AI Risk Assessment Template: The Essential 2026 Guide to Building a Usable One

An AI risk assessment template is a structured document for recording the risks created by each AI system you build or buy, scoring them with consistent scales, and tracking what you do about them. It turns scattered concerns about bias, accuracy, security and misuse into a register that managers can act on.

The template matters because AI risk behaves differently from ordinary IT risk. Models drift, outputs vary, and harm can land on people who never used the system. A template built only for servers and passwords misses those points.

This guide covers the fields to include, the risk categories to prompt for, and a worked example. Figures are illustrative, so calibrate scales to your own organization.

Why you need an AI risk assessment template

Most organizations now run AI in more places than they realize: a vendor chatbot, a scoring model in finance, a coding assistant, a document summarizer. Each one carries risks, and without a common template those risks are assessed differently, or not at all.

A shared template gives you comparability. You can see which systems carry the highest exposure, which owners have overdue actions and where the same weakness repeats. It also gives auditors and customers a single artifact that shows you manage AI risk on purpose, not by accident.

ISO/IEC 42001 expects an AI risk assessment process within the management system, and the NIST AI Risk Management Framework, described on the NIST AI Risk Management Framework page, organizes the work as govern, map, measure and manage. A template is how either approach becomes routine.

Start the AI risk assessment template with an inventory

You cannot assess systems you do not know about. The first tab of the template should be an inventory listing every AI system, its owner, its purpose, the data it uses, whether it is built in-house or bought, and who it affects. The approach in AI model inventory shows how to collect this without a long project.

Include shadow AI. Ask each department what tools staff use, and check expense reports and browser extensions. Unapproved use is common and is often where sensitive data leaks.

Tier each system by potential impact. A model that helps draft internal emails needs a light review. A model that influences lending, hiring, medical or safety decisions needs a deep one. Tiering keeps the workload realistic.

Field	What to record	Example
System	Name, owner, purpose	Resume screening model, HR
Risk	Cause, event, consequence	Training data skews results for some groups
Category	Bias, accuracy, security, privacy, misuse	Bias
Score	Likelihood x impact, 1-25	Inherent 15, residual 8
Treatment	Action, owner, due date	Quarterly bias test, data lead, Q2

Risk categories to prompt in your AI risk assessment template

People fill blank registers with the risks they already know. Prompts widen the view. Include at least these categories: data quality and provenance, bias and fairness, accuracy and robustness, security and adversarial attacks, privacy, transparency and explainability, human oversight, third-party dependency, misuse and legal or regulatory exposure.

Add rows for lifecycle risks too: model drift after deployment, which AI model drift risk covers, and incident handling, where AI incident management sets out what to do when something goes wrong.

A AI harm taxonomy can supply more specific harm types, so each risk is described in terms of who could be hurt and how.

Scoring scales for an AI risk assessment template

Keep the scoring simple and defined. Use a one-to-five likelihood scale and a one-to-five impact scale, multiplied to a score out of 25. Write the meaning of each level. For impact, consider harm to individuals, legal exposure, financial loss, operational disruption and reputation.

Score twice: inherent risk before controls and residual risk after. The gap shows whether your controls do real work. For AI, controls might be testing, human review, access limits, monitoring and documented approval gates.

Be candid about uncertainty. If you cannot estimate likelihood because the model is new, say so, score conservatively and plan to gather data. The guidance in AI risk assessment scoring goes deeper on scales and calibration, and AI risk scoring shows how to handle cases where evidence is thin.

Treatment columns and ownership

Each risk above your appetite needs a decision and an owner. Columns should capture the chosen response, the specific action, the named owner, the due date, the status and the expected residual score. Responses follow the usual pattern: reduce, avoid, transfer or accept.

For AI, typical treatments include adding a human review step, narrowing the use case, retraining on better data, adding monitoring thresholds, tightening vendor contracts and blocking certain inputs. Link each action to a control in your management system so evidence is easy to find. Our guide to the AI risk treatment plan explains how this maps to ISO 42001.

Insist on named individuals, not teams. Shared ownership usually means no ownership.

A worked AI risk assessment template example

Take a customer-facing support chatbot built on a third-party model. Risk: the bot gives confident but wrong refund information, causing customer loss and complaints. Category: accuracy. Likelihood 4, impact 3, inherent 12. Existing controls: a disclaimer only.

Treatment: restrict the bot to an approved knowledge base, add a handoff to a human for refund questions, log conversations for weekly sampling and set an alert when the error rate passes a threshold. After these controls, likelihood falls to 2 and residual risk to 6. Owner: head of support. Due: next quarter.

Notice that the row is specific, scored, owned and dated. That is the standard every entry in your AI risk assessment template should meet.

Keeping the AI risk assessment template alive

A template is only useful if it is updated. Review the register quarterly, and trigger a review for a new system, a major model update, a vendor change, an incident or a change in law. Keep a changelog so you can show how risk moved over time.

Report a small set of indicators to leadership: number of high residual risks, overdue actions, incidents in the period and systems without a current assessment. Oversight practices in AI human oversight and testing practices in AI bias testing provide measurable evidence for these indicators.

Finally, tie the template to procurement. No new AI tool should be approved until it has a row in the register and an owner.

Using a ready-made AI risk assessment template

You can build the template from scratch, but the effort goes into scales, formulas, heat maps and report layout rather than into your real risks. A prepared structure saves that effort.

The AI Risk Assessment Report includes a register, heat maps, an ISO 42001 treatment plan and a live workbook, so your team can enter systems and risks and see ranked results immediately. Test it on two or three real systems before rolling it out.

Whichever route you take, agree the scales and categories with legal, security and the business first, so everyone scores from the same definitions.

Who fills in the register

Completion works best as a short workshop rather than a solo task. The system owner describes purpose and data, an engineer explains how the model is built, tested and monitored, and a privacy or legal colleague flags regulatory exposure. A risk lead facilitates, challenges the scores and records decisions as they are made.

Two hours per system is usually enough for a medium-tier tool. High-impact systems need more, including testing evidence and input from the people affected. Record who attended and what evidence was reviewed, because that record shows the assessment was a real exercise, not a form completed in isolation.

When the session ends, every open risk should have an owner and a date. Anything left undecided goes to a named person for follow-up within a week, so momentum is not lost.

Evidence to attach to each risk

Attach proof where you can: test reports, vendor attestations, screenshots of configuration, approval emails and monitoring dashboards. Evidence turns a claim into something a reviewer can verify.

Common mistakes with an AI risk assessment template

The first mistake is scoring only technical risk. Harm to people and legal exposure often matter more than model accuracy. The second is a one-off exercise: models change, so the register must too.

The third is relying on vendor claims. Ask for evidence, such as test results and security attestations, and record what you received. The fourth is copying a generic list of risks without tying each to a real system.

The fifth is forgetting the decision. A register that never leads to a change in how systems are built or used is paperwork. Make sure each review ends with approvals, conditions or rejections that someone can point to.

AI risk assessment template FAQ

What is an AI risk assessment template?

It is a structured register for recording AI systems, their risks, scores, treatments, owners and review dates in a consistent format.

Is it the same as an AI impact assessment?

No. A risk assessment focuses on risk to the organization and its objectives, while an impact assessment focuses on effects on people and society. Many teams run both.

Which framework should it follow?

Either ISO/IEC 42001 or the NIST AI Risk Management Framework works. Many organizations align to both.

How often should we review it?

Quarterly at minimum, and whenever a system, model, vendor or law changes.

Who should own it?

A risk or governance lead should maintain the template, while each system owner owns the entries for their own systems.