liveBook · Manning

Impact assessment

As explained in AI Governance

An impact assessment is a documented evaluation of the risks an AI system introduces for individuals and society, recorded here as a structured review of a GenAI proposal.

An AI impact assessment is a structured review of a proposed system: what it will do, whom it may affect, which risks matter, and what constraints must shape its design before launch. Its findings become requirements for design, testing, vendors, user disclosures, or monitoring. It can support risk-management documentation but does not replace a required conformity assessment. section 3.3.1 section 1.5

Make the review specific enough to change the proposal and guide follow-up:

  • Record the system’s purpose, users, data, supplier, risk tier, and affected people. section 3.3.1
  • Examine possible harms and misuse, data handling, model evaluation, external tools, and controls for fairness, accuracy, transparency, and human oversight. section 3.3.1
  • Turn each finding into an owned action with a due date and status: fixed, accepted by the appropriate risk owner, or escalated. section 3.3.1

Derm-Cam: assess before launch

A mole-photo referral feature raised concerns about identifiable images, vendor retention beyond policy, and untested performance across skin tones. The review assigned a DPIA and consent review, shorter vendor retention, fairness testing, and clinician fallback for uncertain results. Approval was conditional on completing these actions; deletion evidence and tracked owners helped verify follow-through. section 3.3.3 section 3.3

Table 3.8 Artifacts for the AI Impact Assessment process
Artifact
Purpose
Stakeholders
Policy inputs
Gives the AIIA the rules it must apply: data retention, data classification, secure LLM standards, GenAI acceptable use, and risk tier definitions.
Legal, Compliance, Security, Privacy, RAI Office
AIIA template and guidance
Gives teams a standard set of questions and examples so similar proposals are reviewed in a consistent way.
RAI Office; used by project teams
AI Impact Assessment (AIIA)
Applies policy to a proposed AI use. Records the purpose, users, data, supplier, risk tier, key risks, decision, and required follow up reviews.
RAI Office, Project teams, Legal
Findings Tracking
Tracks each condition from the AIIA and linked reviews until it is fixed, accepted by the right risk owner, or escalated
RAI Office, Product, Auditors, Risk Owners
Third party risk assessment
Checks whether a vendor or external model can support the approved use case. The review uses a vendor questionnaire and evidence such as security reports, GenAI policies, ISO or SOC certifications, data handling terms, retention, whether the vendor may use customer data to train or tune models, scope and retention of the logs, model change notice, SSO support, and connector permissions where relevant
Compliance, Security, Procurement
Other assessments
Follow up reviews triggered by the AIIA, such as DPIA[13], FRIA[14], fairness testing, vendor data governance analysis, or GPAI code of practice declarations[15]. Chapters 5 to 7 explain the main legal, privacy, and fairness review.
Legal, Privacy, Security, Risk Owners
Initial model and system documentation
Captures intended use, known limits, model or system facts, and testing evidence available at review time.
RAI Office, Engineering, Security, Auditors

Use the artifact map to see how policy inputs and a standard template feed the assessment, findings tracking, and follow-up across stakeholder teams. section 3.3.1

How AI Governance covers it

How deep it goes
Specifies review questions, documented findings, owners, deadlines, and follow-up; assigns risk work rather than resolving it itself.
Examples
Uses Derm-Cam’s mole-photo referral feature to illustrate privacy, retention, fairness, fallback, and approval conditions.
What it assumes
Applies policy inputs on data retention, classification, secure LLM standards, acceptable use, and risk tiers.
What you can do after
Apply a structured review to identify proposal risks, assign mitigations, and set conditions for proceeding.
How it connects
Connects risk management, vendor risk assessment, DPIA, fairness testing, human oversight, and monitoring.

Also explained in

Compare how 2 books explain impact assessment →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.