Team roles by stage

What you will be able to do

Write down who decides, who does the work and who is consulted at each stage of an AI product. Without this map, work stalls between people, and the person who built a change is often the only one who checks it.

Before you startWhat to know and to have readyAbout 25 minutes

What you need to know

• The four stages and their gates: feasibility, pilot, limited launch and full launch, from Milestones and roadmaps.
• The evaluation set and the definition of done, from Define done with evidence.
• Helpful, not required: what each AI job does all day, from The main job families in AI Career Foundations: Roles, Skills and Roadmaps.

What you need

• The course files, from Read the evidence pack. In this lesson you use the template templates/role-map.md. Copy its text into your own document.

Larkfield, its people and its role map are made up for the courses.

A kitchen renovation

A family renovates its kitchen. At the start, they talk to a designer, and they measure the room. Later, a plumber and an electrician do their work, and a painter comes at the end. An inspector, who did not do the wiring, checks it before the walls close. The family decides every big question: the budget, the plan, the day the walls close.

Two things make this work. First, the people change from one stage to the next: nobody needs the painter in week one. Second, at every stage, everybody knows who decides. When the electrician and the designer disagree about a socket, the family decides, and the work goes on. And the electrician never signs his own inspection.

The problem: work that nobody owns

An AI product also passes through stages, and it also needs different people at each one. If nobody writes down who does what, work falls between people. You only see the gap later, when it is expensive.

The course’s third case study is a project like that. A sponsor started a machine-learning project to warn customers about late deliveries. After six months, the status reports were still green, and nothing had launched. Here is what three of the people said at the end:

• Omar, the product manager: “It’s not on my roadmap. I thought operations owned it.”
• Priya, the data scientist: “Every month the test set or the measure changed, so the numbers are not comparable.”
• Luis, the delivery team lead: “Nobody from the project spoke to us.”

The quotes come from a project file made up for the course. No one in them is lazy: each person did their own work. But nobody owned the decisions after the sponsor left, and nobody owned the evaluation set. And nobody asked the people who would use the warnings. Those are three missing roles, not three missing skills.

The idea: a role map

A role is a piece of work with its responsibility, not a person. One person can hold several roles, and one role can need several people. A role map is a table with the roles down the side and the stages across the top. Each cell holds one of three words:

• Decides: the one person who makes the final call at that stage. Only one, so a decision cannot get stuck between two people.
• Does: the people who do the work.
• Consulted: the people whose knowledge is needed before the decision.

Many companies use a longer version of this idea, called a RACI chart. The career course explains it in One project, many roles. The role map adds one thing that a project plan often misses: the stage. The person who decides at feasibility is often not the person who decides when real customers see the product.

A tiny example

A three-person shop wants an assistant that drafts answers to customer emails. Ana owns the shop, Ben answers the emails today, and Chris builds the website. Here is their map for the first two stages:

Role: person	1. Feasibility	2. Pilot
Owner: Ana	decides	decides
Domain expert (knows the answers): Ben	does: writes 30 example emails and their right answers	does: uses the drafts and marks each one
Builder: Chris	does: tries the assistant on the 30 emails	does: connects it to the inbox
Evaluator: Ben	does: scores Chris’s drafts against his answers	does: counts the drafts that he sent unchanged

Read the map column by column, then row by row. Each column has exactly one “decides”. Ben appears twice, as the domain expert and as the evaluator. That is fine in a small shop. The important part is that the person who builds the assistant (Chris) is not the person who scores it (Ben).

Larkfield’s role map

Larkfield’s support reply assistant needs more roles. In her first week, Nadia, the new head of customer operations, asks Omar for the map. Omar has none, so they build it together from the template. The diagram shows the result. It is constructed for the course from the jobs that these people hold in the earlier courses.

Larkfield's role map for the support reply assistant (constructed). Each column has one person who decides, and the people change from stage to stage.

Text description of the diagram

The rows are the roles of the template, each with a person’s name, plus a business owner. Feasibility: Omar (product lead) decides. Priya (data scientist and evaluator) and Tomás (AI engineer) do the work. Nadia, Grace, Mei and Kwame are consulted. Pilot: Omar decides. Priya, Tomás, Mei and Amira (developer) do the work. Nadia, Grace and Kwame are consulted. Limited launch: Nadia decides. Omar, Grace, Tomás, Mei, Priya as evaluator and Amira (developer and operations) do the work. Priya as data scientist and Kwame are consulted. Full launch: Nadia decides. Omar, Grace, Mei, Priya as evaluator and Amira as operations do the work. Priya as data scientist, Tomás, Amira as developer and Kwame are consulted. The ML engineer row is empty at every stage.

Two people hold two roles each, so their names appear twice in some columns. Priya does the evaluation at every stage, and as the data scientist she is consulted at both launch stages. Amira runs the service at both launch stages, and as the developer she is consulted at the full launch.

Three cells in this map are choices, and each has a reason:

1. Omar decides at the first two stages, and Nadia at the launch stages. Before launch, the question is “Does this product work?”, and that is the product lead’s question. At launch, the agents’ work and the customers’ trust are at risk, and both are in Nadia’s part of the business.
2. Grace does work at both launch stages. She owns the user promise and leads the agents who use the drafts. So she owns adoption: whether the people who should use the product really use it, and use it well. Change and adoption showed why adoption needs an owner.
3. The ML engineer row is empty. An ML engineer trains and serves models. Larkfield uses a hosted model with a written prompt, so it trains nothing today. If Larkfield ever runs its own model, this row fills, and the map changes.

Now check the map with the five questions at the end of the template. Every column has exactly one “decides”. The domain expert is consulted or does work at every stage. Priya owns the evaluation set, and Tomás, whose changes it judges, does not. Grace owns adoption. And Amira answers when the service fails at night. The course MLOps and LLMOps: Operating AI in Production made her the owner of the service, in Ownership and support.

Why the builder does not judge

The map keeps two roles apart on purpose: the person who builds a change, and the person who evaluates it. They are the electrician and the inspector again.

When one person builds, tests and recommends, nobody checks the choice of examples. Three roles, kept apart, put a second pair of eyes between the change and the decision.

Text description of the diagram

In the top row, one person does every job. They build the change, choose the examples, test the change on them, read the result and recommend shipping. Nobody else sees the evidence before the decision. In the bottom row, the jobs are split. The AI engineer builds the change. The data scientist evaluates it on the evaluation set that the team agreed, labelled with the domain expert’s policy. One owner decides at the gate, from the written evidence: continue, change or stop.

A builder is not dishonest. But a builder chooses the examples that the change was made for, and reads the results with hope. Larkfield saw the value of a separate judge with prompt v3. Omar wanted the cheaper-looking prompt shipped. But the release rule was written before the test, on an evaluation set that Priya kept apart. It found that prompt v3 failed all six of its checks, and it stopped the release. Define done with evidence lists those six checks.

Discuss: who decides at the limited launch?

Nadia’s first gate is near: the assistant’s limited launch, on a small share of real tickets. Before she accepts the map, the team gives its views.

• Omar: “I am the product lead. I decided at feasibility and at the pilot, so I should decide at launch too. It’s one product.”
• Grace: “My agents carry the risk. If a safety ticket reaches nobody, my team hears from the customer. I should decide.”
• Tomás: “I’ll own the evaluation set from now on. I know the prompts better than anyone, so I can keep the set up to date.”
• Kwame: “Put me down as consulted at every stage. The provider contract and the cost are my area, but I don’t need to decide.”

Before you read on, answer these:

1. Who should decide at the limited launch, and why only one person?
2. Should Tomás own the evaluation set?
3. What does Grace need, if she does not decide?

The course’s answer. Nadia decides at both launch gates. The launch puts customer operations at risk, and that is her area. One person decides, so that a disagreement ends with a decision, not with a delay. Omar still does the product work, and his evidence goes to her. Tomás should not own the evaluation set, because it judges his own changes. Priya owns it, and Grace’s policy labels it. With prompt v3, a gate that the people who wanted the change did not control stopped a release that failed all six checks. Grace does not need the decision. She needs two things: to be consulted before every gate, and the right to pause the launch when a safety ticket reaches nobody. Write that pause condition into the definition of done.

What would change the answer: in a team of five, one person holds several roles, and the founder may decide at every stage. The builder-and-judge rule still holds there. Borrow a reviewer from another team, or ask the domain expert to score the examples.

When to use a role map, and what it costs

Use it when a product moves from one stage to the next, when someone joins or leaves, and before every gate. The map is short, so it is quick to keep current. A map that is six months old is as bad as none: in the stalled project, the sponsor left, and nobody took his decisions.

Do not use it as an organisation chart, or to rank people. It says who does what for one product, not who reports to whom.

What it costs: about an hour to make, with the people in the room, and five minutes to check before each gate. The real cost is the conversation that it forces. Two people who both think that they decide find out now, not at the gate.

Common mistake

Two people decide

Symptom: at the pilot gate, Omar says “go” and Grace says “not yet”. Nobody knows whose answer counts, so the team waits two weeks for a meeting.

Cause: the pilot column of the map has “decides” in two cells, or the map does not exist.

Fix: put one “decides” in each column, and make the other person “consulted”, with a written way to stop the work. For example: Omar decides, and Grace can pause the pilot if a safety ticket reaches nobody. Write the stop rule in the definition of done.

Common mistake

A role without a person

Symptom: the map has a row for operations, but its cells are empty. At 2 a.m. the assistant stops answering, and the first person to notice is a customer.

Cause: the team copied the roles from the template, but did not put a name in every row that has work.

Fix: a row with work needs a name at every stage where it has work. If nobody can take it, write “nobody yet” and the date by which it must have a name. Larkfield’s ML engineer row says “nobody yet” because it has no work today, not because someone forgot it.

Your turn

Copy templates/role-map.md into your document. Fill it for an AI product in your own work, or for the three-person shop above, for all four stages. Then check it with the template’s five questions. Then change one condition and fill it again: the company decides to train its own model. Which row fills, and from which stage?

Show a worked solution for the shop

Role: person	1. Feasibility	2. Pilot	3. Limited launch	4. Full launch
Owner: Ana	decides	decides	decides	decides
Domain expert: Ben	does	does	does (owns adoption)	does
Builder: Chris	does	does	does	consulted
Evaluator: Ben	does	does	does	does
Security and privacy: an outside adviser	–	consulted	consulted	consulted
Operations: Chris	–	–	does	does

All five checks pass. One person decides in each column (Ana). Ben, the domain expert, works at every stage. The evaluator (Ben) is not the builder (Chris). Ben owns adoption, and Chris answers when it fails. The outside adviser is new: customer emails hold personal data, so someone must be consulted before real emails go through the assistant.

If the shop trains its own model, the ML engineer row fills from the pilot onward. Chris probably cannot do that work alone, so it is a role to borrow or to hire. The next lesson shows how to choose.

See it end to endRecover a stalled AI projectCase study

The third case study gives you the stalled project’s full file. You find the empty cells of its role map and write a 90-day recovery plan with stage gates.

Knowledge check

Check your understanding

Why does each column of a role map have exactly one person who decides? Give an example of what goes wrong with two.

Show the answer

One person decides, so that a disagreement ends with a decision. With two, each can block the other, and the work waits. For example, Omar says “go” at the pilot gate and Grace says “not yet”, and nobody knows whose answer counts. The fix is not to ignore Grace. Make her consulted, with a written right to pause the work on a named condition, such as a safety ticket that reaches nobody.

Check your understanding

Tomás offers to own the evaluation set because he knows the prompt best. What do you answer, and why?

Show the answer

No. The evaluation set judges Tomás’s own changes, so the builder would choose and read the test of his own work. He would not cheat, but he would choose the examples that his change was made for. Priya owns the set, with Grace’s policy labels, and Tomás can propose new cases to her. A gate that the people who wanted the change did not control stopped prompt v3, which failed all six release checks.

Check your understanding

Larkfield's ML engineer row is empty at every stage. Is that a mistake? What would change it?

Show the answer

It is not a mistake. Larkfield uses a hosted model with a written prompt, so nobody trains or serves a model, and the role has no work. The row says “nobody yet” on purpose. It changes if Larkfield decides to run its own model, for example a fine-tuned small model for routing. Then the row fills from the pilot onward, and the team must borrow, hire or grow that skill.

Recap

• A role is a piece of work with its responsibility, not a person. A role map lists the roles by stage, with who decides, who does the work and who is consulted.
• One person decides in each column. At Larkfield, Omar decides at feasibility and the pilot, and Nadia at both launch stages (constructed).
• The builder does not judge: Priya owns the evaluation set, not Tomás. Grace owns adoption, and Amira answers when the service fails.
• An empty row is fine when it has no work (the ML engineer). A row with work and no name is a gap.
• The stalled project shows the cost of no map: a sponsor left, and nobody took his decisions.

Next, you fill a role that the map shows as missing. You write a role profile from the work, and you choose whether to hire, grow or borrow.

Next: Hire and grow skills