To do an AI risk assessment before launch, describe exactly what the AI will do, list the ways it could go wrong, score each one on impact, likelihood, reversibility and who gets affected, then decide on controls and have a named person sign off. For most small and mid-sized businesses that's an hour-long exercise per use case, recorded on one page, and repeated whenever the use case changes. The goal isn't paperwork. It's making sure you launch with the right checks for the actual risk, rather than too few for a risky use case or too many for a harmless one.

This post walks through the method step by step, gives you a scoring matrix and a template table, works through an illustrative example, and explains how to scale the effort to the stakes.

Why assess risk before launch rather than after?

Because the cheapest time to add a control is before anyone depends on the system. Adding a human approval step to an AI workflow on day one costs a design decision. Adding it after the AI has sent 400 wrong quotes costs an incident, an apology and a redesign.

A pre-launch assessment also forces useful clarity. Many AI projects drift because nobody wrote down exactly what the AI is allowed to do, who it affects, and what "good enough" means. The assessment answers those questions as a side effect.

Finally, it's increasingly expected. Larger customers ask suppliers how they govern AI. Regulations such as the EU AI Act take a risk-based approach that classifies uses by the harm they could cause, and other jurisdictions are moving in similar directions. Our post on the EU AI Act and AI agents covers that framework. Whether or not a specific law applies to you, a written assessment is the evidence that you thought about it. (This isn't legal advice; check which rules apply to your business.)

Step 1: How do you describe the use case?

Write a short description that someone outside the project could understand:

  • Purpose: what problem it solves and for whom.
  • What the AI does: reads, summarizes, drafts, classifies, recommends, decides, or acts.
  • Inputs: what data it sees, including any personal or confidential data.
  • Outputs and actions: what it produces and which systems it can change.
  • Human involvement: who reviews what, and when.
  • Who's affected: customers, employees, applicants, suppliers, the public.

The single most important word in that list is the verb. An AI that drafts a reply for a person to send is a very different risk from one that sends it. An AI that recommends a credit limit is different from one that sets it.

Step 2: What could go wrong?

List failure modes specific to this use case. Brainstorm with someone who does the work today; they'll know the edge cases. Common categories:

  • Wrong output: inaccurate facts, wrong figures, misread documents, hallucinated details.
  • Unfair output: different quality or outcomes for different groups of people.
  • Data exposure: personal or confidential data shown to the wrong person or sent to the wrong place.
  • Unsafe action: the AI does something it shouldn't, such as sending, deleting, paying or changing records.
  • Manipulation: someone deliberately tricks it, for example through prompt injection.
  • Over-reliance: staff stop checking because it's usually right.
  • Availability: the AI or its provider is down when you need it.

Aim for five to ten specific failure modes, not fifty generic ones.

Step 3: How do you score each risk?

Use four factors. Simple three-point scales are enough; false precision doesn't help.

FactorLow (1)Medium (2)High (3)
ImpactMinor inconvenience, easily absorbedNoticeable cost, complaint or reworkFinancial loss, legal exposure, harm to a person, serious reputational damage
LikelihoodRare given the design and dataPlausible; would expect it occasionallyExpected regularly without controls
ReversibilityEasily undone before anyone is affectedCan be corrected, with effort and some disruptionHard or impossible to undo (money sent, data leaked, decision communicated)
Affected peopleInternal only, small groupCustomers or staff in limited numbersMany people, vulnerable people, or decisions about individuals' rights or livelihoods

A simple way to combine them: multiply impact by likelihood, then bump the result up one band if reversibility or affected people scores high. In practice, treat any failure mode with high impact and low reversibility as high risk regardless of likelihood. Rare but irreversible harms deserve controls.

Step 4: Which controls match which risks?

Controls should target the specific failure mode. Typical options, from lightest to heaviest:

  • Scope limits: narrow what the AI handles; route everything else to people.
  • Grounding: make it answer from approved sources only, and cite them.
  • Validation rules: automated checks on outputs, such as amounts within range, required fields present, no personal data in external messages.
  • Permissions: read-only access, or actions limited by amount, frequency and type.
  • Human review: a person approves outputs before they take effect, for all cases or above a threshold. Our guide to human-in-the-loop AI covers how to design review so it doesn't become a rubber stamp.
  • Monitoring and alerts: watch error rates, complaints, costs and unusual patterns after launch.
  • Kill switch: a tested way to stop the system quickly.
  • Testing before launch: an evaluation set of realistic cases, including tricky and adversarial ones.

More detail on technical controls is in our post on AI agent guardrails.

Step 5: Who signs off, and when do you review?

Every assessment ends with a decision: launch, launch with conditions, or don't launch yet. Record who decided and why.

  • Low-risk use cases: the system owner signs off.
  • Medium: the owner plus the relevant department head.
  • High: leadership sign-off, with legal or data protection input where personal data, regulated activities or decisions about people are involved.

Set a review trigger: a date (every six or twelve months), plus any significant change to the model, data sources, permissions or scope, plus any incident.

The one-page AI risk assessment template

Copy this into a document or spreadsheet. One row per failure mode.

Failure modeImpactLikelihoodReversibilityAffected peopleRatingControlsResidual ratingOwner
[What could go wrong]1-31-31-31-3Low / Med / High[Specific controls]Low / Med / High[Name]

Above the table: use case description, systems and data involved, date and version. Below it: decision, conditions, sign-off name, next review date.

A worked example (illustrative)

For illustration, take an AI assistant that reads incoming refund requests for an online store, checks them against the order system and policy, and drafts a reply with a recommended action.

Failure modeImpactLikelihoodReversibilityAffectedRatingControlsResidual
Recommends refund outside policy2222MediumPolicy rules checked in code; human approves every refundLow
Quotes wrong policy to customer2222MediumAnswers grounded in the current policy page; human reviews draftLow
Includes another customer's details3132HighOrder lookup scoped to the requesting customer only; output check for other customers' dataLow
Manipulated by text in customer email2222MediumNo direct refund permission; drafts only; injection tests before launchLow
Staff stop reading drafts2222MediumWeekly sampling of approved drafts; track edit ratesMedium

Decision: launch with human approval for every refund for the first two months, then revisit whether low-value refunds within policy could be auto-approved. That's the kind of conclusion a good assessment produces: specific, conditional and revisitable.

How much effort should you put in?

Match the effort to the stakes:

  • Internal productivity uses (summarizing meeting notes, drafting internal documents): a ten-minute checklist is enough.
  • Customer-facing or data-heavy uses: the full one-page assessment.
  • Decisions about people (hiring, credit, insurance, access to services) or regulated activities: the full assessment plus legal and data protection review, bias testing, and documented human oversight.

If you want a broader structure for organizing this work, our explainer on the NIST AI Risk Management Framework shows how a widely used framework approaches the same problem.

Frequently Asked Questions

How long does an AI risk assessment take?

For a typical SMB use case, about an hour with the system owner and someone who does the work today, plus time to document controls. High-stakes uses involving personal data or decisions about people take longer and benefit from legal or data protection input.

Do we need a risk assessment for using ChatGPT or similar tools?

For general use, your AI usage policy and approved tools list usually cover the risk. A specific assessment makes sense when AI becomes part of a defined process, touches customer or personal data, or produces outputs that go to customers without heavy editing.

Who should do the AI risk assessment?

The person who will own the system, together with someone who knows the current process well. Bring in IT for access and data questions, and legal or data protection advice where personal data or regulated activities are involved.

What's the difference between a risk assessment and testing?

The assessment decides which risks matter and what controls are needed. Testing checks whether the system and controls actually work, using realistic and adversarial cases. You need both; the assessment tells you what to test.

Is this the same as a data protection impact assessment?

No, though they overlap. A data protection impact assessment focuses on risks to individuals from processing personal data and may be legally required in some jurisdictions for high-risk processing. An AI risk assessment covers broader risks such as wrong outputs and unsafe actions. Where personal data is involved, do both or combine them.

Want a second pair of eyes before you launch?

We run risk assessments as part of designing every agent and automation, and build the controls they call for into the system itself. It's part of our agentic AI development work. Tell us about the use case you're planning and we'll help you size the risk and the safeguards.