Product craft and leadership12 min read

The Modern Product Manager's Competency Profile

A practical competency model for product managers working with AI, with observable behaviours, role-specific targets and a way to map strengths.

A product toolkit brings together customer understanding, technical judgement and leadership.
On this page
  1. 1.The ten competencies
  2. 2.A working technical baseline
  3. 3.Build fluency through real work
  4. 4.Rate behaviour, not confidence
  5. 5.Make your profile
  6. 6.Map the team as well as the person
  7. 7.Try it: make a practice plan

Two product managers are considering the same role. One is excellent at finding customer problems and bringing people together, but has never inspected an agent trace. The other builds convincing prototypes in a day, but cannot explain which customer would change their behaviour to use one.

Their strengths and gaps matter differently depending on the role. A team needs to understand which capabilities the work calls for, where each person can contribute and where they need help.

The model below describes ten competencies and a common scale for comparing your current profile with the needs of your role. Use it to guide development and discuss team coverage, not as an automated hiring decision.

The ten competencies

CompetencyThe question it helps you answerEvidence to look for
Customer and market understandingWhose problem is this, and why does it matter now?Observed work, customer conversations, alternatives and evidence of demand
Strategy and commercial judgementWhich outcome is worth pursuing, and can the business sustain it?A choice between bets, pricing or cost analysis, and an explanation of what the team will not pursue
Discovery and experimentationWhat do we need to learn before committing?A testable hypothesis, appropriate experiment, contradictory evidence and a decision that changed
Product definition and experienceCan people understand and use what we are building?A clear problem frame, prototype or specification, interaction test and design trade-off
Model and system fluencyHow does the AI system produce an answer or take action?A diagram or trace explaining the model, context, tools, state, permissions and failure points
Evaluation and qualityHow will we know whether the system is good enough?Representative cases, a quality rubric, failure analysis and a release or monitoring decision
Building and technical collaborationWhat can we make to learn, and where does specialist depth matter?A working artefact with stated limits, a technical discussion, or an inspected and corrected failure
Operating judgement and riskWhat can the system do, and how do we recover when it goes wrong?An authority boundary, review plan, escalation path, owner and recovery procedure
Delivery and influenceCan the people involved make a good decision and carry it through?A decision record, resolved dependency, launch plan and evidence of follow-through
Team leadership and learningAre the right capabilities present, and can the team keep improving?Coaching, shared context, a decision about missing expertise or a change that reduced unnecessary work

The categories overlap. A customer interview might reveal a product risk, change an eval and alter the commercial case. Use the distinctions to spot strengths and gaps rather than argue over which column owns a piece of work.

A working technical baseline

A product manager needs enough direct experience with AI systems to ask meaningful questions and recognise when an answer is incomplete. Training a model or owning production infrastructure is a different level of technical responsibility.

For the workflow they own, a useful baseline is the ability to:

  1. Explain the path. Describe what reaches the model, what context is supplied, which tools it can use, where state is kept and what happens to its output. A harness is the surrounding system that manages these interactions and constraints; the model is one part of it.
  2. Work through examples. Run representative tasks, inspect the result and, where available, the trace. Distinguish a model error from missing context, a failed tool, an unclear instruction or an incorrect expectation.
  3. Define quality with others. Help select real tasks and meaningful edge cases, specify what a good result looks like, inspect disagreements and understand what an eval does and does not establish.
  4. Discuss the operating trade-offs. Explain the consequences of errors, review effort, latency, cost, permissions and recovery. Know which decisions need engineering, design, security or domain expertise.
  5. Make something to learn. Create or configure a bounded prototype, test or analysis appropriate to the problem. State what it simulates and what would need to change before real use.

This is a working standard, not a claim that every PM must personally implement every component. A person can become technically fluent by inspecting, experimenting and making decisions alongside specialists. Building production systems, researching models and owning security controls require their own depth.

A PM who relies on AI to summarise research has different responsibilities from one whose product lets an agent change customer records. The latter needs a much deeper understanding of permissions, eval coverage, operational failure and recovery. Set the target from the work and its consequences.

Build fluency through real work

Start with a task you understand well enough to review. Compare the result with your usual approach, keep examples of success and failure, and inspect what caused the difference. Was a source missing? Did an instruction leave an ambiguity? Did a tool fail, or was the expected answer itself wrong? Changing the prompt will not fix every problem.

The responsibility changes when others rely on your method. A personal analysis may become a shared workflow that needs an owner, representative tests, maintained context and a way to report failures. Redesigning how a team works adds questions of access, training, review capacity and evidence of value. These are different scopes of work, not ranks that everyone must pass through.

Practise with colleagues when the task is unfamiliar or consequential. Explaining why a result is correct is part of learning to judge the next one.

Rate behaviour, not confidence

Use the same four levels for each competency. Assess against the evidence you have, and mark an area as not yet observed when you do not have enough. A lack of opportunity is not proof of a lack of ability.

LevelWhat it means
1. DevelopingUnderstands the basic concepts and can contribute with guidance. Needs help to recognise important gaps or select an approach.
2. WorkingHandles a representative task in the current role, explains the reasoning, checks the result and knows when to ask for help.
3. LeadingHandles ambiguous or difficult work, makes trade-offs across functions, improves the approach and helps others apply it.
4. ShapingEstablishes effective practice across multiple teams or complex contexts, makes hard boundaries visible and develops other practitioners.

These levels describe demonstrated work, not seniority. Someone can be shaping practice in discovery and still developing in model systems. A senior title does not make every axis a four. Likewise, someone can show deep craft without managing a portfolio or holding a particular financial title.

Apply the scale to an eval

Consider a team testing an assistant that drafts responses to support requests. An evaluation should cover whether the response follows the current policy, uses the right customer context and flags cases that need a person.

  • Developing: Reviews a small set of examples with a teammate and can describe why a result is wrong or uncertain.
  • Working: Helps choose representative cases, records expected behaviour, examines failures and uses the results in a decision about a bounded pilot.
  • Leading: Works with domain experts to resolve disagreement, identifies important gaps in coverage, and connects release criteria to monitoring and escalation.
  • Shaping: Helps several teams establish maintainable evaluation practices, calibration and ownership, while allowing different workflows to have different quality thresholds.

The thresholds and release decision depend on the actual workflow. The evaluation chapter explains how to build and maintain the suite.

Recognise progress in other areas

The same scale can describe different kinds of work. The examples below are prompts for discussion, not a requirement to perform every activity in every role.

LevelModel and system fluencyCustomer understandingOperating judgement
DevelopingWith a teammate, follows one request through the model, supplied context and a tool; identifies a point of uncertainty.Joins a customer conversation and separates what the person said or did from an interpretation.Recognises a consequential action in a workflow and asks who can approve or stop it.
WorkingInspects a trace, explains plausible sources of an error and works with an engineer to test a cause.Investigates a recent task, compares it with an alternative and uses the evidence to frame a problem.Defines a task's permitted actions, handoff and recovery with the people responsible for the system.
LeadingCompares system approaches against the task's quality, latency and cost; helps the team decide what to test or change.Resolves conflicting signals across research and behaviour, then changes a product bet or its scope.Sets a proportionate release boundary for an ambiguous workflow and makes sure monitoring and escalation are owned.
ShapingHelps multiple teams establish ways to inspect system behaviour and choose appropriate technical approaches.Builds a practice for bringing customer evidence into decisions across teams and develops other practitioners.Improves how teams assign authority and respond to failures across several workflows, adapting controls to their consequences.

Look for a decision the person made and what happened afterwards. For example, to practise model fluency, take a permitted trace of a wrong answer, identify what reached the model and compare at least two plausible causes with an engineer. To practise customer understanding, observe a task and write down one observation that challenges the team's proposed solution. To practise operating judgement, trace a proposed record change through authorisation, approval and recovery with the system owner. These exercises give a reviewer something concrete to discuss; completing one does not by itself establish a level.

Make your profile

Start with the ten rows above. For each one, write down:

  • Your current level, or not yet observed.
  • The level your current role requires in the next six to twelve months.
  • One piece of evidence and any uncertainty about it.
  • One action that would improve the capability or put the missing expertise alongside you.

Compare the two profiles in a table first. A radar chart can make their shape visible, but its area is not a meaningful total. Do not average the axes, rank people by the size of the shape or treat a self-assessment as a hiring or promotion decision.

Map your profile

Choose a current level and a target for each area. Leave either blank when there is not enough evidence. Your choices stay in this page and reset when you leave.

12345678910

Blue solid: current. Amber dashed: target. An outline appears once all ten levels in that column are assessed. The numbers correspond to the rows below.

1. Customer and market understanding

Current: Not yet observed; target: Not yet observed

2. Strategy and commercial judgement

Current: Not yet observed; target: Not yet observed

3. Discovery and experimentation

Current: Not yet observed; target: Not yet observed

4. Product definition and experience

Current: Not yet observed; target: Not yet observed

5. Model and system fluency

Current: Not yet observed; target: Not yet observed

6. Evaluation and quality

Current: Not yet observed; target: Not yet observed

7. Building and technical collaboration

Current: Not yet observed; target: Not yet observed

8. Operating judgement and risk

Current: Not yet observed; target: Not yet observed

9. Delivery and influence

Current: Not yet observed; target: Not yet observed

10. Team leadership and learning

Current: Not yet observed; target: Not yet observed

Use the shape to discuss evidence and development. It is not a score or a hiring decision.

A worked profile

A generalist PM working on an established service might set a level-three target for customer understanding and delivery, and a level-two target for model and system fluency. The PM is using AI to explore research and prototype interactions, but the product itself does not make autonomous decisions. Their immediate development action is to inspect the provenance and failure modes of an AI-assisted research workflow with an analyst.

An AI-product PM responsible for an agent that updates customer records may need level-three capability in evaluation and operating judgement as well as the working baseline across the other areas. They still need engineering and security partners to own the controls that sit within those disciplines.

A technical platform PM may place greater weight on model systems and evaluation. A product leader may need deeper team leadership and commercial judgement while retaining enough direct technical fluency to question the evidence. These profiles are illustrations, not universal targets.

Map the team as well as the person

A team profile answers a different question: is the work covered? List the consequential activities, the person accountable for each one and the specialists who contribute. Check the evidence that the activity is happening.

For example, a PM may frame the quality decision, a domain specialist may define unacceptable outcomes, and an engineer may build the test infrastructure. If nobody owns the decision to pause a failing agent, a combined team score will not reveal the gap.

Use the profile in coaching and planning. Pick a small number of gaps that matter to current work, create an opportunity to practise, ask for feedback, and revisit the evidence. The team design chapter helps decide whether to develop a capability, bring in a specialist or change how work is shared.

Try it: make a practice plan

Choose one or two areas where the needs of your current work exceed the evidence in your profile. Pick a real task you can attempt with appropriate support. Agree with a colleague what you will do, what evidence you will keep, how they will give feedback and when you will review it. For example, inspect one permitted failure trace with an engineer and explain two possible causes before looking at the fix.

Prompt
Help me design a small practice plan for a product competency.

My current work: [describe the task and my responsibility]
Capability I want to develop: [choose one]
Evidence of my current level: [examples or not yet observed]
Support available and constraints: [people, time and access]

Suggest a task I can practise in my work, what I should attempt myself,
what feedback to request and what evidence would show progress. Offer a
smaller alternative if the task depends on access I do not have. Do not
infer my ability from my job title or treat one exercise as a rating.

Choose a plan with the colleague who can support it. After the work, record what you understood, where you needed help and what you would do differently. Update the profile from demonstrated work rather than from the assistant's assessment.

v3.1 · Updated Oct 2026