The Near Miss Nobody Logged

A junior reviewer stayed late on a Friday. The agent had drafted seventy-four client notes. Seventy-three were dull in the way good operations are dull. The seventy-fourth used the right tone and the wrong benefit amount, close enough to pass a skim, wrong enough to become a complaint, a clawback or a quiet loss of a household's trust. She caught it. She corrected it. She went home.

On Monday, her manager said the new workflow was "stabilizing." Nothing had shipped that shouldn't have. There was no ticket. There was no five-whys. There was a person, a Friday and a number that almost left the building.

That's a near miss. In aviation and medicine, near misses are treated as the most expensive free data you will ever get. In agentic work, they're treated as proof the system is fine. The catch is credited to the model. The labor is credited to no one.

What almost left

A near miss, here, is a finished artifact that was wrong, was caught by a human and didn't reach the person it would have harmed. The harm didn't happen. The conditions did. If you only study the incidents that escape, you're studying luck and calling it reliability.

The benefit amount wasn't a hallucination in the cartoon sense. It was a confident completion from a table the agent was allowed to see and a policy it wasn't required to re-read. The draft sounded like a person who had checked. No one had. The junior had.

I have watched this pattern in claims, in faculty communications, in vendor emails that "just needed a once-over." The once-over is the whole control. When the volume rises, the once-over thins, as The Human Overhead of Agentic Systems already named. A near miss is what the thinning looks like on the day you still got away with it.

Who noticed, and why that's not a strategy

The catcher is almost never the person who signed the contract. It's the person close enough to the work to feel the wrongness in their body, the amount that "looks off," the name that belongs to last quarter's cohort, the sentence that would be cruel if a human had meant it.

That noticing is expertise. It's also a single point of failure. If she had been out, if the queue had been worse, if she had trusted the calm of the draft, the note would have gone. Next time, some of those conditions will be true. Folklore doesn't scale. Named people get the flu.

Leaders sometimes answer this with more model evaluation. More evaluation won't replace a person who can be asked why. Explainability as a Leadership Skill is about what a non-technical leader actually needs in order to decide. A near miss is one of those needs, in story form: this almost happened. This is who stopped it. This is the decision right that was missing.

If your only record of that story is a Slack thread that scrolls away, you haven't learned. You have been entertained by your own escape.

The ritual that would have caught it in public

You don't need a new platform. You need a short, slightly inconvenient rite that makes the catch count.

Four lines, written the day of the catch, visible to the people who own the workflow:

1. What almost left. One sentence. No hero language.
2. Who noticed. A name, not "the team."
3. What would have been owed to the person it would have reached.
4. Which control was thin. Review minutes, exception staffing, a policy the agent isn't required to re-read, a decision right that had already moved to the draft.

This is a cousin of the Friday rite in Rituals That Keep the Human as Moral Agent: one decision, four sentences, one name, one stop condition. The near-miss version is for the decisions that didn't have to be made because someone still had enough attention to flinch.

Keep the log somewhere boring. A shared page is enough. If you put it in a tool that requires a ticket type, people will stop writing. The point isn't compliance theater. The point is that the next person who signs a volume target can see the catches that volume is already consuming.

Why the dashboard will miss it

Dashboards count what escaped or what was generated. They're almost silent on what was caught. Caught work looks like success. Success often means a human is still subsidizing the metric. That won't tell you about the next failure.

Trust moves on this. A client who never sees the wrong number doesn't send a complaint, so Trust Capital on the Balance Sheet records no withdrawal. The withdrawal was avoided. The conditions of the withdrawal remain. If you only book the complaints, you will believe the account is healthy until it's not.

A recovery isn't "we fixed the prompt." A recovery is: the class of error has an owner, a thinner control has been staffed or the volume has been cut, and the catcher isn't the only person who would notice. Anything less is a reprint of Friday.

Keep the human as the moral agent

Don't congratulate the system for a catch it didn't make. Don't punish the catcher with more volume because "see, review works." Don't let a near miss expire into a story you tell at offsites.

The moral fact is simple. Someone was almost owed an apology you wouldn't have wanted to write. A person prevented that. The organization now owes that person a change in the work, not a thank-you in a channel.

If you can't say who would catch the next one, you don't have a control. You have a mood.

This week

Open a one-page near-miss log. Write down one catch from the last thirty days that never became a ticket. Name the person. Name what almost left. Name the thin control. Then do one of two things before Friday: staff the control, or cut the volume until the existing staff can actually see.

The five questions in the Ethical AI Leadership Decision Toolkit are built for this kind of page. Field notes continue in EI Leadership Insights.