One score, one action. The rule every confidence surface I ship has followed since a roomful of traders taught me why.
An unexplained 87% is a shrug with decimals. It reports the model's mood and assigns the user all of the work: is 87 high for this kind of call? High enough to spend money on? Higher than my own hunch? While the person does that math, the moment to act expires — and a recommendation nobody acts on rounds down to zero, whatever the model's accuracy says.
The rule: a confidence score may never reach the screen alone. It resolves into exactly one of three verbs before the user sees it — act, review, or ignore.
1. The score resolves to a verb — not a number
Act. Confidence clears the threshold where the cost of checking exceeds the cost of being wrong. The interface makes acting one motion. Review. The model has a position but not the standing to hold it alone — the interface opens the reasoning, not the spinner. Ignore. Below the floor, a recommendation is noise wearing a number; the interface says so and steps aside.
| Confidence band | Action | What the interface does |
|---|---|---|
| High | Act | Offers a default the user accepts in one motion, undo always in reach. |
| Medium | Review | Opens the reasoning and hands the call to a person. The system declines to decide alone. |
| Low | Ignore | Holds the recommendation back rather than guess, and points to the manual path. |
Where each band starts is a product decision someone owns, argues for, and revisits against real outcomes — that ownership is the point a raw score lets nobody take. The rule isn't the cutoffs. The rule is that every score lands on exactly one verb.
2. "Review" forces the reasoning surface to exist
A verb of review is empty without something to review. So the moment a score lands there, the interface owes the user the handful of signals that drove it, in language the person can test against their own judgment. Build the rule and you have to build the reasoning panel; there is no skipping it.
Even when the user overrules the model, they should see how it got there. A review without a reason is a shrug with extra steps. (See the ML Explainability and Confidence Score patterns for the surfaces that do this.)
3. The override stops being a failure mode
An act the user can't undo is a bet they never agreed to make. So every recommended action carries a first-class override — not a buried setting, but a path visible at the point of decision.
A user correcting the model inside a review state is not the system breaking; it is the training data the next version needs. Log it, feed it back, and show people their fingerprints on next week's calls. The first time someone watches their own pushback sharpen the model, the override flips from a fight into coaching. (This is the Human-in-Loop pattern in practice.)
Where it has run
Media buying: planning fell from two weeks to three hours once buyers acted instead of re-deriving the model. PE screening: 60% faster across 42 deals in 90 days. Technical diligence: three-week cycles to four days. Different stakes, same three verbs. If a score on your screen doesn't bind to act, review, or ignore, with a reason for every review and an override for every act, it's decoration. Delete it or finish it.
See This Rule In Action
A roomful of traders taught me this rule. It now runs across these case studies:
- Programmatic Advertising Platform: each buyer recommendation resolves to one verb, with named reasoning and an override that trains next week's model. Planning fell from 2 wks → 3 hrs.
- AI-Assisted Private Equity Investing: a score that opens its own reasoning when it hits review, never a bare verdict the analyst has to take on faith.
- VC Technical Diligence: the same three verbs took the diligence cycle from 3 wks → 4 days.
One pattern per month with tradeoffs and the templates I use.