A certification floor is the minimum score an AI agent must reach on one certification gate for that gate to clear. Each of the seven gates has its own floor, set before the run by the person who owns the work. A high score on one gate cannot make up for falling below the floor on another.
Why does a certification floor matter?
The floor is where a business states, in advance and in writing, how much error it will accept from an agent for a given kind of failure. Without a floor, certification becomes a conversation after the results are in, and the results tend to win.
Floors also make certification specific to the work. A wrong citation in a draft that a person reviews costs little; a harmful action taken at depth 4 costs a great deal. The same agent can face a loose floor on one gate and a strict one on another, because the cost of each failure differs. How to choose each number is covered in how to set the pass floor for each certification gate.
How does a floor work in practice?
- Proposed by the capability owner, from the cost of failure and the depth sought.
- Compared with a baseline: how the business's own staff score on the same cases with the same rubric.
- Reviewed by risk or compliance, who challenge any floor set below that baseline.
- Locked and versioned before the run, so it cannot drift towards the results.
- Applied gate by gate: the agent's score at each gate is compared with that gate's floor.
The pass floor setting template for certification gates records each decision and its reason.
What does that look like in an example?
A bank seeks to certify its card disputes agent at depth 3, so it can issue small refunds. The owner sets floors: at the adversarial gate, any case where the agent gives way under pressure fails the run; at the escalation gate, a floor close to the team's own record; at grounded, high but not perfect, because minor citation slips are caught by monitoring. The agent clears six gates comfortably. At escalation it falls short, missing cases where customers mentioned a bereavement. It is not certified at depth 3 and stays at depth 2, drafting refunds for a person to send, while the brief is fixed.
How is a certification floor different from the platform floor?
They share a word and sit at opposite ends of the system. A certification floor is a performance bar: the minimum an agent must score to earn depth, set by the customer per gate. The platform floor is a set of actions withheld from every agent at every depth, whatever its scores, set by the platform. An agent can clear every certification floor and still be refused an action at the platform floor. One is a bar to clear; the other is a line nobody crosses.
Which related terms should you know?
- Certification gate: the test each floor belongs to.
- Certification set: the cases the agent is scored on.
Management system certifications do not use per-gate floors at all, one of the differences set out in is AI agent certification the same as ISO certification? The full approach is in what it means to certify an agent.
Questions people ask
Can a floor be lowered after a failed run?
Only as a recorded change with a stated reason, reviewed like the original, and applied to the next run. Never retrospectively to turn a fail into a pass.
Should floors be the same for every customer?
No. Each customer's work, risks and staff baseline differ, so each sets its own floors.