One score, one action.
A confidence number that doesn't tell the user what to do next is an opinion in percentage points. Every confidence surface I ship maps to exactly one of three actions — act, review, or ignore — never a bare 87%.
This is the rule the four pattern libraries below all serve. Confidence scores, failure states, explainability, and human-in-the-loop are the instruments. This is the contract they enforce: a score the user can act on, not just read.
1. The score commits to an action — not a number
A bare percentage pushes the decision back onto the user: "87% — now you figure out whether that's good enough." That's the algorithm refusing to take a position. The rule flips it: the model has to translate its own certainty into one of three recommended moves, and the interface states the move before — or instead of — the number.
| Confidence band | Action | What the interface does |
|---|---|---|
| High | Act | Recommends a default the user can accept in one move — with an undo always in reach. |
| Medium | Review | Surfaces the reasoning and routes the call to a human. The system declines to decide alone. |
| Low | Ignore | Withholds the recommendation rather than guess, and offers the manual path instead. |
Where each band starts is a domain decision, set with the team and calibrated against real outcomes — not a universal cutoff. The rule isn't the numbers. The rule is that every score lands on exactly one action.
2. Every "review" needs a reasoning surface
"Review" is the dangerous band — it asks a human to make a call without saying why the system is unsure. So the moment a score lands on review, the interface owes the user the model's reasoning: the handful of signals that drove the score, in language the person can check against their own judgment.
Even when the user overrules the model, they should understand how it got there. A review without a reason is just a shrug with extra steps. (See the ML Explainability and Confidence Score patterns for the surfaces that do this.)
3. Every action is reversible — the override loop
An act the user can't undo isn't trust — it's a bet they didn't agree to make. So the third part of the rule: every recommended action carries a first-class override. Not a buried setting; the override path is visible at the point of decision.
And every override is signal. When a user rejects a high-confidence recommendation, that's the most valuable feedback the system gets — it feeds the next round of calibration. Designing the override well is how the trust layer keeps earning its confidence bands instead of assuming them. (This is the Human-in-Loop pattern in practice.)
Your AI ships a number and hopes the user knows what to do with it?
That gap — between what the model knows and what a person will stake a decision on — is the work. I've designed this rule into production AI products across AdTech, FinTech, and EdTech. Let's apply it to yours.
Book a 30-min callThe rule in one line
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
This rule was forged on, and shows up across, these case studies:
- Programmatic Advertising Platform: media-buyer recommendations bound to an action, with reasoning and override built in.
- AI-Assisted Private Equity Investing: analyst scores that route to review with the model's reasoning, not a bare verdict.
One pattern per month with tradeoffs and code examples.