# AI Risk Assessment Template: Worked Example and Checklist A standard assessment you can run on any AI system: a worksheet that covers every dimension, a scoring rubric that produces a tier, a matrix of what each tier requires, and a sign-off page. The output is one signed assessment per system with a rationale and a review date. **Who this is for:** The risk owner or governance analyst assessing a specific AI system before deployment, and the approver who signs the residual risk. Source playbook: https://aigovernance.com/playbook/how-to-perform-ai-risk-assessment --- ## AI risk assessment worksheet _The questionnaire. One per system. Covers purpose, data, model, human oversight, failure modes, and third-party dependencies so nothing gets skipped._ ### Template | Section | Question | Response | Evidence / notes | |---|---|---|---| | System | What does the system do, and what decision or output does it produce? | | | | System | Who are the affected people, and how many? | | | | Data | What data is used for inference, and what are its sensitivity categories? | | | | Data | What is the provenance and lawful basis for that data? | | | | Model | Built in-house, fine-tuned, or a third-party model or API? | | | | Model | What is known about training data, evaluations, and known limitations? | | | | Oversight | Where is the human in the loop, and can they realistically override? | | | | Failure | What are the top three failure modes, and what is the impact of each? | | | | Failure | How would a failure be detected, and how fast? | | | | Third party | Which vendors or upstream models is this system dependent on? | | | | Legal | Which regulatory obligations apply, and is a specific assessment (for example an EU AI Act FRIA) required? | | | ### Worked example | Section | Question | Response | Evidence / notes | |---|---|---|---| | System | What does it do? | Ranks inbound job applicants 1-5 for recruiter triage | Product spec v3 | | System | Affected people, how many? | External applicants, roughly 40,000 per year | ATS export | | Data | Inference data and sensitivity | CV text, work history, education; special-category data likely present | Data map row 12 | | Data | Provenance and lawful basis | Submitted by applicants; basis is legitimate interest with notice | Privacy notice section 4 | | Model | Build type | Third-party base model, fine-tuned in-house on 5 years of hiring outcomes | MLOps ticket 812 | | Model | Training data, evals, limitations | Historical hiring data carries known demographic skew; adverse-impact eval run quarterly | Fairness report Q2 2026 | | Oversight | Human in the loop | Recruiter sees the score and can ignore it; no gate forces a second look at low scores | Workflow diagram | | Failure | Top failure modes | 1: systematic under-scoring of a protected group. 2: score treated as decisive despite guidance. 3: model stale after role mix changes | Risk workshop notes | | Failure | Detection speed | Adverse-impact eval is quarterly, so bias could run for up to a quarter undetected | | | Third party | Dependencies | Base model vendor; ATS platform | Vendor register PRC rows 4, 9 | | Legal | Obligations | EU AI Act Annex III high-risk; FRIA required; NYC Local Law 144 bias audit required | Legal memo 2026-03 | ### Acceptance criteria - Every section has a response, and "not applicable" is justified rather than left blank. - Claims in the responses point to evidence (a document, a ticket, a dataset), not memory. - Failure modes are specific to this system and each has a stated impact and a detection method. - Where a regulation requires a named assessment, the worksheet records whether it has been done and where it lives. --- ## Risk scoring rubric _Turns the worksheet into a tier. The same three-axis rubric as the inventory kit, kept identical on purpose so inventory and assessment never disagree._ ### Template | Dimension | Score 1 | Score 2 | Score 3 | |---|---|---|---| | Data sensitivity | | | | | Decision impact | | | | | Regulatory exposure | | | | **Tier:** any 3 gives at least High. All 1s is Minimal. Otherwise the highest score sets the tier (2 is Limited). Article 5 prohibited practice is Unacceptable. ### Worked example | Dimension | Score | Rationale | |---|---|---| | Data sensitivity | 3 | Special-category data present in CVs | | Decision impact | 3 | Score shapes recruiter triage order | | Regulatory exposure | 3 | EU AI Act Annex III plus NYC LL144 | | **Tier** | **High** | | ### Acceptance criteria - The rubric wording matches the inventory kit rubric exactly, so a system gets the same tier in both places. - The score for each dimension cites the worksheet response it came from. --- ## Risk tier control requirements matrix _What each tier obliges you to have in place. Use it as the checklist between "assessed" and "approved to deploy"._ ### Template | Requirement | Minimal | Limited | High | Unacceptable | |---|---|---|---|---| | Entry in AI inventory register | Yes | Yes | Yes | Do not deploy | | Named accountable owner | Yes | Yes | Yes | | | This risk assessment, signed | Optional | Yes | Yes | | | Human oversight design documented | No | Recommended | Yes | | | Pre-deployment evaluation (performance, and bias where people are affected) | No | Recommended | Yes | | | Ongoing monitoring with thresholds | No | Basic | Yes, with alerting | | | Regulator-specific assessment (for example EU AI Act FRIA) | No | If triggered | If triggered | | | Board or committee visibility | No | No | Yes | | ### Worked example | Requirement | Status for Resume Screener (High) | Gap? | |---|---|---| | Inventory register entry | Present | No | | Named owner | Dana Portillo | No | | Signed risk assessment | This document, pending sign-off | Closing | | Human oversight design | Recruiters can override, but no gate on low scores | Yes: add a review step for bottom-quartile scores | | Pre-deployment evaluation | Quarterly adverse-impact eval exists | Partial: move first eval before each model update, not only quarterly | | Ongoing monitoring | Quarterly only | Yes: add monthly adverse-impact check with an alert threshold | | EU AI Act FRIA | Not started | Yes: Legal to schedule | | Committee visibility | Not on the agenda | Yes: add to next AI governance committee pack | ### Acceptance criteria - The matrix is tailored to your control set, with each requirement pointing at the internal control or standard that satisfies it. - The assessment records the status of every requirement for the system's tier, and every gap has an owner and a date. - A system is not approved for deployment while any High-tier requirement is an open gap without a documented, time-boxed exception. --- ## Risk assessment sign-off _The page that makes the assessment a record. Names the residual risk, who accepted it, and when it gets looked at again._ ### Template > One page. Attach to the completed worksheet. - **System:** - **Assessment date:** - **Assessed by:** - **Risk tier:** : - **Key residual risks after controls:** - **Required controls confirmed in place:** - **Open gaps and exceptions:** - **Decision:** Approved to deploy / Approved with conditions / Not approved - **Conditions:** - **Accepted by:** - **Next review date:** ### Worked example - **System:** Resume Screener (INV-0044) - **Assessment date:** 2026-09-02 - **Assessed by:** R. Nkemelu, Risk Analyst - **Risk tier:** High: special-category data, consequential decision, Annex III use - **Key residual risks after controls:** Demographic skew could persist up to a quarter between evals (medium likelihood, high impact). Recruiters may over-rely on the score (medium likelihood, medium impact). - **Required controls confirmed in place:** Inventory entry, named owner, adverse-impact eval, override capability. - **Open gaps and exceptions:** Monthly bias monitoring (owner: MLOps, due 2026-10-15). FRIA (owner: Legal, due 2026-10-31). Low-score review step (owner: Talent, due 2026-10-10). - **Decision:** Approved with conditions - **Conditions:** All three gaps closed by 2026-10-31 or the system is paused. - **Accepted by:** D. Portillo, Head of Talent (risk acceptance authority confirmed for High tier per governance charter) - **Next review date:** 2026-12-02 ### Acceptance criteria - The person accepting the risk has documented authority to accept risk at that tier under the governance charter. - Residual risks are stated with likelihood and impact, not just listed. - Every conditional approval names what must be true by when, and what happens if it is not. - The next review date is set and lands in the inventory register. --- ## Governance controls this kit produces evidence for - **HOC-001**: The worksheet plus scoring rubric is a completed, evidenced risk classification for the specific system. - **BRD-006**: The sign-off page records the residual risk and a named risk acceptance against defined authority, which rolls up into risk appetite reporting. - **BRD-009**: Each signed assessment is a source row for the multi-framework risk register, with tier, rationale, and review date. - **CMP-007**: The worksheet flags when an EU AI Act conformity assessment or FRIA is triggered and tracks whether it has been completed. - **MON-003**: The failure-mode and monitoring sections define what bias and fairness monitoring the system needs and at what cadence.