Governance & guardrails
AI risk assessment template: a free scoring rubric
An AI risk assessment template gives you a repeatable way to score each AI system by likelihood and impact, sort it into a risk tier, and decide what controls it needs. This page is the full rubric, ungated: the questions to ask, the scoring scales, the tiers mapped to the NIST AI RMF and the EU AI Act, and the trap to avoid.
How to use this template
Run it once per AI system. Capture the context, score likelihood and impact, combine them into a tier, and apply the controls that tier calls for. Use the same rubric for every system so the results are comparable; the point is not a perfect number, it is a defensible, consistent ranking that tells you where to spend your effort. This is the scoring step at the center of any AI governance framework and the Map-and-Measure work the NIST AI RMF asks for before you manage anything.
Step 1: capture the context
Before scoring, write down what you are assessing. These five answers drive the scores:
- What the system does, who owns it, and whether you built or bought it.
- Who it affects: employees, customers, the public, and whether it touches a decision about a person.
- What data it uses, including any personal, sensitive, or regulated data.
- How autonomous it is: does a human act on its output, or does it act on its own?
- What happens if it is wrong, unavailable, or misused.
Step 2: score likelihood (1 to 5)
How likely is it that this system causes a harm, given the controls actually in place today?
| Score | What it means |
|---|---|
| 1: rare | Hard to trigger; strong controls already in place; little exposure. |
| 2: unlikely | Possible but not expected; some controls; limited use. |
| 3: possible | Could happen in normal operation; partial controls; regular use. |
| 4: likely | Expected to happen without intervention; weak controls; broad use. |
| 5: almost certain | Happening or clearly imminent; no meaningful controls; heavy use. |
Step 3: score impact (1 to 5)
Rate the worst realistic outcome across these dimensions, and take the highest as the impact score. A system can be low on four dimensions and high on one; the one that is high is what decides the tier.
| Dimension | Low (1–2) | High (4–5) |
|---|---|---|
| People and rights | No effect on individuals | Unfair or harmful outcomes for people (bias, denial of a service, safety) |
| Legal and compliance | No regulatory exposure | Breaches law or a binding obligation (privacy, EU AI Act, sector rules) |
| Financial | Negligible cost | Material financial loss or liability |
| Operational | Easily absorbed | Disrupts a core process or decision at scale |
| Reputation and trust | Not visible externally | Public failure that damages customer or market trust |
Step 4: combine into a tier
Multiply likelihood by impact for a rating from 1 to 25, then read the tier. The tiers map to how the NIST AI RMF and the EU AI Act treat a system, so the same score tells you both how much internal oversight to apply and which regulatory questions to ask.
| Tier | Score | NIST AI RMF | EU AI Act | Controls |
|---|---|---|---|---|
| Low | Score 1–4 | Light-touch: register it, basic policy applies | Minimal-risk territory | Inventory entry, acceptable-use rules, no formal review |
| Moderate | Score 5–9 | Map and Measure it; document known risks | Often the transparency-obligation band (disclose AI use) | Named owner, documented risks, human review of outputs, periodic check |
| High | Score 10–15 | Full Measure and Manage; evidence and monitoring | Maps to high-risk-style scrutiny where the use qualifies | Committee approval, testing, human-in-the-loop, production monitoring, incident plan |
| Critical | Score 16–25 | Highest scrutiny; consider not deploying | Check against prohibited practices and high-risk obligations | Executive sign-off, strong controls or do not deploy; independent review |
The EU AI Act mapping is a prompt, not a verdict. Its categories (prohibited, high-risk, limited, minimal) are defined by use, not by a score, so a high or critical tier is your cue to check the system against the Act's actual high-risk and prohibited definitions, which our framework crosswalk lays out.
Step 5: assign the controls
The tier decides the controls, and writing down what you are not doing matters as much as what you are. For a low-tier tool, an inventory entry and the acceptable-use rules are enough. A high-tier system earns testing, human review, monitoring, and committee approval. A critical-tier system needs executive sign-off and strong controls, and sometimes the right decision is not to deploy it. Record the deferrals: which controls a system does not have yet, why, and what compensates.
The over-tiering trap
The most common mistake is treating everything as high risk. It feels safe and it is the opposite. When every system gets the full treatment, the committee drowns, reviews become rubber stamps, and the credit model gets the same attention as the meeting-notes summarizer. Governance theater is a documented failure mode, not a hypothetical one: the research on fragmented frameworks warns specifically about "superficial compliance" and decision paralysis when companies pile on controls indiscriminately. Tier honestly, put your real effort on the systems that affect people, money, safety, or the law, and let the low-stakes majority stay low-stakes.
What to do with the result
You now have a ranked list: every system, its tier, and its gaps. That is the input to everything downstream. The committee works the high and critical tiers, the acceptable use policy covers the low ones, and agents get the extra weight described in agentic AI governance.
If you would rather have this run for you, with a stated scope and price, that is the fixed-scope AI risk assessment ($20,000 to $80,000 depending on scope: how many systems and teams, company size, and regulatory exposure), and if you have been told to align with a specific framework, that is NIST AI RMF implementation. This template is the free sample of that work.
One line of hygiene: this template is general information, not legal advice, and it does not classify a system under the EU AI Act for you. For obligations that apply to your specific systems, talk to counsel.
Questions people ask
- What is an AI risk assessment?
- An AI risk assessment is a repeatable way to rate each AI system by how likely it is to cause a problem and how bad that problem would be, then sort it into a risk tier that decides how much oversight and control it gets. It is the step that lets you spend scrutiny where it matters instead of treating a drafting tool and a credit model the same way.
- How do you score AI risk?
- The common method is likelihood times impact. Score how likely a harm is on a 1-to-5 scale, score the worst realistic impact on a 1-to-5 scale across dimensions like people, legal, financial, and reputation, and multiply for a 1-to-25 rating that maps to a tier. The exact numbers matter less than applying the same rubric to every system so the results are comparable.
- What should an AI risk assessment include?
- Per system: context (what it does, who it affects, what data it uses, how autonomous it is), a likelihood score, an impact score across several dimensions, a resulting risk tier, and the controls that tier requires. The template on this page covers all of it, and it maps the tiers to the NIST AI RMF and the EU AI Act.
- How does this map to the NIST AI RMF and the EU AI Act?
- The tiers line up with both. The NIST AI RMF asks you to Map and Measure risk before you Manage it, which is exactly this scoring step, and the higher tiers pull in more of its Measure and Manage work. The EU AI Act sorts systems into prohibited, high-risk, limited (transparency), and minimal categories; our high and critical tiers are where you check a system against its high-risk and prohibited obligations. The full side-by-side is in our framework crosswalk.
- Do we need to assess every AI tool?
- Register every tool, but assess in proportion to risk. A quick inventory entry and the default acceptable-use rules cover most low-stakes tools. Reserve the full scoring, testing, and monitoring for the systems that affect people, money, safety, or compliance. Assessing everything to the same depth is the over-tiering trap below: it burns the effort you need for the systems that actually matter.
Tell us about the work.
A few lines is enough. We read every enquiry ourselves and reply within one business day.