Skip to content
Français

Who is accountable

  • Lesson 2
  • 30 min
  • Module 1: What governance is for

Where we are

In the last lesson, we set four goals for governance, and one of them was clear owners. But a refund workflow touches support, finance, privacy and the customer. If everyone is a little responsible, nobody is. In this lesson, we give each system one accountable owner and fill a responsibility table, starting with a kitchen that three friends renovate together.

Plays from YouTube. Nothing loads from YouTube until you press play. Privacy

What you will be able to do

Give every AI system one accountable owner, and write down who approves, who does the work, who is asked and who is told. When something goes wrong, you then know who answers within minutes, not after a week of meetings.

Before you startWhat to know and to have readyAbout 30 minutes

What you need to know

What you need

Three friends share a flat, and they renovate the kitchen together. One of them found the builder. One of them paid the deposit. One of them was at home when the plumber came. Two months later, water drips through the ceiling of the flat below.

The neighbour knocks and asks: “Who is in charge of this?” Each friend points at another one. The first found the builder, but did not check the work. The second paid, but never saw the pipes. The third let the plumber in, but thought the others had agreed on the plan. Everyone was a little responsible, so nobody can answer.

Now think of a house that one owner renovates. The owner decides and pays, and answers to the neighbour. A builder does the work. An electrician is asked about the wiring before the walls close. A building inspector signs off the wiring, and the owner cannot sign it alone, because the owner wants the job finished. The neighbours are told when the noise starts. Each person has a clear part, and one person answers.

Larkfield’s support-resolution workflow proposes refunds, replacements and return labels. A person approves each action before it runs. Its system card (constructed) names the accountable owner as “Support team”.

Imagine a customer complains that the workflow refused a fair refund (constructed for this lesson). Who answers the complaint? Who decides whether the rule was wrong? Who can switch the workflow off while the team looks? The card names a team of many people. Each team lead can say, truthfully, that the decision was not theirs alone. This is the kitchen again.

The accountable owner of an AI system is the one named person who answers for its results and can say “stop”. “Accountable” comes from “account”: the person who must give the account of what happened, and why. The owner does not do all the work. The owner makes sure the work happens, and answers when it goes wrong.

Three rules make the owner real:

  • One person, not a team. A team cannot answer a question or take a decision. A person can.
  • A name, not a job title. “The finance lead” is empty on the day that person leaves. Write the name, and change it when the job changes hands.
  • Someone with the power to act. The owner must be able to stop the system and to change its rules. An owner who must ask permission to say “stop” is not the owner.

The approver is a different role. An approver says yes or no to one specific action or change before it happens. The two are often different people, for the same reason as in the house. The person who wants a result should not be the only check on it. An expense claim works the same way: you do not approve your own.

Around the owner and the approver, more people have a part. The picture shows them for one system.

One AI system in the middle. On the left, two groups point to it. Decide: the board, which sets the rules and decides high-risk requests; the accountable owner, one named person who answers for the results and can say stop, with a heavy border; the approver, who says yes or no to one action or change. Build, advise and check: the technical owner, who builds and runs it; risk, with the method and the challenge; legal, for questions about the law; privacy, for personal data and its records; security, for access and attacks. On the right, the system points to a third group: the people who use it, staff who read its output, and the people it affects, customers and suppliers, who need a way to complain.
The roles around one AI system. Only the accountable owner (heavy border) answers for all of it.
Text description of the diagram

The picture puts one AI system in the middle. Two groups on the left point to it. The first group decides. The board sets the rules and decides the high-risk requests. The accountable owner, with a heavy border, is one named person who answers for the results and can say stop. The approver says yes or no to one action or change. The second group builds, advises and checks. The technical owner builds and runs the system. Risk brings the method and challenges the answers. Legal answers questions about the law. Privacy looks after personal data and its records. Security looks after access and attacks. On the right, the system points to the people who use it, staff who read its output, and to the people it affects, such as customers and suppliers. The people it affects need a way to complain.

Two groups are easy to forget. The people who use the system are the agents and clerks who read its output. They see its mistakes first, so they need a simple way to report them. The people it affects are often outside the organisation: the customer who gets a refund decision, the supplier who gets a payment. They are never asked before a decision, but they need a way to complain and to have a wrong result corrected.

A responsibility table lists the decisions about one system, and who does what in each. Many organisations use a table called RACI (responsible, accountable, consulted, informed). This course’s table adds one column, the approver, because AI governance often turns on the difference between owning and approving. Five letters:

Letter Meaning
A accountable Says yes or no, and answers for the result. Exactly one named person per row.
P approves Approves a specific action or change before it happens.
D does Does the work.
C consulted Asked before the decision, and listened to.
I informed Told after the decision.

Farida starts with the simplest system, the document assistant. It answers staff questions from Larkfield’s own documents, and a person decides what to do with each answer. Mei (she), the data engineer who owns Larkfield’s data records, runs it. Farida fills the table in four steps.

  1. List the decisions. Farida writes the five decisions about this system. They are a launch or major change, new or removed documents, a model or prompt change, a wrong answer to a customer, and the review.
  2. Write one A per row. Mei answers for all five. She owns the documents and the system, so the A column is “Mei” five times.
  3. Write the approver. Mei approves a launch and a model change. The data team lead approves new documents, a small and frequent change. Nobody approves an incident or a review in advance, so those cells get a dash.
  4. Fill D, C and I. The data team does the work. Farida is consulted on big changes, and Kwame (he) on the platform. Grace (she) is consulted about policy documents and about wrong answers to customers, because she owns the promise to customers.
A table with five rows and five role columns, for the document assistant. Launch or a major change: A Mei, P Mei, D data team, C Farida and Kwame, I support team leads. Add or remove documents: A Mei, P data team lead, D data team, C Grace, I staff who use it. Change the model or the prompt: A Mei, P Mei, D data team, C Farida, I staff who use it. A wrong answer reached a customer: A Mei, P none, D data team, C Grace, I Farida. Review the system: A Mei, P none, D data team, C Farida, I governance board. The A column has heavy borders and the P column is shaded.
The responsibility table of the document assistant (constructed), as in the course's template. One name in every A cell.
Text description of the diagram

The table has one row per decision and five role columns. For a launch or a major change, Mei is accountable and approves. The data team does the work, Farida and Kwame are consulted, and the support team leads are informed. To add or remove documents, Mei is accountable, the data team lead approves, the data team does it, Grace is consulted and the staff who use it are informed. To change the model or the prompt, Mei is accountable and approves, the data team does it, Farida is consulted and the staff are informed. When a wrong answer reached a customer, Mei is accountable, nobody approves in advance, the data team does the work, Grace is consulted and Farida is informed. For the review, Mei is accountable, the data team does it, Farida is consulted and the governance board is informed.

Now Farida checks the table, row by row. Every A cell has exactly one name, and it is a person, not a team. Then she reads the A and P columns together. Mei is both accountable and approver for two rows. Is that a problem?

Here it is not. The document assistant is low risk with the method of Risk tiers and proportionate controls, later in the course. It decides nothing about customers, and Mei owns its data. A second approver would add waiting and no safety. For a system that acts on customers, such as a refund, the approver of an action is usually not the accountable owner.

Guided practice: two rows for the workflow

Section titled “Guided practice: two rows for the workflow”

Open the support-resolution workflow’s card. Fill only the first two rows of a new table: “Launch, or a major change” and “Approve an action that affects a customer”.

  1. Read the card’s “Accountable owner” line. Decide whether you can copy it.
  2. Choose one named person for the A column. Write one sentence of reason.
  3. Read the card’s “Who approves changes and actions” line.
  4. Fill the P column of both rows.
  5. Fill D, C and I.
  6. Check: one name per A cell, and a person in every cell where you can name one.
Show one possible answer

You cannot copy “Support team”: it is a team. Grace is the best accountable owner, because she owns the refund policy and the promise to customers.

Decision A P D C I
Launch, or a major change (for example more autonomy) Grace the governance board Amira, the developer Farida, Mei, Kwame team leads
Approve an action that affects a customer Grace the agent or team lead within their limit; Grace above it Amira builds the approval step — the customer, by the reply

Grace is accountable on both rows, but she does not approve every refund. The agent or team lead approves each action at the moment it happens, within a limit. A major change, such as more autonomy, goes to the board, because the owner who wants the change should not be its only check.

Farida shows the document assistant’s table at a team meeting.

  • Tomás: “Why is legal not the A on every row? Legal knows the law.”
  • Kwame: “I built the platform under three of these systems. Should I be their owner?”
  • Grace: “If I am the owner of the workflow, do I approve every refund myself?”
  • Sam asks: “Why does the customer appear in the table at all? They do not decide anything.”

Before you read on, answer these four questions.

The course’s answer. Legal is consulted, not accountable. A lawyer can say what a law asks, but cannot answer for how support uses a tool every day. Kwame is the technical owner: he answers for the platform working. The business decision, such as “should this system refund money?”, belongs to the person who owns that business. Grace owns the workflow but approves only what goes above the agents’ limit, as the guided practice showed. And the customer is in the table because the customer is affected. Being “informed” is the minimum, and the incident and complaint rows need a way to correct a wrong result.

What would change the answer: in a ten-person company, the founder may be the owner of every system. That is fine if it is written down, and if a second person approves the actions that touch customers’ money.

Write a responsibility table for every system that touches customers, staff or money, and keep it on the system’s record. A low-risk system needs five short rows; it takes about as long as reading this section. Update it when a person changes job, the same day.

The cost is a little time and one hard conversation: somebody must agree to be the owner. That conversation is the point. An owner who did not agree will not act.

Common mistake

An owner who is a committee

Symptom: a system makes a visible mistake. A meeting is called, then a second one. Three days later, the system is still running, and nobody has told the affected customers.

Cause: the A column says “AI working group” or “Support team”. A group needs a meeting to decide, and nobody may decide alone.

Fix: write one name in every A cell. Give that person the power to stop the system without a meeting. The group can still be consulted (C). It cannot be accountable.

Check your understanding

What are the three rules that make an accountable owner real? For each one, give a Larkfield entry that breaks it.

Show the answer

One person, not a team: the workflow’s card says “Support team”. A name, not a job title: the invoice reader’s card says “the finance lead”. Someone with the power to act: an owner who must wait for a committee before stopping the system breaks it. At Larkfield, Kwame would break it as owner of the support MCP server. He can stop its technology, but he cannot change the refund policy that it serves: that policy is Grace’s.

Check your understanding

Mei is both accountable and approver for the document assistant’s model changes. Predict whether Grace should be both for each refund the workflow proposes, and explain why.

Show the answer

No. The document assistant is low risk and decides nothing about customers, so one person is enough. A refund acts on a customer’s money. Grace owns the refund policy and wants the workflow to work, so she should not be the only check on each action. The agent or team lead approves each refund within a limit, and Grace approves the ones above it.

Check your understanding

In your own organisation, or one you know, name one tool or system and write the A and P for its launch. What would make the two different people?

Show the answer

Answers differ. A good answer names one person, not a team, for A, and an approver with a reason. The two should differ when the system acts on customers, money or personal data, because the person who wants the launch should not be its only check. They can be the same for a low-risk tool that decides nothing about people.

Recap

  • The accountable owner is one named person who answers for a system’s results and can say “stop”. It is a person, not a team, and a name, not a job title, with the power to act.
  • The approver says yes or no to one action or change. For actions on customers or money, it is usually not the owner.
  • Around them: the board, the technical owner, risk, legal, privacy, security, the people who use the system and the people it affects.
  • A responsibility table lists each decision with A, P, D, C and I. One name in every A cell.

You can now fill a responsibility table for one system. But Larkfield has eight systems, and some of their cards are wrong or empty. Next, you open the portfolio pack and the one-page policy, and you find what nobody owns.

Next: Meet Larkfield's AI portfolio