Impact assessments
- Lesson 10
- 35 min
- Module 3: Inventory, risk tiers and impact assessment
Where we are
In the last lesson, we gave each system a risk tier. The refund workflow becomes high if it refunds small amounts alone. But a high tier only says “look closely”. It does not say who could be hurt, how, or what to do about it. In this lesson, we write an impact assessment and decide who signs it, as a town does before it changes the road outside a school.
Lesson 10 of 19 · AI Governance: Policies, Risk and AccountabilityImpact assessments
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
Before a high-risk launch or change, write down who could be affected, how, and what you will do about it. Keep it on one page, and let the accountable owner sign it.
Before you startWhat to know and to have ready
What you need to know
- The tiers and the controls of the high tier, from Risk tiers and proportionate controls.
- The responsibility table and the accountable owner of the refund workflow, from Who is accountable.
- Quebec’s rules on decisions based only on automated processing, from Canada, the US, the UK and others.
What you need
- The impact-assessment template, copied into your document tool.
- Request R-12 in requests for review, the system card S2 and the evidence from earlier courses.
A new road outside a school
Section titled “A new road outside a school”A town wants to make the road outside a school faster, with a wider lane and fewer stops. Before it builds anything, it asks some questions. Who uses this street? Drivers, but also children, parents with buggies and older people who cross slowly. What gets better? The drive is shorter. What gets worse, and for whom? Faster cars near children.
Then the town decides what it will do about the harms: a crossing, a speed limit at school times, a way for parents to complain. And a named official signs the plan, so that somebody answers for it.
The town does not do this for a new bench. It does it for changes that can hurt people.
The problem: a tier is not a plan
Section titled “The problem: a tier is not a plan”Omar’s request R-12 asks to let the refund workflow refund up to 25 dollars without approval (a constructed request). The last lesson showed that this change moves the workflow from medium to high: 9 controls become 12. The board could now say yes or no.
But a yes or no on a tier leaves the important questions open. Which customers are most exposed? What happens when the workflow refunds when the policy says “a person must decide”? How does a customer complain about an automatic decision? Who checks that the 25-dollar limit works? Without written answers, the board approves a label, not a plan. And when something goes wrong, nobody can show what was considered.
The idea: an impact assessment
Section titled “The idea: an impact assessment”An impact assessment is a short document, written before a launch or a major change. It says who could be affected and how, and what the benefits and the harms are. It names the rights that could be touched, and how people can get a wrong result corrected. Then it lists the controls and ends with a decision. It is the town’s plan for the road.
The course’s template follows the public outline of ISO/IEC 42005:2025, AI system impact assessment (official source, checked 22 October 2026). ISO/IEC 42005 is a paid guidance standard on how and when to assess the effects of an AI system on people and society. The template’s headings and questions are this course’s own words, not the standard’s text.
The template has seven sections. Each one answers one question:
- Scope: what is assessed, and who is responsible?
- The system: what does it do, what may it do alone, and what must it never be used for?
- People affected: who does it touch, and who is more exposed?
- Benefits and harms: what could happen, how likely is it, and how bad?
- Fairness, rights and redress: could the result differ for some groups? Which rights apply? How does a person get it corrected?
- Controls: what reduces each harm, who owns it, and what proves that it works?
- Decision: go, go with conditions, or no go, and what would change it?
A tiny example: the triage service on one page
Section titled “A tiny example: the triage service on one page”Larkfield’s only high-tier system today is the ticket triage service. Farida writes its assessment with Amira, its developer. The page is constructed for this lesson; the numbers in it are recorded by MLOps and LLMOps: Operating AI in Production.
Text description of the diagram
The page has seven sections. Scope: the routing of the triage service as it is today. Grace, the accountable owner, signs it. Farida wrote it with Amira. The next review is in three months. The system: it sorts every ticket and says whether it needs a person, alone. Agents read and send the drafts. It is never used to decide refunds. People affected: every customer who writes. The most exposed are customers whose ticket needs a person, for example about a safety problem. Benefits and harms: the benefit is faster sorting. The harm when it is wrong: a ticket that needs a person reaches nobody. That happened 10 times in one recorded week. Fairness, rights and redress: whether routing differs by language is not measured yet; a customer who writes again reaches a person. Controls: the release gate before every prompt or model change, a daily sample of tickets marked “no person needed”, and a pause if one of them needs a person. Decision: go with conditions. Stop if another ticket that needs a person reaches nobody.
Look at three things in this page. Section 4 names the harm from the system’s own record, not a general fear. Section 5 admits a gap: nobody has measured the routing by language. So “measure it” becomes a condition in section 7, with an owner. And section 6 ties each control to the harm that it reduces, so a reader can check that no harm is left without a control.
When you need one, and who signs it
Section titled “When you need one, and who signs it”The template and the risk-tier table answer both questions. An assessment is needed before a launch, before a major change (new users, new data or more autonomy), and at each review. A high-tier system needs the full assessment, which the board reviews. A medium-tier system that affects customers needs one too, signed by the owner, without the board review. Both are controls of their tier.
Text description of the diagram
The flow starts with a launch or a major change: new users, new data or more autonomy. The question is the tier after the change. If it is high, write the full assessment; the accountable owner signs it, and the board reviews it. If it is medium and the system affects customers, write the assessment; the accountable owner signs it, with no board review. If it is low, or medium for staff only, no assessment is needed; the inventory row and the tier’s controls are enough. A note applies to every path: if one of the two questions before scoring was a yes, the legal specialist reads the assessment before anyone signs. And every system is assessed again at each review and after an incident.
Some laws ask for their own assessments, and they are narrower than people think. The EU AI Act asks for a fundamental rights impact assessment only from some deployers of high-risk systems, for example public bodies and some credit and insurance uses. It is not required of every company. Canada’s Directive on Automated Decision-Making asks federal institutions, not private companies, to publish an algorithmic impact assessment. So Larkfield writes its assessments because its own method says so. The lawyer says whether a law also asks for one.
Discuss: who signs the refund assessment?
Section titled “Discuss: who signs the refund assessment?”Farida has drafted the assessment for R-12. Now somebody must sign it.
- Omar: “Farida wrote it, and she knows the risks best. She should sign.”
- Tomás: “It talks about rights and the law. That is the legal team’s job, so the lawyer signs.”
- Grace: “I own the refund policy and the promise to customers. But I am not sure I want this change.”
- Sam: “Can Omar sign? It is his request.”
Before you read on, answer these:
- Who should sign, and why?
- What is wrong with each of the other three choices?
- What protects the decision if the signer has a conflict?
The course’s answer. Grace signs. She is the accountable owner of the refund workflow: the one person who answers for its results and can say stop. Her doubt is not a reason to avoid signing. It is the reason the assessment exists: she can sign “go with conditions” and name the conditions.
Farida writes and checks the assessment, but she does not own the refunds. If she signed, the owner would never have to read it. The lawyer is consulted on section 5, for example on Quebec’s rule for decisions based exclusively on automated processing (official text, section 12.1, checked 22 October 2026), which Canada, the US, the UK and others explains. Whether an automatic refund counts is a question for the lawyer, not a conclusion for this course. Omar must never sign: a person who asks for a change cannot be the only check on it. Because the tier is high, the board reviews the signed page, and that review protects the decision if Grace has a conflict.
What would change the answer: if the refund workflow had no named accountable owner, as the inventory found for other systems, nobody could sign yet. The first step would be to name one.
When to write one, and what it costs
Section titled “When to write one, and what it costs”Write the assessment before the work starts, not after the launch. A first page takes a few hours with the system’s developer and owner. Most answers are already in the inventory row, the system card and the evidence. Keep it to one page and link the evidence; a long assessment is one that nobody reads.
Do not write one for every low-tier change: that is how assessments turn into forms. And do not copy last year’s page at a review. The questions stay the same, but the answers change with the system.
Common mistake
An assessment that says nothing
Symptom: every section is filled, the page is signed, and the board approves in five minutes. But section 4 says “possible errors, low likelihood”, and section 6 says “monitoring”.
Cause: the sections were filled with general words, not with this system’s evidence. Nobody could disagree with the page, so nobody had to read it.
Fix: every harm names a group, an example and its evidence, or says “not measured”. Every control names an owner and what proves that it works. A reviewer who cannot find one harm to question sends the page back.
Guided practice: sections 1 to 3 for R-12
Section titled “Guided practice: sections 1 to 3 for R-12”-
Open your copy of the impact-assessment template, request R-12 and the system card S2.
-
In section 1, write the system and the change: “Support-resolution workflow: refunds up to 25 dollars without approval (R-12)”.
-
Write the accountable owner who signs: Grace. Write who wrote it and who was consulted.
-
Set the next review: 3 months after a launch, or after the first incident. A high tier means a review every 3 months.
-
In section 2, write what the workflow may do alone after the change, and what still needs a person.
Check: refunds above the limit, refusals and tickets that the policy sends to a person still go to a person.
-
In section 3, list the groups: customers who ask for refunds, team leads, and finance.
-
Mark the more exposed customers. Use the evidence from earlier courses.
Check: you found the German ticket. In the course AI Agents and Workflow Orchestration, one of the changes that nobody asked for was on a ticket in German.
The recorded evidence for section 4 is in the same file. That course tried three designs of the workflow. With every proposal approved as it was, the design with one agent and a small model made 3 changes nobody asked for in 2 tickets, out of 70. The router design, where a small model only chooses between fixed steps, made 0, and a strict approver removed all of them in every design.
Your turn: sections 4 to 7
Section titled “Your turn: sections 4 to 7”Write sections 4 to 7 for R-12. Name each harm with its evidence, give each control an owner, and end with a decision and what would change it.
Show one possible answer
- 4. Benefits and harms. Benefit: faster refunds and less approval work. Harms: a wrong refund (money, and it can be corrected); a refund made where the policy says “a person”. Evidence: only the recorded test set of 70 tickets; there is no production measure yet. A wrong refusal is not part of the change, because refusals still go to a person. Write that down as a condition.
- 5. Fairness, rights and redress. Check the results by language. A customer who disagrees reaches a person in one step, and the reply says that a person can review it. Quebec: ask the lawyer whether a refund that is only ever granted automatically counts as a decision based exclusively on automated processing.
- 6. Controls. A limit of 25 dollars per refund and a daily total; only the router design; refusals always go to a person; Grace samples 20 automatic refunds a week; a pause if the sample finds one refund that the policy forbids; the release gate before any model or prompt change.
- 7. Decision. Go with conditions, as a pilot: one queue, 4 weeks, with the limit and the pause condition above. The board decides again with the sample results. What would change it: a forbidden refund in the sample (stop), or an incident with the approvals (stop, and fix the approvals first).
Knowledge check
Section titled “Knowledge check”Check your understanding
The board receives a signed assessment for R-12. Section 5 says 'no fairness issues expected'. Why should the board send it back, and what should the section say instead?
Show the answer
“No issues expected” is a claim without evidence. Nobody has measured the workflow’s results by language, and the agents course recorded a change that nobody asked for on a German ticket. The section should say what is not measured yet, and how it will be measured: results by language in the pilot. It should also say how a customer who disagrees reaches a person and gets the refund corrected.
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: the meeting-notes tool (low tier) gets a new feature that sends meeting summaries to customers. Does it need an impact assessment now?
Show the answer
Score it again first: this is a major change, because the tool’s output now reaches customers. People affected is already 2 (customers are in some calls), and harm may become 2: a wrong summary can tell a customer something false. Then the sum is 2 + 2 + 1 + 2 = 7: medium. A medium-tier system that affects customers needs an impact assessment that Omar, the accountable owner, signs, with no board review. If the change made it high, the board would review the assessment too. The tier after the change decides, not the old 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
Module completion check: justify the tier of the refund workflow, today and with request R-12, and show which controls follow from it.
Show the answer
Today: harm 2 (a wrong refund costs money and can be corrected), people 2 (customers), autonomy 2 (it proposes; a person approves each action), data 2 (personal data). The sum is 8: medium, with 9 controls. Medium drops the fast path and the yearly review. It adds, for example, a system card, an evaluation before each release, a person who approves every action that affects a customer, and a review every 6 months. It also adds an impact assessment that Grace signs, because the workflow affects customers. With R-12, autonomy becomes 3, and the sum is 9: high. Rule 2 (acts alone with a harm of 2 or more) gives the same answer. It then needs 12 controls. The full impact assessment, reviewed by the board, replaces the owner-only one, and a review every 3 months replaces the one every 6 months. The new ones are a release gate with written thresholds, monitoring with pause conditions and an incident plan, and a privacy or legal review. Autonomy moves the tier, not the topic.
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 impact assessment says who could be affected and how, the benefits and harms, fairness, rights and redress, the controls and the decision, on one page.
- The course’s template follows the public outline of ISO/IEC 42005, in its own words.
- Write one for a high-tier system, and for a medium-tier system that affects customers. Write it before a launch or a major change, and again at each review. The tier after the change decides.
- The accountable owner signs. The risk lead writes it, specialists are consulted, the person who asked never signs, and the board reviews high-tier pages.
In this module, you listed every system, gave each one a tier and assessed the riskiest change. But all of this is work for a few people. Most of Larkfield’s staff use AI tools every day, and they need rules they can follow without asking. Next, you write those rules.
Next: Acceptable use for staff