Risk tiers and proportionate controls
- Lesson 9
- 35 min
- Module 3: Inventory, risk tiers and impact assessment
Where we are
In the last lesson, we built Larkfield’s AI inventory: eight rows, one for each system. But the inventory treats a meeting-notes tool and a service that routes every ticket alone in the same way. If both get the same checks, the checks are too slow for one or too weak for the other. In this lesson, we give each system a risk tier and the controls that match it, as a city does with food inspections.
Lesson 9 of 19 · AI Governance: Policies, Risk and AccountabilityRisk tiers and proportionate controls
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
Give each AI system a risk tier from four plain questions. Then a low-risk tool gets a quick yes, and a high-risk system gets a close look. Then read the controls that each tier needs.
Before you startWhat to know and to have ready
What you need to know
- Larkfield’s inventory, from The AI inventory. You need the autonomy, the data and the people affected of each system.
- That some laws put a use in a category of their own, from The EU AI Act and Canada, the US, the UK and others.
What you need
- Your inventory from the last lesson.
- The risk-tier table, copied into your document tool.
Food inspections
Section titled “Food inspections”A city’s food inspectors cannot check every kitchen every week. So they do not visit a shop that sells sealed chocolate bars as often as a restaurant that serves raw fish. A hospital kitchen that feeds patients gets the most visits of all.
The inspectors ask a few questions about each place. How bad is a mistake: a stale biscuit, or food poisoning? Who eats there: healthy adults, or sick people and children? Does anybody check the food before it is served? And what does the kitchen handle: packets, or raw meat? The answers put each place in a group, and each group has its own number of visits.
The problem: the same checks for every system
Section titled “The problem: the same checks for every system”Without groups, a team has two bad choices. It can check everything lightly, or check everything closely.
Light checks for everything fail on the systems that matter. Larkfield’s ticket triage service was registered as tier low by its own team, so it got light checks. In the “last week” of MLOps and LLMOps: Operating AI in Production, a new prompt went to every customer, and no rule made it pass the release gate first. 10 tickets that needed a person reached nobody. The answers are recorded by that course; the situation is constructed.
Close checks for everything fail in the other direction. If the meeting-notes tool needs a board decision, a security review and a three-month review, the team waits for weeks. People then stop asking, and they buy tools on a company card. That is how MinuteMoss (made up for this course) arrived at Larkfield.
The idea: risk tiers and proportionate controls
Section titled “The idea: risk tiers and proportionate controls”A risk tier is a label, low, medium or high, that says how closely a system must be checked. Proportionate controls are the checks that each tier needs: few for low, more for medium, the most for high. A control is anything that reduces a risk, such as a person who approves each action, a test before each release or a review date.
If you took Leading AI Products and Teams, you scored risks by impact × likelihood: low 1 to 4, medium 5 to 9, high 10 to 25. Here, we score systems. The inventory has one row per system, so the tier goes in that row.
This course uses its own method. It is constructed for the course: one reasonable method, not a law or a standard. It has four factors, the food inspector’s questions for AI:
- Harm of a wrong result: 1, small and easy to undo; 2, money or time lost, and it can be corrected; 3, safety, legal rights or access to a service, or hard to undo.
- People affected: 1, only staff; 2, customers, suppliers or other people outside the company; 3, vulnerable people, or a decision about a person’s access to a service, a job, credit or insurance.
- Autonomy: 1, it suggests or summarises and a person decides; 2, it proposes an action and a person approves each one; 3, it acts alone.
- Data: 1, no personal data; 2, personal data such as names and messages; 3, sensitive personal data, or data sent to a company with no contract.
Add the four scores. A sum of 4 to 6 is low, 7 to 8 is medium, and 9 to 12 is high. Two rules then change the result:
- A harm of 3 is never low. One severe harm must not hide behind three small scores.
- A system that acts alone with a harm of 2 or more is high. Nobody can stop a wrong refund that nobody sees first.
Before you score, ask two questions. Is it a use that a law bans or calls high-risk, for example hiring, credit or education? The EU AI Act lists such areas in its Annex III (Regulation (EU) 2024/1689, official source, checked 22 October 2026). Does it make a decision about a person based only on automated processing? A yes to either question goes to the legal specialist first, whatever the score. The method sorts the work; it does not answer legal questions.
A tiny example: the document assistant
Section titled “A tiny example: the document assistant”The template’s worked example is S3, the document assistant. Every value comes from its inventory row:
| Factor | Score | Why |
|---|---|---|
| Harm | 1 | A wrong answer goes to a member of staff, who can check the cited document. |
| People | 1 | Only staff use it and read its answers. |
| Autonomy | 1 | It answers a person; the person decides what to do. |
| Data | 2 | Staff-only documents can contain personal data. |
| Sum | 5 | 4 to 6 is low. Neither rule applies. |
The tier is low, so the document assistant needs the 4 low-tier controls and a review once a year. The whole path, from the inventory row to the controls and back, is one flow:
Text description of the diagram
The flow starts with an inventory row: its purpose, the people affected, its autonomy and its data. First come two questions. Is it a use that a law bans or calls high-risk? Does it make a decision about a person based only on automated processing? If yes, ask the legal specialist first. If no, score the four factors from 1 to 3. Add them, for a sum from 4 to 12, and apply the two rules. The result is the tier: low for 4 to 6, medium for 7 to 8, high for 9 to 12. The tier gives the controls: 4 for low, 9 for medium, 12 for high. The last step sets the review date, in 12, 6 or 3 months, and writes it back into the inventory row.
The controls of each tier
Section titled “The controls of each tier”Each tier keeps the controls of the tiers below it, except the ones it replaces with a stricter one. A medium system loses the fast path and the yearly review, and gets a review every 6 months. A high system replaces the owner’s impact assessment with a full one that the board reviews. So the list grows from 4 to 9 to 12 lines.
| Tier | The controls that it adds | The lower-tier controls that it replaces |
|---|---|---|
| Low | Listed in the inventory with a named accountable owner; Acceptable-use rules apply; Fast path: the owner and the risk lead decide within 2 working days; Review once a year | none |
| Medium | A system card, kept current; A vendor check before buying or switching on; An evaluation before each release, with the result in a decision record; A person approves every action that affects a customer; A change log; If it affects customers: an impact assessment, signed by the accountable owner (no board review); Review every 6 months | Fast path: the owner and the risk lead decide within 2 working days; Review once a year |
| High | A full impact assessment, signed by the accountable owner and reviewed by the board; A release gate with written thresholds; Monitoring with pause conditions, and an incident plan; A privacy or legal review where personal data or the law is involved; Review every 3 months | If it affects customers: an impact assessment, signed by the accountable owner (no board review); Review every 6 months |
The fast path is the low tier’s way to a decision. The owner and the risk lead decide within two working days, with no board meeting. It is for the low tier only. It keeps the checks quick where a mistake is small. So a team has a reason to register a tool instead of buying it quietly.
Explore: the refund workflow
Section titled “Explore: the refund workflow”The explorer applies the method. Load one of Larkfield’s systems, or change one factor and watch the tier and the controls. Make your prediction first.
Explore
Score a system and see its tier and controls
Predict first: The refund workflow proposes refunds, and a person approves each one. Which tier is it? And which tier if it refunds small amounts alone?
- Is it a use that a law bans or calls high-risk (for example hiring, credit, insurance prices, education, essential public services)?
- Does it make a decision about a person based only on automated processing?
- Sum
- 1 + 1 + 1 + 2 = 5
- Tier
- low
- Decided by
- the sum
- Controls
- 4
- Review
- every 12 months
The sum 5 is in the low band (4 to 6). No rule changes it. Tier low: 4 controls, and a review every 12 months. These are the course's scores for S3 Document assistant.
- Needed at this tier
- Replaced by a stricter control at this tier
- Not needed at this tier
| Control | Added at |
|---|---|
| Listed in the inventory with a named accountable owner (Needed at this tier) | low |
| Acceptable-use rules apply (Needed at this tier) | low |
| Fast path: the owner and the risk lead decide within 2 working days (Needed at this tier) | low |
| Review once a year (Needed at this tier) | low |
| A system card, kept current (Not needed at this tier) | medium |
| A vendor check before buying or switching on (Not needed at this tier) | medium |
| An evaluation before each release, with the result in a decision record (Not needed at this tier) | medium |
| A person approves every action that affects a customer (Not needed at this tier) | medium |
| A change log (Not needed at this tier) | medium |
| If it affects customers: an impact assessment, signed by the accountable owner (no board review) (Not needed at this tier) | medium |
| Review every 6 months (Not needed at this tier) | medium |
| A full impact assessment, signed by the accountable owner and reviewed by the board (Not needed at this tier) | high |
| A release gate with written thresholds (Not needed at this tier) | high |
| Monitoring with pause conditions, and an incident plan (Not needed at this tier) | high |
| A privacy or legal review where personal data or the law is involved (Not needed at this tier) | high |
| Review every 3 months (Not needed at this tier) | high |
The method is constructed for this course: one reasonable method, not a law or a standard. The scores of the eight systems are the course’s answers.
Text description of the explorer
The explorer has a note with the two questions to ask before scoring. Below it, buttons load the scores of each of Larkfield’s eight systems. Four groups of three buttons set the factors: harm, people affected, autonomy and data, each from 1 to 3, with the words of each level. The readout shows the sum, the tier, what decided it (the sum, rule 1 or rule 2), the number of controls, and the review interval. A table lists every control of the method, with the tier that adds it. Each row is marked as needed at this tier, replaced by a stricter control of this tier, or not needed. It starts with the document assistant: 1 + 1 + 1 + 2 = 5, low, 4 controls, a review every 12 months. The refund workflow today scores 2 + 2 + 2 + 2 = 8: medium, 9 controls, and the fast path and the yearly review are replaced. If it refunds alone, autonomy becomes 3: 9, high, 12 controls, and rule 2 also applies.
What this tells us. Today, the refund workflow is medium: 2 + 2 + 2 + 2 = 8, with 9 controls. If it refunds small amounts alone, only autonomy changes, from 2 to 3. The sum becomes 9, and rule 2 gives the same answer: high, with 12 controls. Many people predict “high” for both, because refunds are money. But the topic does not move the tier. Autonomy does. The same refunds with a person who approves each one are medium.
Larkfield’s eight systems
Section titled “Larkfield’s eight systems”Farida scores all eight rows with the team. The result, calculated by the course’s method:
Text description of the chart
A column chart with three columns. Low: 2 systems (document assistant, meeting-notes tool). Medium: 5 systems (support-resolution workflow, support MCP server, weekly digest script, reply suggester in the helpdesk tool, invoice reader in the accounting tool). High: 1 system (ticket triage service).
What this tells us. Only one system needs the full set of controls. Two can use the fast path: the document assistant and the meeting-notes tool. So most of Larkfield’s governance work goes to the 5 medium systems, and to one high system that its own team called low. The invoice reader is medium although a clerk checks each amount: a wrong amount reaches a supplier, who is outside Larkfield.
Two results need a word. The weekly digest script is medium (2 + 2 + 1 + 3 = 8), but its tier does not matter yet. It sends customer data to a company with no contract, which breaks the policy before any tier applies. The first decision is to stop it. And the triage service needs a discussion.
Discuss: the registry says low
Section titled “Discuss: the registry says low”Farida’s table says that the triage service is high. The registry says low.
- Amira (developer of the triage service): “It only sorts tickets and drafts replies. An agent reads every draft before it goes out. That is low.”
- Tomás: “And it uses a model from a big provider. They test it.”
- Grace: “But the routing goes to nobody first. If it says a ticket does not need a person, no person sees that ticket.”
- Sam: “Which matters more: the drafts that agents read, or the routing that nobody reads?”
Before you read on, answer these:
- Score the four factors for the routing, not for the drafts.
- Which rule applies?
- What evidence would you show Amira?
The course’s answer. Score what the system does alone, not what a person checks. The routing acts alone: autonomy 3. A wrong route can leave a ticket about a safety problem with nobody, so harm is 3. Customers are affected (2), and the tickets hold personal data (2). The sum is 3 + 2 + 3 + 2 = 10: high, and both rules agree. The evidence is recorded: 10 tickets that needed a person reached nobody in one “last week”. The provider’s tests do not change the score; they test the model, not Larkfield’s routing.
Recorded numbers like these come from one model on one test set. They show what the records look like, not how good a model is today. A registry stores the tier that a team gives it. So the risk lead checks every registered tier against the method, at least when a system first registers.
What would change the answer: suppose a person checked every ticket that the router marks “no person needed”. Autonomy would drop to 2. But the sum would still be 3 + 2 + 2 + 2 = 9, which is high, because the harm stays 3. A tier moves down only when a real change lowers the harm or the autonomy. A new label does not move it.
When to use tiers, and what they cost
Section titled “When to use tiers, and what they cost”Tier every system in the inventory, and tier it again at every major change: new users, new data or more autonomy. A tier that nobody updates is like the triage service’s registered tier: a label from the day it launched.
The method costs about ten minutes per system, once the inventory row is good. The real cost is in the controls. A high tier means an impact assessment, a release gate, monitoring and a review every three months. So keep the method simple, and argue about the scores, not about the arithmetic. If a score is unclear, take the higher one and write why.
Do not use the tiers as a legal classification. “High” here means “check closely”, not “high-risk under a law”. None of Larkfield’s systems is high-risk under the EU AI Act as this course reads it, and only the specialist can confirm that.
Common mistake
Everyone scores their own system low
Symptom: every team’s system comes back low or medium, and nobody has a high system, even the ones that act alone.
Cause: each team scores its own system and sees what a person checks, not what the system does alone. The tier decides the work, so a low score saves the team work.
Fix: the risk lead scores with the team, factor by factor, and writes one reason per score. Check the two rules separately: anything that acts alone with a harm of 2 or more is high, whatever the sum.
Guided practice: the reply suggester
Section titled “Guided practice: the reply suggester”-
Open your copy of the risk-tier table and your inventory.
-
Ask the two questions for S6, the reply suggester from TicketHarbor (made up for this course). Both answers are no.
-
Score harm. A wrong suggestion can give a customer wrong information about money or a delivery: 2.
-
Score people. Customers read the replies: 2.
-
Score autonomy. The agent edits and sends each reply: 1.
-
Score data. Customer messages and the helpdesk history: 2.
-
Add the scores and check the rules.
Check: 2 + 2 + 1 + 2 = 7, medium, 9 controls. Neither rule applies.
-
Write the controls that S6 does not have yet. It has no named owner, no system card and no vendor decision.
Your turn: tier every system
Section titled “Your turn: tier every system”Score the six other rows: S1, S2, S4, S5, S7 and S8, with one reason per score. Then compare your tier with the registry’s tier for the four registered systems.
Show the answer
| System | Harm | People | Autonomy | Data | Sum | Tier | Registry says |
|---|---|---|---|---|---|---|---|
| S1 Ticket triage service | 3 | 2 | 3 | 2 | 10 | high | low |
| S2 Support-resolution workflow | 2 | 2 | 2 | 2 | 8 | medium | medium |
| S3 Document assistant | 1 | 1 | 1 | 2 | 5 | low | low |
| S4 Support MCP server | 2 | 2 | 2 | 2 | 8 | medium | medium |
| S5 Weekly digest script | 2 | 2 | 1 | 3 | 8 | medium | not registered |
| S6 Reply suggester in the helpdesk tool | 2 | 2 | 1 | 2 | 7 | medium | not registered |
| S7 Meeting-notes tool | 1 | 2 | 1 | 2 | 6 | low | not registered |
| S8 Invoice reader in the accounting tool | 2 | 2 | 2 | 1 | 7 | medium | not registered |
Only the triage service disagrees with the registry. If your score differs by one point elsewhere, check your reason against the level’s words. A different reason is worth a discussion; a different sum with the same reasons is an arithmetic slip.
Knowledge check
Section titled “Knowledge check”Check your understanding
Tomás says: 'The refund workflow handles money, so it is high risk.' Use the method to say when he is right and when he is wrong.
Show the answer
He is wrong today: the workflow proposes, and a person approves each refund. Harm 2, people 2, autonomy 2, data 2: 8, medium, 9 controls. He is right for Omar’s change. If it refunds small amounts alone, autonomy becomes 3 and the sum is 9. Rule 2 (acts alone with harm 2 or more) also says high: 12 controls. The topic is the same in both cases; autonomy moves the tier.
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
Predict: a new tool reads job applications and ranks the candidates for a manager, who decides. Before you score anything, what do you do, and why?
Show the answer
Ask the two questions first. Hiring is a use that a law can call high-risk (the EU AI Act lists employment among its high-risk areas), so the answer to the first question is yes. The tool goes to the legal specialist first, whatever its score. The method sorts the governance work; it does not decide the legal category.
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
Why does the method have rule 1 (a harm of 3 is never low)? Give scores that would be low without it.
Show the answer
Without it, one severe harm can hide behind three small scores. For example: harm 3, people 1, autonomy 1, data 1 adds up to 6, which is in the low band. A tool used only by staff, that only suggests, with no personal data, could still give advice that hurts someone’s safety. Rule 1 lifts it to medium, so it gets a system card, an evaluation before each release and a person who approves each action that affects a customer.
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
- A risk tier says how closely a system must be checked. Proportionate controls give each tier its checks: 4 for low, 9 for medium, 12 for high. A higher tier replaces the fast path, the yearly review and the owner’s assessment with stricter controls.
- The course’s method scores harm, people affected, autonomy and data from 1 to 3, adds them, and applies two rules. Ask the two legal questions before you score.
- The fast path lets the owner and the risk lead decide low-tier requests within two working days.
- At Larkfield: low 2, medium 5, high 1. The triage service is high, although its registry entry says low. The refund workflow is medium today, and high only if it refunds alone: autonomy moves the tier.
You can now say how closely each system must be checked. But a high tier only says “look closely”, not what to look at or who could be hurt. Next, you write an impact assessment for a high-tier change, and decide who signs it.
Next: Impact assessments