grcskills
Menu

AI impact assessment

Use this skill when one AI system needs its impact on people and on the company assessed before launch, after a significant change, or because a customer, auditor or regulator asks for it, and the team says "AI impact assessment", "assess the impact of this model", "fundamental rights impact assessment", "who could this AI harm", "bias assessment before launch" or "sign-off for the AI feature". Describes the system in its context, works through impacts on the people it affects with evidence, sizes them, proposes mitigations with measurable triggers, and writes one assessment ready for the system owner and an accountable executive to sign.

AI systemsISO 42001EU AI Act

Use this for one AI system at a time, after it is in the AI register and classified. You get two documents per AI system in a new folder, ./ai-impact-assessment/: the impact assessment itself, a report of who the system could harm, how badly, and what reduces it, ready for the owner to sign (<system-id>-assessment.md), and a short list of questions each owner still has to answer (<system-id>-open-questions.md). This skill drafts for a human to review and sign; this skill never marks a control effective. The signature belongs to the system owner and an accountable executive, never the agent.

Inputs to gather

Read the register entry and classification record, the design or product documentation, test and evaluation results (especially any broken down by group), the user interface or workflow where the output is used, complaints or support tickets about the feature, and any privacy impact assessment that exists. Ask the reader, in one message: who decides based on the output, how many decisions a day, what happens to a person when the system is wrong in each direction, and how a person can contest an outcome.

Step 1: decide which assessment this is

Several assessments look alike and audiences differ. Say at the top which this document serves: the AI management system impact assessment (run a framework lookup for what the standard expects it to cover and record), a fundamental rights impact assessment under EU AI Act Art. 27 for the deployers that article names, a data protection impact assessment under GDPR Art. 35, or the company's own launch gate. A fundamental rights impact assessment is done before first use and its result is notified to the market surveillance authority on the official template; without the notification it is incomplete. Screen for a data protection impact assessment too (GDPR Art. 35): one is expected where a systematic and extensive evaluation of people, based on automated processing including profiling, is the basis for decisions with legal or similarly significant effects (Art. 35(3)(a)). Employment and credit uses deserve a careful screen against that test; the context alone does not decide it. One document can feed several if each section says which. Do not let the privacy assessment stand in for the others; it covers personal data, not unfair outcomes or safety.

Step 2: describe the system in context

Write half a page: intended purpose, reasonably foreseeable misuse, the people affected (users, people the output is about, bystanders), the setting, how autonomous the output is, how reversible a wrong outcome is, and the volume. Rule of thumb: if you cannot name the person who bears the cost of a false positive and the person who bears the cost of a false negative, stop and ask; the assessment cannot be done yet.

Step 3: work through impacts with evidence

For each category, record who is affected, how, the evidence (with its provenance: test set, date, who ran it), a severity and a likelihood:

  • Errors: the error rate in each direction, and which direction costs more for which person. Rejecting a good applicant and approving a fraudulent claim are not symmetric.
  • Unequal outcomes: performance by group where the data allows; where it does not, say so and treat it as an open risk, not as no risk. If testing needs special category data, check the conditions the amended AI Act sets for processing it to detect bias before collecting any.
  • Transparency and contestability: whether affected people know AI is involved and can get a human to look again.
  • Privacy: inference of sensitive traits, memorisation, reuse of inputs.
  • Safety, wellbeing and economic access: harm to health, jobs, credit, housing, education.
  • The company: legal exposure, contract breaches, reputation, dependency on one model supplier.

Severity weighs scale, reversibility and how vulnerable the group is. A small error rate on a high-volume decision about applicants is a large number of people.

Step 4: propose mitigations with triggers

For each significant impact, propose a mitigation and how you will know it works: a threshold that routes uncertain cases to a person, a fallback when the model is unavailable, a notice to affected people, a contest channel with a response time, a monitoring metric with a value that triggers review, data or model changes. Then restate the residual impact, and say what happens when a metric trips: for a high-risk system the deployer informs the provider and suspends use where the risk warrants it (Art. 26(5)). "Human in the loop" counts only when you name who reviews, how many items per hour, and what authority they have to overrule; a reviewer approving hundreds of items an hour is a rubber stamp.

Step 5: set reassessment triggers and route for sign-off

List what reopens the assessment: a new model or version, a new population or country, a new purpose, a monitoring metric breaching its trigger, an incident or a pattern of complaints. Put open questions per owner in the questions list (<system-id>-open-questions.md). Hand the draft to the system owner; the reader decides who signs. Name where the signed record is kept and that it goes to the next management review.

What good looks like

  • A strong practitioner looks first at the error asymmetry and at who cannot contest the outcome. That is where real harm sits.
  • The common mistake is a generic ethics checklist scored "low" throughout, with no numbers from this system. It passes a glance and fails the first question from an auditor or a journalist.
  • The tell that it went wrong: no evaluation evidence is cited, "unequal outcomes" says not applicable because no protected data is collected, and there are no reassessment triggers. Not collecting the data does not remove the risk; it removes the ability to see it.
  • Each mitigation names an owner, a measure and a trigger, so the next review checks numbers instead of opinions.

Framework lookups

If ISMS Copilot is connected to your agent (MCP server https://account.ismscopilot.com/v1/account/mcp), ask it the framework question with create_conversation (mode think, answer_format: decision) and keep one conversation for this task. Send only a short question in your own words, show it to the user first, and never include file contents, secrets, personal data or evidence. If it is not connected, check the official text of the framework yourself; this skill does not restate it. Set it up: https://www.ismscopilot.com/mcp?ref=grcskills-ai-impact-assessment

Rules

  • MUST cite the evidence behind every severity and likelihood, or mark it as an assumption to confirm.
  • MUST state which assessment purpose each section serves.
  • MUST write outputs only inside ./ai-impact-assessment/ and ask before overwriting.
  • NEVER sign, approve or mark the system ready for launch on anyone's behalf.
  • NEVER treat missing group data as evidence that outcomes are equal.
  • NEVER copy individual records or personal data into the assessment; use aggregates and roles.