One project, many roles

What you will be able to do

See how AI roles divide and share the work on one real project. In a job interview and in your first week, the question is rarely “What is an ML engineer?” It is “Who does this, who decides, and who must be asked?”

Before you startWhat to know and to have readyAbout 35 minutes

What you need to know

• The six job families, from The main job families: Six job families.
• The roles around the core, from Technical and adjacent routes: The roles around the core.
• Larkfield’s document assistant, from AI in everyday products: The example: a document assistant at Larkfield.

What you need

• Your notes.md file in the ai-career folder. You made it in AI in everyday products: Set up your course notes. If it is missing, make it again as in that section. This lesson adds the heading ## Task cards.

The team

Larkfield decides to build the document assistant from AI in everyday products. These people work on it. All of them are fictional.

Person	Role	Main question that they ask
Ruth	Director of customer service	Is this worth the money and the risk?
Omar	AI product manager	What do we build first, and how do we know that it works?
Grace	Support lead, 11 years at Larkfield (subject-matter expert)	Is the answer correct, and will agents trust it?
Mei	Data engineer	Do we have the right documents, and do they stay current?
Kwame	Solution architect	Where does it run, what does it cost, is it secure?
Tomás	AI engineer	How do we get correct answers with sources from the model?
Sam	Software engineer, help desk tool team	How does it fit in the tool that agents use all day?
Priya	Data scientist	How good is it, measured, not guessed?
Lena	Privacy officer	What personal data does it touch, and is that legal?

Ruth is a stakeholder: a person or group who is affected by the project or can make decisions about it. Ruth does not build anything, but she pays for the project and can stop it. The 40 support agents are stakeholders too. They are the users.

Four kinds of responsibility

For each piece of work, the team agrees on four things:

• Does: the person who does the work.
• Decides: the one person who makes the final decision. Only one person decides, so a decision cannot get stuck between two people.
• Asked: people whose knowledge is needed before the decision.
• Told: people who must know after the decision.

Many teams use a version of this called a RACI chart (Responsible, Accountable, Consulted, Informed). The names change between companies. The idea is the same.

The project, phase by phase

The phases of Larkfield's project and who leads each one. Evaluation decides if the project goes forward or back.

Text description of the diagram

1. Frame the problem (weeks 1–2): led by the product manager and the support lead.
2. Prepare the data (weeks 2–3): led by the data engineer. The privacy officer is asked.
3. Design (week 3): led by the solution architect.
4. Build (weeks 3–5): led by the AI engineer and the software engineer.
5. Evaluate (weeks 4–5): led by the data scientist and the support lead.
6. Decision: is it good enough for a pilot? If no, the team goes back to Build, fixes it and tests again. If yes, the project goes to phase 6.
7. Pilot and operate (from week 6): led by the product manager and the support lead. When a problem is found, the team goes back to Build.

1. Frame the problem (weeks 1–2)

Omar meets Grace and four agents, and watches them answer questions for one morning. Grace brings numbers from her quality checks: an answer takes about four minutes, and about 1 answer in 12 quotes an old rule.

Omar writes a one-page problem statement. Grace corrects one sentence: agents do not need “an answer”, they need “the rule, with its source”. Agents get warnings when they promise a customer something wrong, so they will not trust an answer that they cannot check.

Output: the problem statement and two success measures: time per question under one minute, and old rules in fewer than 1 answer in 100. Decides: Ruth (go or no go). Asked: Grace, Kwame (rough cost). Told: the support team.

2. Prepare the data (weeks 2–3)

Mei finds the policy documents in three places: the support wiki, a shared drive and email attachments. She finds 61 files, not 40. Some are old versions, and two are drafts that never became official.

Mei cannot know which files are official. Grace can. Grace marks 40 files as current and official, and names the owner of each policy. Mei builds a pipeline that copies only those files every night.

Lena asks a question that nobody else asked: “Will agents paste customer names and addresses into the questions?” They will. So Lena sets two rules: questions are kept for 30 days only, and the tool reminds agents not to paste personal details.

Output: a pipeline of 40 current documents, and a privacy note. Decides: Mei (how the pipeline works), Lena (privacy rules). Asked: Grace. Told: Tomás.

3. Design (week 3)

Kwame compares two options: Larkfield runs the whole system in its own cloud account, or it uses a model provider’s hosted service. He chooses a provider that Larkfield already has a contract with, because the contract already covers customer data. He estimates the cost: about 4,000 questions a month, at a few cents each.

Output: a two-page design, with a cost limit for each month. Decides: Kwame. Asked: Lena, Omar, Sam. Told: everyone.

4. Build (weeks 3–5)

Tomás builds the assistant. It finds the right sections in the 40 documents and sends them to the model with the question. It asks for a short answer with the section number. If the documents do not cover the question, the assistant must say so instead of guessing.

Sam adds the panel to the help desk tool. Grace asks for one change after a test: show the date of the policy next to each answer. Agents trust a rule more when they can see that it is current.

Output: a working assistant in a test version of the help desk tool. Decides: Tomás (how the assistant works), Sam (how the panel works). Asked: Grace. Told: Omar.

5. Evaluate (weeks 4–5)

Priya needs a test set: realistic questions with known correct answers. She takes 120 real questions from last year. Grace writes the correct answer and the correct policy section for each one. Priya cannot do this part. Grace knows that “broken on arrival” and “stopped working after a month” go to two different policies.

Priya runs all 120 questions through the assistant. The results:

Result	Questions
Correct answer, correct source	106
Correct answer, wrong source	6
Says “the policy does not cover this”, correctly	3
Wrong answer	5

All 5 wrong answers are about refunds, and all 5 promise more than the policy allows. Grace marks them as serious: a promise to a customer costs money, and taking it back costs trust.

Decides: Omar, after a meeting with Grace and Priya: no pilot until the serious errors are fixed. Tomás changes how refund sections are found. Priya runs the test again: no serious errors, and 112 answers correct with the correct source. Told: Ruth.

6. Pilot and operate (from week 6)

Ten agents use the assistant for four weeks. Every answer has a “thumbs down” button. Every Friday, Grace and Priya read every answer that got a thumbs down, and Omar decides what to fix next.

The measures after four weeks: 50 seconds per question, and no answers with old rules. Ruth agrees to give the assistant to all 40 agents.

The week something goes wrong

In week 12, an agent gives the assistant a thumbs down on Monday morning. The assistant said that delivery to the islands is free. On 1 September, Larkfield started to charge for it.

Here is what each person does that week:

When	Who	What
Monday 09:10	An agent	Marks the answer as wrong, and writes: “Island delivery is not free since 1 September.”
Monday 10:00	Grace	Confirms that the agent is right. Tells all agents to check island delivery answers by hand until the problem is fixed.
Monday 11:30	Tomás	Finds that the assistant used the delivery policy from June. The September version is not in its documents.
Monday 14:00	Mei	Finds the cause: the new policy was saved in a new folder, which the pipeline does not copy. She adds the folder, and the pipeline takes the new version that night.
Tuesday	Priya	Adds three island delivery questions to the test set, so this mistake is caught next time. She runs the test set: all pass.
Tuesday	Grace	Searches the saved questions for island delivery: 14 customers got the wrong answer.
Wednesday	Omar	Decides two changes. First, the pipeline reports any new file in the policy folders. Second, each policy owner tells Mei before a change.
Wednesday	Ruth	Decides to honour the free delivery for the 14 customers, because Larkfield’s agents promised it.
Wednesday	Lena	Checks: no personal data was exposed. No action needed.

Notice four things:

1. A user found the problem, not a computer. The thumbs-down button and the Friday review are part of the product.
2. The model was not the cause. It wrote a good answer from an old document. The cause was in the data and in the process for changing policies.
3. The fix had a technical part and a human part. Mei fixed the pipeline. Omar changed the process, so policy owners now tell the team about changes.
4. The domain expert was needed at every step. Grace confirmed the error, warned the agents and found the 14 customers.

Roles are work, not people

Larkfield has a person for every role. A company of 12 people does not. There, the founder may be the product manager. One engineer may be the data engineer, the architect and the AI engineer. The best support person may be the SME and the evaluation specialist. An outside lawyer may answer the privacy questions.

The work in each phase is the same. When you join a small company, you may do three roles. When you join a large one, you may do one part of one role. In both, you must know who does, who decides, and who must be asked.

Guided practice: match the task cards

Here is a worked example.

Card 0: “Decide whether the assistant may answer questions about payments.”

• Decides: Omar, the product manager. It is a decision about what the product does.
• Asked: Grace (are payment answers risky for agents?), Lena (payments involve personal data and maybe special rules), Priya (how good are the answers today?).
• Overlap: three people give information, and only one decides.

Now match six cards. In notes.md, under ## Task cards, write who does each task, who decides, and one overlap that you see.

1. Write 120 past customer questions with the correct answer and the correct policy section.
2. Find why the nightly copy of the documents missed a file.
3. Decide whether all 40 agents can use the assistant next month.
4. Check whether customer names in the questions are stored, and for how long.
5. Make each answer cost half as much, with no fewer correct answers.
6. Choose between running the assistant in Larkfield’s own cloud account or at a model provider.

Show the answers

1. Does: Grace writes the answers. Priya chooses the questions and sets up the test. Overlap: the test set needs both, Priya’s method and Grace’s knowledge.
2. Does and decides: Mei. Overlap: Tomás often finds the symptom first, as in week 12.
3. Decides: Ruth, from Omar’s recommendation. Asked: Grace and Priya, with the pilot results. Overlap: Omar recommends, but Ruth owns the budget and the risk.
4. Does: Lena, with Mei, who knows where the data is stored. Decides: Lena. Overlap: a privacy question is also a data question.
5. Does: Tomás. Asked: Priya measures that the answers are still correct. Kwame checks the cost limit. Overlap: a cost change can be a quality change. It must pass the test set.
6. Decides: Kwame. Asked: Lena (data rules), Omar and Ruth (cost). Overlap: an architecture choice is also a legal and a budget choice.

Different answers can be correct if you explain them. For example, in some companies the product manager decides on card 3.

Independent variation: a dental clinic with three people (one dentist, one nurse, one receptionist) wants an assistant that answers patient questions about appointment rules. In notes.md, write who would cover each of the nine roles from the team table. Which roles must the clinic buy from outside, for example from a software company?

Common mistake

Everybody thought that somebody else owned it

Symptom: a problem continues for weeks. In meetings, each person says, “I thought the other team handled that.” At Larkfield, before week 12, nobody owned the question “Who tells the team when a policy changes?”

Cause: the work was divided by role, but the space between two roles had no owner. The policy owners thought the pipeline was automatic. Mei thought that every policy was saved in the same place.

Fix:

1. List every piece of ongoing work, not only the build work. Include “keep documents current” and “review bad answers”.
2. Give each piece one person who does and one person who decides.
3. When a problem crosses two roles, write down who owns the handoff.

Check your understanding

In week 5, the five refund errors were found. Why did Omar decide, and not Tomás, who fixed them?

Show the answer

The question was not “how do we fix it”, which is Tomás’s work. It was “is the product ready for real agents”, which is a product decision with business risk. Omar decided after he asked Grace (how serious are these errors?) and Priya (how often do they happen?). Tomás then decided how to fix them.

Check your understanding

Name two moments in this project where the project would have failed without Grace, the subject-matter expert.

Show the answer

Any two of these: she corrected the problem statement (agents need the rule with its source); she chose the 40 official documents out of 61; she wrote the correct answers for the test set; she marked the refund errors as serious; she confirmed the island delivery error and found the 14 customers. In each case, the technical people could not know the answer themselves.

Recap

• An AI project goes through phases: frame, data, design, build, evaluate, pilot and operate. Evaluation can send it back to build.
• For each piece of work, agree who does, who decides, who is asked and who is told.
• When something goes wrong, the cause is often in the data or a process, not in the model. The fix has a technical part and a human part.
• Roles are work, not people. In a small company, one person does several roles.
• The subject-matter expert was needed in every phase.

Next, in Module 2, you find your own starting point, starting with an honest list of your skills.

Next: Skills inventory