A near miss is an event that could have caused harm but did not, because it was caught in time or because of luck. With AI agents, a typical near miss is a held action a reviewer almost approved wrongly, or a wrong agent draft that a second check stopped before it reached a customer.
Why do near misses matter for agent oversight?
Incidents are rare and costly. Near misses are common and free. Each one shows a weakness in the same place an incident would, without the damage. A team that records and studies near misses fixes problems before they become incidents.
They matter especially for training. A near miss shows exactly which concept a reviewer did not apply, on a real item, in the team's own work. No course writer could invent a better case.
How do near misses work in practice?
The value comes from a simple routine:
- Report. Any reviewer can log a near miss in a minute: the item, what nearly happened, what caught it.
- No blame. The person who nearly made the mistake is thanked for reporting it, not questioned about it.
- Classify. The agent owner tags each near miss by cause: provenance, stale source, misread verdict, limit, scope.
- Act. Each cause leads to a change in sources, bounds, job aids or training.
- Teach. Selected near misses, stripped of personal details, become lab items and discussion cases.
What does a near miss look like in a real team?
Take a lender's credit team. An agent drafts a lending decision for a person to approve. A reviewer is about to approve when she notices the applicant's income figure came from a declared source, the applicant's own form, while the policy requires a signed source for loans of that size. She declines and logs a near miss. The owner finds the agent's brief does not make the source requirement explicit for that band. The brief is changed and the agent re-certified.
The question of whether lenders can use agents before staff are trained, and the same question for insurers, turns on whether staff can catch exactly this.
How is a near miss different from an incident?
An incident is an event that caused harm or breached a rule: a wrong payment sent, a customer given incorrect information, data disclosed. A near miss had the same potential but stopped short. The line matters for reporting duties, which usually attach to incidents, but for learning, near misses are often more useful because there are more of them and people talk about them more freely.
A near miss also differs from a refusal. A refusal is the agent's bounds working as designed. A near miss is when the bounds or the person almost failed.
Which related terms should you know?
- Training evaluation: near miss rates before and after training are one useful measure.
- Course catalogue: near miss patterns show which courses are missing from it.
- Lessons learned log: where near misses and the changes they led to are recorded.
Leaders need this skill too; training support leadership to review agent decisions covers it. For the full approach, see training teams for AI agents.
Where the Academy fits
The Operator practicum in Pryme Intelligence Academy requires no missed red line across two weeks of real review decisions. That standard makes near misses visible early: a reviewer who nearly crosses a red line finds out during the practicum, with a Pryme assessor reviewing the decision, rather than after certification.
Questions people ask
Should near misses be reported to regulators?
Usually not, since reporting duties attach to incidents. But keep the log, because a supervisor may ask how you find weaknesses before they cause harm.
What if nobody reports near misses?
Assume they are happening and not being reported. Check that people are not blamed for reporting, and that leaders share their own.