AI Glossary / Governance, ethics and safety / Intermediate

Risk tolerance

Definition

Risk tolerance is the amount and type of uncertainty or potential harm that an organization or person is prepared to accept while pursuing an objective. For AI, it helps determine when a system may operate, when safeguards are needed, and when a risk requires escalation or stopping.

What it means

A business might accept occasional incorrect product tags that staff can easily correct, while accepting far less uncertainty about whether an automated action could expose confidential records. The difference involves consequences, reversibility, affected people, and the organization's obligations, not simply different accuracy targets.
NIST definitions describe tolerance in several related ways, including acceptable uncertainty and willingness to bear remaining risk after a response. Organizations should therefore state how they use the term instead of assuming that every policy uses an identical definition. Tolerance is a decision boundary, not evidence that a system actually meets it.

How it works

Start with the objective and the people or assets that could be affected. Describe plausible adverse outcomes, their severity, their likelihood where that can be estimated, and important uncertainties. Decide who has authority to accept particular risks and which matters cannot be accepted locally.
Translate those decisions into usable conditions: permitted tasks, minimum evidence, operational limits, review triggers, and escalation rules. Evaluate the remaining risk after controls are applied. Monitor changes in use, populations, performance, and consequences. A threshold that was sensible for an internal drafting tool may be inappropriate when the same model starts sending messages directly to customers.

Three practical examples

• A catalog team allows a small rate of reversible tagging mistakes, provided a correction route and periodic sampling remain available.
• An organization prohibits its assistant from sending confidential documents to unapproved destinations, even when the assistant usually follows instructions.
• A maintenance team permits advisory predictions but requires an engineer's approval before equipment is taken out of service, because interruption has significant operational consequences.

Think of it this way

Think of setting conditions for driving in poor weather. The purpose of the trip, visibility, road conditions, and ability to stop influence what you are prepared to attempt. The analogy has limits: organizational AI decisions can expose other people to consequences, so an individual manager's comfort does not establish an acceptable boundary for everyone affected.

Worked example

Consider a fictional internal document-tagging pilot. Its owner sets a provisional condition: no more than 2% of reviewed tags may require substantial correction before an unattended rollout is considered. In a representative evaluation of 1,000 tags, reviewers identify 15 substantial corrections. The observed proportion is 15 divided by 1,000, or 1.5%.
That point estimate is below 2%, but it does not prove that the underlying rate is below the threshold. Sampling uncertainty, reviewer consistency, and differences between evaluation and production still matter. The team might require more evidence and specify how uncertainty enters its decision. It also checks a separate confidentiality requirement; good tagging performance cannot compensate for an unacceptable disclosure pathway. The percentages are teaching assumptions, not recommended operating limits.

Illustrative case study

In a fictional company, Meadow Repairs introduces an assistant to summarize service notes. Managers initially say that they have a low tolerance for AI risk, but staff cannot tell what that means. A workshop separates incorrect summaries, missed urgent issues, and unauthorized disclosure.
The team assigns different controls and decision owners to each outcome, then documents conditions for continuing the pilot. A change that gives the assistant access to additional records triggers a fresh review. The case illustrates how a broad statement becomes operational guidance; it does not report a measured improvement or demonstrate that every remaining risk is acceptable.

Common misconceptions and mistakes

Do not confuse tolerance with the risk currently present. A desired boundary and a measured exposure are different things. Do not treat a single accuracy percentage as a complete expression of tolerance, because errors can have very different consequences.
Risk appetite, tolerance, capacity, and limits are sometimes distinguished in organizational frameworks, but their exact relationships should be defined locally. Most importantly, an internal acceptance decision does not remove legal duties, contractual requirements, or the rights of affected people. A lack of evidence about harm should not silently become evidence of acceptability.

Business use

Documenting tolerance helps teams prioritize testing, set approval authority, and decide which capabilities should remain restricted. It also prevents every difficult judgment from being improvised during an incident. A useful record names the objective, adverse outcome, evidence required, controls, decision owner, expiry or review trigger, and unresolved uncertainty. Senior leaders should understand what is being accepted rather than approving an unexplained risk score.

People also ask

Is risk tolerance just a numerical threshold?

No. It can include qualitative limits, prohibited activities, required safeguards, and evidence conditions.

Can a team accept risk on behalf of everyone affected?

Its authority is limited by governance, obligations, and people's rights; internal willingness is not unlimited permission.

Does tolerance stay fixed after launch?

It should be reviewed when context, consequences, evidence, or organizational objectives change.

Check your understanding

A model's average error rate meets an internal target, but a newly discovered failure can disclose restricted information. Is the deployment automatically within tolerance?

Answer: No. The disclosure risk requires its own assessment against applicable conditions and obligations. An acceptable average error rate does not settle a different category of harm.

Related terms

• Risk control
• Uncertainty
• AI governance
• risk appetite
• residual risk
• risk assessment
• risk acceptance

Learn next

Take one proposed AI use and write separate acceptance conditions for ordinary errors, sensitive-data exposure, and loss of human oversight. Then identify the evidence and decision authority needed for each condition. This exercise reveals where a vague statement such as low risk needs a concrete interpretation.

More in Governance, ethics and safety

• Responsible AI - Responsible AI is the practice of developing, deploying, using, and governing AI with attention to its effects on people, society, and the environment.
• Retention limit - A retention limit is a defined boundary on how long data or records are kept for a particular purpose.
• Retraining - Retraining updates a previously trained machine-learning model through another training process using different training data.
• Risk-averse - Risk-averse describes a decision maker who prefers a certain outcome to an uncertain prospect with the same expected outcome value.
• Risk-neutral - Risk-neutral describes a decision maker who compares uncertain prospects by their expected value in a specified outcome measure, without an additional...
• Risk-seeking - Risk-seeking describes a decision maker who prefers an uncertain prospect to a certain outcome with the same expected value.

Sources

Sources: NIST CSRC, Risk tolerance, multiple contextual definitions: https://csrc.nist.gov/glossary/term/risk_tolerance . NIST AI RMF, Framing risk: https://airc.nist.gov/airmf-resources/airmf/1-sec-risk/ . NIST AI RMF Core, governance and documented risk tolerance: https://airc.nist.gov/airmf-resources/airmf/5-sec-core/ .

Previous: Risk stratification Next: RMSNorm