The AI inventory
- Lesson 8
- 30 min
- Module 3: Inventory, risk tiers and impact assessment
Where we are
In the last module, we saw what the laws and standards ask of an organisation. Every one of them starts from the same question: which AI systems do you have? Larkfield cannot answer it yet. The platform registry shows only four systems, and nobody has looked at the bills or at the tools it buys. In this lesson, we build the AI inventory, like the catalogue of a library.
Lesson 8 of 19 · AI Governance: Policies, Risk and AccountabilityThe AI inventory
1. What governance is for
2. Laws and standards in plain words
3. Inventory, risk tiers and impact assessment
4. Policies people can follow
5. Controls, documentation and evidence
What you will be able to do
Find every AI system that your organisation builds, buys or uses without telling anyone. Write each one down with the few fields that later decisions need.
Before you startWhat to know and to have ready
What you need to know
- Larkfield’s eight systems and their owners, from Meet Larkfield’s AI portfolio.
- What an accountable owner is, from Who is accountable.
- Why the rules that apply depend on the system, from A map of the landscape.
What you need
- A browser and a spreadsheet or a document tool.
- The inventory template and its spreadsheet copy, inventory.csv.
- The registry extract, the usage report and the portfolio folder with the eight system cards.
A library with no catalogue
Section titled “A library with no catalogue”Think of a small school library. Books arrive as gifts, teachers bring their own, and some books leave and never come back. Nobody wrote any of this down. Then a parent asks a simple question: “Do you have a book about first aid, and is it up to date?” Nobody can answer without searching every shelf.
The library needs a catalogue: one card per book, with the title, the shelf, who borrowed it and when it was last checked. The catalogue does not make the books better. But every other job starts from it: repairs, a list of old books, an answer for the parent.
The problem: you cannot govern what you cannot see
Section titled “The problem: you cannot govern what you cannot see”Go back to Farida’s first week. Farida, Larkfield’s new risk and compliance lead, asks Kwame for the list of AI systems. Kwame owns the platform controls. He opens the platform registry, the list of systems that call a model through Larkfield’s platform. It shows 4 systems and a total of US$9.33 for the month. Everything is within its budget. It looks complete.
Then Mei forwards a bill from another AI provider. Larkfield has no contract with that company. The key belongs to a contractor who has left. The line says US$63.40, and the service received 1,150 customer messages. It is the weekly digest script that the platform team found in a shared folder. The registry did not show it, because the script calls the AI service with its own key and never goes through the platform.
All of this is constructed for this course: Larkfield, its registry extract and the bill. The cost of one ticket or one question is recorded by earlier courses. The number of tickets and questions in a month is an assumption.
Text description of the chart
A column chart of one month of AI spend. The four registered systems: triage-service US$8.41, resolution-workflow US$0.32, doc-assistant US$0.60 and support-mcp US$0.00, because it calls no model. The unregistered script: US$63.40, in a different colour. It is 87% of all AI spend.
What this tells us. The biggest line is the one that the registry cannot see: 87% of all AI spend. But the money is small everywhere. The risk is not in the bill. It is in what the systems do, and where the data goes: 1,150 customer messages went to a company with no contract. And the bill shows only AI that somebody pays for separately. The AI features inside the tools that Larkfield buys appear on no AI bill at all.
The idea: an AI inventory
Section titled “The idea: an AI inventory”An AI inventory is a list with one row for each AI system that an organisation builds, buys or uses. It includes the systems that nobody registered. It is the library’s catalogue, for AI.
What counts as an AI system? The EU AI Act defines it as a machine-based system that infers, from its input, how to produce outputs such as predictions, content, recommendations or decisions (Regulation (EU) 2024/1689, Article 3, official source, checked 22 October 2026). The EU AI Act explains the law. For the inventory, a plain test is enough: does it write, summarise, sort, predict or decide from data? If yes, it gets a row. The legal question, whether the law calls it an AI system, comes later and needs the specialist.
Every row has one of three kinds:
- Built: a system that your teams made, such as Larkfield’s ticket triage service.
- Bought: an AI feature inside a tool that you pay for, such as a reply suggester that a helpdesk vendor added in an update.
- Unregistered: AI that somebody found but nobody registered, such as the digest script.
Each kind needs different checks. A built system has a team that can change it. A bought feature has a vendor and a contract. An unregistered system has neither, so the first question is whether it may run at all.
Where the rows come from
Section titled “Where the rows come from”No single source sees everything. Larkfield needs three, and each one finds what the others miss.
Text description of the diagram
The diagram shows three sources with arrows into one AI inventory of eight rows. First, the platform registry. It finds every system that calls a model through the platform, S1 to S4, and it misses AI inside bought tools and AI that uses its own key. Second, bills, contracts and company cards. They find bought tools and unknown AI services: the digest script on another provider’s bill, and the three bought tools in contracts and card payments. They miss free tools and AI features that a vendor switches on in an update. Third, asking each team one question: “Which tools do you use that write, summarise or decide?” This finds the rest.
The registry is the best source for built systems, because it fills many fields by itself. A course on governing AI agents builds such a registry. Each registered system has a record with its owner, purpose, risk tier, data classes, tools, autonomy and review date. So for S1 to S4, Farida copies most fields from the registry extract instead of asking the teams.
But a registry stores what it is told. The triage team registered its service as tier low. Nobody checked that against a method. The next lesson does, and the answer is different.
The fields that matter
Section titled “The fields that matter”A catalogue card has few fields, and each one answers a question that somebody will ask. The inventory works the same way. Every field must feed a later decision; a field that feeds none is extra work.
| Field | The decision it feeds |
|---|---|
| Kind: built, bought or unregistered | which checks apply: release gate, vendor check, or “may it run?” |
| Accountable owner, technical owner | who can say yes or no, and who can change or stop it |
| People affected | whose life a wrong result touches; they are not always the users |
| Data, vendor and model | which privacy rules and contracts apply, and where data goes |
| Autonomy: suggests, proposes with approval, acts alone | the risk tier, in the next lesson |
| In the registry | whether the platform fills and checks the row |
| Tier, last and next review | which controls apply, and which reviews are late |
The template has these fields and a few more, for example the purpose and the status. Keep them short: one line each.
A tiny example: one row for the document assistant
Section titled “A tiny example: one row for the document assistant”The template’s worked example is S3, the document assistant from Portfolio Capstone: Build, Evaluate and Ship an AI Product. Here is its row, with every value from the S3 system card and the registry extract (constructed):
| Field | Value | Where it comes from |
|---|---|---|
| Kind | built | the card |
| Purpose | answers staff questions from documents, with citations | the card, and the registry’s purpose |
| Accountable owner | Mei | the card; the registry agrees |
| Technical owner | Mei’s data team | the card |
| People affected | staff; customers indirectly, through the answers staff give | the card |
| Data | internal and staff-only documents, and the partner shop’s documents | the card, and the registry’s data classes |
| Vendor and model | Larkfield’s model provider, with a contract | the card |
| Autonomy | suggests: it answers a person, who decides | the card, and the registry |
| In the registry | yes | the registry extract |
| Last and next review | at launch; overdue | the card, and the registry: “overdue by 2 months” |
Even this simple row has a finding. The review is two months late. Nobody noticed, because nobody looked at the review dates of all systems together. One row in a shared list made it visible.
Larkfield: the rows that nobody had
Section titled “Larkfield: the rows that nobody had”The four built systems take minutes, because the registry and the cards agree on most fields. The other four rows take the real work.
- S5, the weekly digest script (unregistered). It has no accountable owner and no technical owner, and its key belongs to a contractor who has left. Each week, it sends customer messages to a company with no contract. Its row is the most urgent in the inventory: somebody must decide now whether it may run.
- S6, the reply suggester in the help desk (a bought add-on). TicketHarbor (made up for this course) added it in an update and switched it on for everyone. Nobody is named as its owner. The finance lead owns the add-on’s contract and did not know about the feature.
- S7, the meeting-notes tool (bought). MinuteMoss (made up for this course) records online meetings, sometimes with customers. Omar’s team bought it on a company card.
- S8, the invoice reader in the accounting tool (bought). LedgerFern (made up for this course) fills in supplier amounts. Its owner is a role, “the finance lead”, with no name, and nobody checked the vendor.
When the eight rows are together, two numbers show the state of governance at once. 4 of 8 systems are in the registry (50%). And 5 systems have no named person as accountable owner. S2 has a team, S8 has a role, and S4, S5 and S6 have nobody.
Discuss: does the meeting-notes tool count?
Section titled “Discuss: does the meeting-notes tool count?”Farida shows the eight rows at the team meeting.
- Tomás: “MinuteMoss is a note-taker, not an AI system. And it comes from a vendor, so the vendor is responsible for it.”
- Omar: “If every tool with some AI inside goes into the inventory, the list never ends. My spell checker uses AI too.”
- Mei: “MinuteMoss records calls with customers and suppliers. Those recordings go to the vendor, and to a model provider that it does not name.”
- Sam: “So how do we decide what gets a row?”
Before you read on, answer these:
- Does MinuteMoss pass the plain test: does it write, summarise, sort, predict or decide from data?
- Is Tomás right that the vendor is responsible?
- Is Omar’s worry real, and what would you do about it?
The course’s answer. MinuteMoss gets a row. It writes notes and a list of actions from recordings of people, and some of those people are customers. Tomás is wrong about responsibility. Canada’s federal privacy law for companies says that an organisation stays responsible for personal information that it sends to another company for processing (PIPEDA, Schedule 1, principle 4.1.3, official source, checked 22 October 2026). Canada, the US, the UK and others explains this law. The vendor is a reason to check the vendor, not a reason to skip the row.
Omar’s worry is real: an inventory that lists every spell checker is too long to keep current. So Larkfield sets a line. A tool gets its own row if it sends Larkfield’s data about people to an AI service. It also gets one if its output reaches customers or a decision. Small tools that do neither are covered by the staff rules on AI use, without a row each.
What would change the answer: suppose MinuteMoss ran only on one computer and kept no recording. Then it could fall under the staff rules instead. It would still need a row the day it starts to record customers.
When to keep an inventory, and what it costs
Section titled “When to keep an inventory, and what it costs”Start the inventory before any other governance work, because tiers, controls, audits and the board all read from it. The first version takes a few days: the registry for built systems, one month of bills and card payments, and one question to each team.
After that, the cost depends on the registry. Built systems update their own rows when they register or change. The other rows need a check of new bills every month and the question to each team twice a year. Do not turn the inventory into a long form: a field that no decision reads is a field that nobody keeps current.
Common mistake
An inventory that only lists what the platform sees
Symptom: the inventory has a row for every system in the registry, and the registry report says “all systems within budget”. But a bill shows an AI service that is in neither.
Cause: the inventory was copied from the registry only. The registry cannot see AI inside bought tools, or AI that uses its own key.
Fix: add the two other sources: bills and contracts every month, and the question to each team. Measure coverage as registered systems out of all known systems. At Larkfield, it is 4 of 8, not 4 of 4.
Guided practice: the first two rows
Section titled “Guided practice: the first two rows”-
Open your copy of
inventory.csvin a spreadsheet. The S3 row is already filled in. -
Open the system card S1 and the registry extract side by side.
-
Fill in the S1 row from the card: kind, purpose, owners, people affected, data, vendor and model, autonomy.
-
Fill in “in the registry” and the next review from the registry extract.
Check: S1 is built, acts alone when it routes tickets, is in the registry, and its next review is due in 4 months.
-
Compare the owners. The card says that Grace is accountable. The registry says that the owner is Amira.
-
Write Grace as accountable owner and Amira as technical owner. Add a note: “the registry names the technical owner”.
-
Leave the tier empty. The next lesson sets it.
When two sources disagree, write what each one says and mark the row. A disagreement is a finding, and the owner of the row decides which source is right.
Your turn: complete Larkfield’s inventory
Section titled “Your turn: complete Larkfield’s inventory”Fill in the six other rows: S2, S4 and S5 to S8. Use the system cards, the registry extract and the usage report. Leave the tier for the next lesson.
Check your rows
- In the registry: yes for S1 to S4; no for S5 to S8.
- Kind: built for S1 to S4; unregistered for S5; bought for S6, S7 and S8, with the vendor’s name.
- No named person as accountable owner: S2 (a team), S4 (nobody), S5 (nobody), S6 (nobody), S8 (a role). That is 5 rows. Do not invent an owner here; write the gap.
- Autonomy: S1 acts alone (routing); S2 proposes and a person approves each action; S4 proposes refunds and an owner approves; S3, S5, S6 and S7 suggest or summarise; S8 fills in amounts and a clerk checks and posts each one.
- Data that leaves Larkfield without a contract: S5 only. For S6, S7 and S8, write “the vendor’s choice” for the model, because the cards do not say more.
- Reviews: S2, S4 and S5 to S8 were never reviewed; S3 is overdue.
Knowledge check
Section titled “Knowledge check”Check your understanding
Larkfield's registry lists four systems, and all of them are within budget. Name two kinds of AI system that the registry cannot see, with one Larkfield example of each.
Show the answer
AI that calls a model with its own key, around the platform: the weekly digest script, found on another provider’s bill. AI features inside tools that Larkfield buys: the reply suggester in the helpdesk tool, the meeting-notes tool and the invoice reader. They do not call the platform, and they appear on no AI bill. A registry only stores what is registered with it.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
The unregistered script is 87% of Larkfield's AI spend. Why is the money not the main finding?
Show the answer
The whole bill is small: US$63.40 for the script and US$9.33 for everything registered. The risk is in what the script does. It sent 1,150 customer messages to a company with no contract, with the key of a person who has left. A cheap system can carry the biggest risk, so the inventory records data and owners, not only cost.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Omar proposes an inventory field 'number of users'. Which test should a field pass before you add it, and does this one pass?
Show the answer
A field must feed a later decision. “Number of users” does not decide much on its own. One user who sends customer data to an unknown company matters more than a hundred users of a spell checker. The field “people affected” already feeds the tier. So add the number of users only if a decision will read it, for example a licence count for the vendor review.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Recap
- An AI inventory has one row for each AI system that you build, buy or use, including the ones that nobody registered. Every other governance task reads from it.
- Use three sources: the platform registry for built systems, bills and contracts for bought and unknown services, and one question to each team for the rest.
- Keep only fields that feed a decision: kind, owners, people affected, data and vendor, autonomy, registry, tier and reviews.
- At Larkfield, the registry saw 4 of 8 systems, and 5 systems have no named accountable person. The biggest risk sits in the cheapest line.
You can now list every system and see what each one touches. But the list treats a meeting-notes tool and a system that routes every ticket alone in the same way. Next, you give each system a risk tier, so that the checks match the harm.