AI risk assessment tool
Most AI incidents are predictable in hindsight — sensitive data pasted into a consumer tool, a hallucinated figure that reached a customer, a biased output that slipped through unreviewed. This tool runs the predictable check before you ship: it walks you through the five risk dimensions that matter for generative AI and turns your answers into a rating plus a prioritised list of mitigations.
Why most AI risk assessments fail to prevent incidents
Generic checklists ask the wrong questions. “Does the system use personal data?” is almost always yes. “Is there a hallucination risk?” is always yes. The relevant questions are:
- Does the system’s output reach someone who cannot reasonably verify it themselves?
- Does the use case involve a consequential decision (a person is hired, denied credit, assigned a medical risk score)?
- Is the data involved regulated (health, financial, biometric) or particularly sensitive (children’s data, political opinion)?
- Does the output go directly to end users without any intermediate review?
Those are the factors that determine whether a given risk is theoretical or genuinely dangerous. This tool’s guided questions are structured around those distinctions.
How it works
You describe the deployment — the use case, the data it touches, where its output goes, and your industry — then answer guided questions for each of five dimensions: data privacy, hallucination, bias, security, and compliance. Every dimension produces a 0–100 sub-score. The overall rating leans toward the worst dimensions rather than a flat average, because one severe risk (exposing regulated data, or shipping unreviewed output to customers) should dominate. The tool then surfaces the specific mitigations tied to your highest risks.
The five dimensions and their critical failure modes
Data privacy. The key risk is regulated or sensitive data being sent to a third-party model without a data-processing agreement or without the model being configured to disable training on your inputs. The mitigation is straightforward: enterprise APIs include DPAs and no-training by default; consumer products typically do not.
Hallucination. The key risk is confident, wrong output reaching someone who acts on it. Low-risk: an internal brainstorming tool. High-risk: medical symptom analysis, legal research, financial calculations shown to customers. Mitigation focuses on grounding (retrieval-augmented generation from trusted sources), human-in-the-loop for consequential outputs, and clear uncertainty signalling in the UI.
Bias. The key risk is systematically different quality or outcomes for different groups, particularly on demographic lines. Most bias risk comes from training data that reflects historical inequities and use cases where the model influences consequential decisions about people. Assessment focuses on whether the use case is high-stakes and whether outputs are auditable.
Security. AI-specific security risks that traditional threat models miss: prompt injection (an attacker supplies content that changes the model’s behaviour), model inversion (repeated querying reconstructs training data), and excessive agency (an agentic system takes harmful actions because its scope was too broad).
Compliance. Regulatory requirements that specifically apply to AI: GDPR Article 22 for automated decisions, the EU AI Act (certain applications become “high risk” or “prohibited”), sector-specific rules for healthcare AI (FDA in the US, MDR in the EU) or financial AI. Industry and geography determine which rules apply.
Tips and notes
Answer honestly about the worst plausible case, not the happy path — risk assessment is for the failure mode. The biggest risk reducers are almost always the same few moves: keep a human in the loop for any consequential decision, ground answers in trusted sources to cut hallucination, restrict and log access, and never feed regulated data into a tool without a data-processing agreement and a no-training endpoint. Re-run the assessment whenever the use case, data, or audience changes, and pair it with the AI usage policy builder so the mitigations become enforceable rules rather than good intentions.