check out my AI Bootcamp for L&D.

Dr Phil's Newsletter

Discussion about this post

Ready for more?