Skip to content
Français

Meet Larkfield's AI portfolio

  • Lesson 3
  • 30 min
  • Module 1: What governance is for

Where we are

In the last lesson, we gave the document assistant one accountable owner and a responsibility table. That was the easy system. Larkfield has seven more, and some of their cards are wrong or empty. In this lesson, we open the portfolio pack and the one-page AI policy, and we find what nobody owns, like a new tenant with a bunch of unlabelled keys.

What you will be able to do

Open the course’s practice files, check that you have all of them, and read Larkfield’s AI portfolio and its one-page policy. Find the owner of each system, and find what nobody owns. This is the setup check of the course.

Before you startWhat to know and to have readyAbout 30 minutes

What you need to know

What you need

  • The course’s folder C33 in the labs repository, open in your browser.
  • Your document tool, with two copies of the responsibility table template.

You move into a flat. The last tenant left a bunch of eight keys on the table, with no labels. You try each door. Some keys open a door, and you write the name on a label. One key opens nothing that you can find. Then you find a door that has no key at all, and the building manager does not know who has it.

At the end of the day, the useful result is not the keys that work. It is the short list of problems: the key that opens nothing, and the door that nobody can open.

  1. Open the folder C33 in your browser.
  2. Open the folder portfolio, then the file README.md.
  3. Count the systems in its first table. You see eight, from S1 to S8.
  4. Open the folder policy, then the file ai-policy-as-found.md. You see six numbered lines.

If you see fewer than eight systems or six lines, download the files again. To work offline, use the Download the files button above: it gives you the folder C33 in one ZIP file. To start again at any time, download a fresh copy; nothing in the folder changes by itself.

Farida found this policy on her first day. It is constructed for the course, with gaps on purpose. Read the six lines:

  1. Staff may use AI tools to work faster.
  2. Do not put confidential information into AI tools.
  3. New software, including AI tools, is approved by IT.
  4. For customer data, follow the privacy policy.
  5. AI output must be checked before use.
  6. Questions: ask your manager.

The header says: “Owner: not stated. Last review: not stated.”

Answer one question in one or two sentences: who may approve a new AI tool that reads customer messages? Say which line you used. If the policy does not answer the question, say what is missing. Write your answer before you read on.

Show the course’s answer

By the words of line 3, IT approves it. But that answers a smaller question than the one asked. IT can judge the software: does it install, is it secure? Nobody is named to approve the use of customer messages, and line 4 only points to another policy. A tool that reads customer messages needs a business owner who accepts the risk (Grace, for support), a privacy check (Mei) and the risk lead (Farida). The policy is missing who approves the data use, and a list of the approved tools.

If your answer was “IT”, you read the policy correctly. The point of the check is to see that the policy does not answer the real question.

The picture marks the gap in every line.

Seven rows. Each row joins a part of Larkfield's policy, on the left, to its gap, on the right in red. Header, owner and last review not stated: nobody keeps the policy current. Line 1, staff may use AI tools to work faster: which tools, and do AI features inside bought tools count? Line 2, do not put confidential information into AI tools: what is confidential, and does it apply to every tool, even approved ones? Line 3, new software including AI tools is approved by IT: IT judges the software, nobody approves the use of customer data; this gap has a heavy border. Line 4, for customer data follow the privacy policy: it points elsewhere, who checks and when? Line 5, AI output must be checked before use: by whom, which outputs, what proves it? Line 6, questions, ask your manager: who answers the manager?
Larkfield's one-page AI policy (constructed) with its gaps marked. The gap of line 3 (heavy border) is the answer to the entry check.
Text description of the diagram

The picture has seven rows. Each row joins a part of the policy to its gap. The header names no owner and no review date, so nobody keeps the policy current. Line 1 lets staff use AI tools to work faster, but does not say which tools, or whether AI features inside bought tools count. Line 2 forbids confidential information in AI tools, but does not say what is confidential, or whether it applies to approved tools too. Line 3 says that IT approves new software. IT judges the software, but nobody approves the use of customer data; this gap has a heavy border, because it answers the entry check. Line 4 points to the privacy policy, but does not say who checks, or when. Line 5 asks that AI output be checked, but does not say by whom, which outputs, or what proves it. Line 6 sends questions to a manager, but does not say who answers the manager.

None of the six lines is wrong. Each one is a good intention with nobody behind it. You rewrite this policy later in the course. For now, keep the gap of line 3 in mind: approving software is not the same as approving what the software does with people’s data.

The portfolio pack lists what the platform team found. Four systems were built by Larkfield’s teams in earlier courses. One script was never registered. Three are AI features inside tools that Larkfield buys, from vendors made up for this course: TicketHarbor (an add-on to the help desk), MinuteMoss (meeting notes) and LedgerFern (the accounting tool). One name may be new. The support MCP server is a connector for AI applications. It gives them three tools: read a ticket, search the knowledge base and propose a refund.

Three groups of systems. Built by Larkfield, in green, all in the registry: S1 ticket triage service, S2 support-resolution workflow, S3 document assistant, S4 support MCP server. Found, never registered, in orange: S5 weekly digest script. AI inside tools Larkfield buys, in orange, none in the registry: S6 reply suggester from TicketHarbor, S7 meeting-notes tool from MinuteMoss, S8 invoice reader from LedgerFern.
Larkfield's AI portfolio (constructed): eight systems, and four of them in the platform registry. The vendors are made up for this course.
Text description of the diagram

The picture shows three groups. Built by Larkfield, in green because all four are in the registry: ticket triage, the support-resolution workflow, the document assistant and the support MCP server. Found and never registered, in orange: the weekly digest script. AI inside tools that Larkfield buys, in orange because none is in the registry: the TicketHarbor reply suggester, the MinuteMoss notes tool and the LedgerFern invoice reader. Four of the eight systems are in the registry.

The registry is the platform team’s list of the systems that run on Larkfield’s own platform. That is why only the built systems are in it. The bought features and the script never passed through the platform, so the registry cannot see them. A list that only one team keeps sees only that team’s systems.

Each system card has the same fields and ends with four questions. Farida reads the card of the ticket triage service with them. The card says (constructed): accountable owner Grace; approver Grace; technical owner Amira (she), the developer; data: customer messages, order numbers, names; “routes every ticket alone”.

Question on the card What Farida finds Problem?
1. Is the owner one named person who can say yes or no? Grace, by name. She owns support. no
2. Who approves a change, and who approves an action on a customer? Grace, for both. not yet clear: the service routes every ticket alone, so no person approves each action
3. What data goes where, and is there a contract? Customer data goes to the model provider, under Larkfield’s contract. no
4. What would a wrong result cost, and who would notice? A ticket sent to the wrong queue may wait, or reach nobody. yes: the card’s note says the registry lists it as “low” risk

So the owner of the ticket triage service is clear. The open point is different: a system that acts alone on every ticket is labelled low risk. Farida notes it and moves on. Later in the course, in Risk tiers and proportionate controls, you check that label with a method.

Open the eight cards from the portfolio README. For each system:

  1. Read the “Accountable owner” line.
  2. Mark it person if it names one person.
  3. Mark it team if it names a group.
  4. Mark it role if it names a job title with no name.
  5. Mark it nobody if the field is empty.
  6. Count the systems that are not marked “person”.
Show the course’s answer
System Owner as written Mark Proposed owner, and why
S1 Ticket triage service Grace person Grace
S2 Support-resolution workflow Support team team Grace: she owns the refund policy
S3 Document assistant Mei person Mei
S4 Support MCP server nobody named nobody Grace: its only business effect is the refund proposal. Kwame stays the technical owner.
S5 Weekly digest script nobody named nobody Nadia, until the script is stopped or adopted
S6 Reply suggester (TicketHarbor) nobody named nobody Grace: it writes to her customers
S7 Meeting-notes tool (MinuteMoss) Omar person Omar, but nobody checked what happens to recordings of people outside Larkfield
S8 Invoice reader (LedgerFern) the finance lead role the finance lead, by name

Five of the eight systems have no named person as accountable owner: one team, one role and three empty fields. Three of the five are not in the registry either.

Look at where the gaps are. The three systems with a named owner are the ones that one team built or bought on purpose. The gaps are at the edges between teams. A server sits under someone else’s workflow. Nobody admits to the script. A vendor switched a feature on, and its contract sits in finance.

The weekly digest script sends customer messages, with names and order numbers, to an AI service with no contract. The key belongs to a contractor who has left. Farida asks who should own it.

  • Omar: “Nobody. Delete it today, and the problem is gone.”
  • Grace: “My team leads read that summary every Monday. It helps them see complaints early.”
  • Kwame: “If I own it, I can move it onto the platform and the contracted model.”
  • Sam asks: “Can a script be owned by nobody for a week while we decide?”

Before you read on, answer these:

  1. Who should own the script this week, and why?
  2. Is Omar’s answer wrong?

The course’s answer. Somebody senior must own it now, because a decision cannot wait: customer data goes to a company with no contract every week. Nadia owns it until it is stopped or adopted. Her first decision is probably to stop it, which Omar wants. But deleting it removes the evidence. It also hides the need that Grace’s team leads have. So Nadia stops it and keeps a copy of the script and its records for the review. Then she asks Grace whether a registered system should do the same job. If it should, Grace becomes its owner and Kwame its technical owner.

What would change the answer: if the AI service had a contract with Larkfield, the danger would be smaller. Even then, the script would still need an owner.

Independent task: two responsibility tables

Section titled “Independent task: two responsibility tables”

Fill a full responsibility table for two systems: the support-resolution workflow and the reply suggester in the helpdesk tool. Use the six rows of the template’s empty table. The reply suggester is the harder one: think about who approves its first switch-on, who approves each suggested reply, and what happens when TicketHarbor changes its model.

Check your tables. Every A cell has one name. Every cell names a person, not a team, wherever you can. Every row where A and P differ has a reason.

Show one possible answer

Support-resolution workflow

Decision A P D C I
Launch, or a major change (for example more autonomy) Grace the governance board Amira 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
Change the model or the prompt Grace Grace, after the release gate passes Amira Farida team leads
Handle an incident Grace — Amira, team leads Farida, Mei Nadia, the board
Answer a complaint Grace — a team lead Farida the board (monthly)
Review the system Grace — Amira Farida the board

Reply suggester in the helpdesk tool (TicketHarbor, made up for this course)

Decision A P D C I
Switch it on, or a major change Grace the board, for the first switch-on (a vendor check is needed) IT (the helpdesk administrator) Mei, Farida, the finance lead (contract) agents
Approve a suggested reply Grace the agent who sends it the agent — —
The vendor changes the model Grace Grace IT Farida agents
Handle an incident Grace — IT, team leads Farida, Mei the vendor, the board
Answer a complaint Grace — a team lead Farida —
Review the system and the vendor Grace — IT Farida, the finance lead the board

Your names can differ and still be right. Mark the reasons: one person in each A cell, and an approver who is not the owner where the action reaches a customer.

Do this owner check once for the whole portfolio, then for every new system on the day it arrives. Ask the finance team for the list of software subscriptions too: the bought features hide there.

It costs an afternoon for eight systems. The expensive part is not the reading. It is the conversations that follow, because each “nobody” becomes a request to a busy person.

Common mistake

The list that only shows what one team sees

Symptom: the registry shows four systems, all with owners. The team reports “100% of AI systems have an owner”.

Cause: the registry only sees the systems that run on the platform. The script and the three bought features never passed through it.

Fix: build the list from more than one source: the registry, the software subscriptions, the vendor contracts, and a question to every team lead. Count coverage against that full list: here, four of eight.

Check your understanding

Name the accountable owner and the approver for a new AI feature, and explain why they are not the same person. Use the reply suggester that TicketHarbor switched on.

Show the answer

The accountable owner is Grace. She owns the promise to customers, and she answers for every reply that the agents send. The approver of each reply is the agent who sends it; the approver of the first switch-on is the board, after a vendor check. They are not the same person for two reasons. The person who wants a result should not be the only check on it. And one owner cannot read every reply, so the approval sits where the action happens.

Check your understanding

The invoice reader’s card says the owner is “the finance lead”. Why is that a problem, and how do you fix it?

Show the answer

It is a job title, not a name. When the finance lead leaves or changes job, the system has no owner, and nobody notices. Write the person’s name, and update it the day the job changes hands. The card also shows no vendor check, so the named owner’s first task is to ask LedgerFern (made up for this course) what it does with suppliers’ bank details.

Check your understanding

By the policy as written, IT approves a new AI tool that reads customer messages. Why is that not enough?

Show the answer

IT can judge the software, but nobody approves the use of customer data. A tool that reads customer messages needs a business owner who accepts the risk, a privacy check and the risk lead’s view. The policy should name who approves the data use and list the approved tools.

Recap

  • Larkfield has eight AI systems: four built, one unregistered script and three features in bought tools. Only four of eight are in the platform registry.
  • Five of eight have no named person as accountable owner: one team, one role and three empty fields. The gaps sit between teams.
  • The one-page policy has a good intention in every line and nobody behind any of them. Approving software is not approving the use of customer data.
  • A system that sends customer data to a company with no contract needs a senior owner today, before anyone decides its future.

In this module, you learned what governance is for and who answers for each system. But Farida still does not know which rules Larkfield must follow, in Canada and in the European Union. In Module 2, you map the laws, standards and frameworks, and see which ones apply.

Next: A map of the landscape