Your portfolio
Most AI PM candidates can talk about AI. Fewer can show a document that demonstrates AI-specific decisions. One strong artifact closes that gap immediately.
An AI PM portfolio is not a collection of project descriptions — it is a collection of decisions. Interviewers look for evidence that you made AI-specific choices, documented trade-offs, and can quantify outcomes. Select an artifact type to see what it must demonstrate.
End-to-end PM judgment: problem scoping, AI-specific decisions, evaluation strategy, outcome ownership
One explicitly documented trade-off (what you descoped and why), one AI-specific decision (confidence threshold, autonomy level, eval approach)
The Acid Test
Can a hiring manager see exactly what decisions you made and why? If the artifact could have been written by a PM with no AI knowledge, it is not demonstrating AI PM craft. The AI-specific section — threshold decision, autonomy level, eval strategy, fallback design — is what interviewers are looking for. Everything else is context.
The AI case study is the highest-signal artifact in your portfolio. Six sections, 400–600 words total. Each section has a specific job — and a specific way to get it wrong. Select a section to see good vs. bad examples.
"Claims processors spent 23 minutes manually routing each claim, handling 80 per day, with a 12% misrouting rate causing SLA breaches on £500+ claims."
"Our team wanted to improve the claims routing process using AI."
Specific numbers (23 minutes, 80/day, 12%, £500+) make the problem concrete and testable. Vague problems produce vague solutions and unmeasurable outcomes.
Four templates for the most common AI PM artifacts. Each is designed to force the decisions that signal AI PM craft — not describe what was built, but document why it was built that way.
PRODUCT REQUIREMENTS DOCUMENT — AI Feature Problem Statement [Who experiences it] + [what the pain is] + [frequency] + [cost/consequence] User Persona Role: [job title, company type] Context: [workflow stage where pain occurs] Primary pain: [one sentence] Current State How it's done today: [current workflow] Where it breaks: [failure modes, error rate, time cost] Proposed Solution What the feature does: [in plain language] Autonomy level: [Assist / Review / Monitor / Full — with definition] What it does NOT do: [explicit scope boundary] Why AI (not rules or manual) Data available: [source, volume, label quality] Pattern learnable: [evidence baseline can be beaten] Baseline: [current accuracy / human accuracy] ── MODEL REQUIREMENTS ────────────────────────────── Input schema Field name | Type | Source | Required | Notes ... Output format Structure: [JSON schema or field list] Constraints: [length, format, value range] Confidence signal: [how confidence is surfaced] Confidence threshold: [value] Below threshold action: [escalate / fallback / queue] Expected queue rate: [% of inputs below threshold] Fallback behavior [What happens when model declines or errors] Evaluation dataset Source: [where it comes from] Size: [n examples] Labeling method: [who labels, IAA target] Update cadence: [quarterly / per sprint] Quality targets Precision ≥ [X]% Recall ≥ [X]% Latency p99 ≤ [X]ms ──────────────────────────────────────────────────── Success Metrics Primary (outcome): [business metric + threshold] Secondary (leading): [model or user behavior metric] Guardrail (must not): [existing metric this must not degrade] Out of Scope [Item 1] — [one-line rationale] [Item 2] — [one-line rationale] Open Questions [Question] | Owner: [name] | Due: [date]
How you walk through an artifact is as important as the artifact itself. Most candidates describe what the feature does. Strong candidates lead with why the AI-specific decisions were made. Use this five-step structure.
Lead with the business problem
Start with the pain and the business cost — not the technology. "Claims misrouting was costing 23 minutes per claim and causing SLA breaches" is more compelling than "we built a routing model."
Name one AI-specific decision
Identify the decision that proves you understood AI — not just product. Confidence threshold, autonomy level, eval dataset design, fallback path. This is where interviewers lean in.
State the explicit trade-off
"We descoped X because Y — and the cost of that was Z." Trade-offs reveal judgment. If you have no trade-offs to describe, either the project was trivial or you were not in the decision-making room.
Quantify the outcome
Specific numbers: 94% accuracy, AHT from 23 to 7 minutes, misrouting rate from 12% to 6%. If you don't have numbers, say why (product was early stage, org hadn't set up tracking) — but never vague success language.
State one learning
"Underestimated labeling cost by 2x. Next time: 50-example pilot before committing." Self-aware, specific, has a fix. This is the part that signals seniority — junior PMs describe what happened, senior PMs describe what they changed.
Common Presentation Mistakes
Starting with the technology ("we used a transformer model") instead of the business problem
Spending 80% of the time describing the feature and 20% on decisions and outcomes — should be the reverse
Describing a successful outcome without acknowledging any risk or learning — signals naivety, not competence
Saying "the team decided" without clarifying your specific role in the decision
Presenting hypothetical work as if it were real — interviewers ask follow-up questions that expose this immediately
Portfolio questions test your ability to narrate decisions, not just describe outcomes. The differentiating moves: naming a specific AI decision, stating an explicit trade-off, and quantifying an outcome against a pre-defined threshold.
Do you have any artifacts you can walk me through?
Yes. I can walk you through a PRD I wrote for a claims routing classifier.
The most interesting decision was the confidence threshold. We had to choose between a higher threshold (fewer errors but more human review) or a lower threshold (more automation but more misrouting). I calculated the cost of each error type — misrouting cost $40 in rework, over-review cost $4 in agent time — so the error cost asymmetry told me to favor precision.
We also descoped auto-close despite being technically feasible, because regulatory rules required a human touchpoint for claims above £500. That was the trade-off: 60% automation instead of 90%, but full compliance.
Outcome: 94% routing accuracy, AHT from 23 to 7 minutes, misrouting rate 12% to 6%.
What's the difference between a PRD and a model spec?
A PRD defines the product — the problem, user, solution, scope, and success metrics. Written for engineers, designers, and stakeholders. Includes why AI is the right approach, but stays at the product level.
A model spec defines the model — exact input schema, output format and constraints, confidence threshold, fallback behavior, evaluation dataset, quality targets, and monitoring triggers. Written for data scientists. It's the contract between PM and data science: what "done" means, measured how, against what data.
An AI PM writes both. The PRD can't substitute for the model spec — it's not precise enough for a data scientist to build against. The model spec can't substitute for the PRD — it has no user context, scope, or business outcome.
Can I include a hypothetical project in my portfolio?
Real is always better. Real projects have genuine constraints that forced real trade-offs — and that's what interviewers are evaluating. A hypothetical where you control all the variables can't demonstrate judgment under real pressure.
If you must use hypothetical work, impose real constraints explicitly: "Assume you have 18 months of data at 60% label quality, a 6-week commitment to a client, and no MLOps infrastructure." Then make decisions against those constraints. The constraint realism is what makes it credible.
The minimum bar: even one real artifact from a side project, open-source contribution, or volunteer work beats five hypothetical documents.
What do hiring managers actually look for when reviewing an AI PM portfolio?
Three things, in order.
First: evidence of AI-specific decisions. Did you make a threshold decision? Define an eval dataset? Document a fallback path? If the document could have been written by a PM with no AI knowledge, it fails this test.
Second: trade-off documentation. What did you NOT build, and why? Trade-offs show you were making decisions, not just executing. A document with no trade-offs is a description, not a PM artifact.
Third: outcome ownership. Can you quantify the result? Not "the model performed well" — but specific numbers against pre-defined thresholds. Hiring managers at AI companies have seen hundreds of vague case studies. Specific numbers are rare and memorable.
I don't have any AI product artifacts yet. How do I build a portfolio from scratch?
Four paths.
Side project: pick a real problem you have personally (not hypothetical), and document the PM process using the templates. Even a small project with real data and a real eval produces a credible artifact.
Reframe existing work: if you have traditional PM experience, write the model spec for what an AI version of that feature would have required. Show you can think through the AI-specific decisions even if you haven't shipped them.
Open-source AI: find an open-source AI project and write a PRD for a missing feature — real codebase, real constraints, real users.
The fastest path: write one detailed case study using the 6-section format for any project — even a failed one. A failed project with documented learning is more credible than a vague success story.