Forward-deployed / Learning zone
Technical product managementa standalone module
Lesson 09

Technical product management for AI

TL;DR

Everything in this module still applies to AI products. Every practice bends the same way, for the same root cause: behaviour is probabilistic and quality is discovered, not designed. So the operating loop reorganizes around evidence. Eval-driven development makes a graded example set the spec, the regression suite, and the launch gate all at once. Discovery starts with feasibility spikes. Delivery mixes experiment-shaped work with feature-shaped work. Releases treat every model and prompt change as a migration. And the product that wins long-term is the one whose data flywheel — production feedback improving evals, evals enabling safe iteration, iteration improving the product — spins fastest. The AI PM's job is to build and protect that loop.

🎯 For the AI PM

Why it matters — This lesson is the briefing. The previous seven taught the general craft. This one assembles the AI-specific deltas into a single operating loop you can run.

What it changes in your decisions — You budget for the loop, not just the feature: eval construction, feedback instrumentation, and model-change management are line items with your name on them, not engineering hygiene you hope happens.

Ask yourself — "If the model provider shipped a new version tomorrow, does my team have a same-day, evidence-based answer to 'should we upgrade?' — or a debate?"

Risk if ignored — An AI feature that demos well, launches loudly, degrades silently, and can never be safely improved — because nobody can tell whether any change makes it better or worse.

The operating loop

The operating loop

Three connected loops · daily, monthly, quarterly

Your leverage as PM is making sure all three arrows back into the eval suite actually exist.

Spec with graded examples
→
Eval suite — the executable spec
→
Build & iterate
↻
Staged rollout — quality-gated
→
← from eval suite when it clears the bar
Production — online quality proxies
Inner loop (daily): build ↔ eval — no prompt, retrieval, or model change lands without a score.
Release loop (weekly–monthly): eval → rollout → production — quality gates the ramp.
Outer loop (quarterly): production → spec — edits, retries, and failures become new eval cases and reshape what "good" means.

Read it as three connected loops. The inner loop (build ↔ eval) is daily: no prompt, retrieval, or model change lands without a score. The release loop (eval → rollout → production) is weekly-to-monthly: quality gates the ramp. The outer loop (production → spec) is quarterly: what users actually do reshapes what "good" means. Your leverage as PM is making sure all three arrows back into the eval suite actually exist.

Eval-driven development

The eval suite is to an AI feature what the test suite is to code — except the PM co-owns it, because it encodes product judgment:

Once the suite exists, hard conversations get easy. "Can we ship?" is a score. "Did the new model help?" is a diff. "Which of these two prompts is better?" takes an afternoon, not a meeting.

What each discipline gains

Context engineering: the PM's real lever over quality

The question that actually predicts an AI feature's quality is "what context does the model need to make this decision meaningful?", not "what data do we have?" — and a clever prompt, on its own, answers that question badly: it's fragile to rephrasing, doesn't scale across use cases, and carries no memory forward between calls. The PM's job isn't to write the prompt — it's to specify what context a decision requires, the same way specs specify an eval threshold. This is developed in full in Context engineering for the product leader — the paradigm shift, the diagnostic for prompt-vs-context failures, and how to write a context requirement into a spec.

The data flywheel

The compounding asset of an AI product isn't the model — everyone can rent the same model. It's the loop: usage → captured feedback → better evals (and sometimes fine-tuning data) → safer, faster iteration → better product → more usage. Two PM jobs make it spin: instrument feedback capture as a launch requirement (accepted / edited / abandoned — the silent signals, since users rarely click thumbs-down), and secure the right to learn from it (privacy policy, enterprise contracts, regional law — a viability question from lesson 2 that quietly determines whether your flywheel is legal). A competitor with the same model and a faster flywheel wins in a year. That asymmetry — not model choice — is usually the real AI strategy question.

The flywheel's strategic twin is workflow capture. The defensible position isn't the model or even a beloved feature — it's owning a control point in the customer's workflow: the step whose outputs downstream work consumes and whose inputs upstream work feeds. From a control point you expand horizontally (each captured workflow adds enterprise context that makes the next one easier to win), and a broader bundle lets you cross-subsidize any component a point-solution competitor tries to commoditize. When you evaluate an adjacent workflow, the deciding question is directional: if another player captured it, would they integrate into us, or we into them?

The long game: models age, debt compounds

Three roadmap truths that only bite in the product's second year, so budget for them in its first:

Vibe coding: the prototype is the new spec

The 2025 shift with the biggest day-to-day impact on the PM craft: AI coding tools (Claude Code, Cursor-class editors) made working software cheap enough that PMs build it themselves — "vibe coding," Karpathy's February 2025 coinage that ended the year as Collins' word of the year. Three changes follow for this lesson's operating loop:

A closing practical note: use these tools in your own workflow. The PMs with the best model intuition use models daily — generating mock-up variations for design discussions, clustering and prioritizing bug reports, critiquing a UX flow, drafting the first pass of a PRD or stakeholder mail. It's the cheapest training available in where models shine and where they quietly fail — and that intuition is exactly what the eval bar and scoping decisions above run on.

Failure modes

Practitioner checklist