How do you verify what an AI agent decides?
You don't let the decision take effect until it passes checks that don't depend on the AI model: schema, business rules, limits and signatures. If a check fails, the action is stopped and handed to a person, never guessed through. Every check and its result is written to a tamper-evident log, and the agent never holds the keys to payments or core systems.
The five parts of a verification layer
| Part | What it does |
|---|---|
| 1. Deterministic checks | Schema, business rules and limits decide whether an action may commit, independent of the model |
| 2. Different mechanisms | Checks use different methods, so one failure doesn't slip past them all |
| 3. Bounded self-correction | An agent may retry a failed step a fixed number of times, then it must hand over |
| 4. Human hand-off | Repeated failures and anything outside the rules go to a named person |
| 5. Record and isolation | Every check is logged; credentials stay outside the agent, used only at settlement after all checks pass |
What verification can't catch
Checks catch actions that break the rules. They don't catch actions that follow the rules but miss the intent. That's why a named owner reviews exceptions and decides what agents aren't allowed to decide. The safety bound also assumes checks fail independently, which is why they use different mechanisms.
Where it comes from
This is the AMEG (governance) and AMES (security) research, applied in production in our AdTech work. For example, invalid-traffic verdicts are decided by rules, and the AI only explains a verdict that has already been decided.
Frequently asked
Can the AI check its own work?
Not as the only check. A model reviewing itself can share the same mistake, so the decisive checks are deterministic.
Does verification slow things down?
Rule checks are fast. In our AdTech work they run inside the real-time bidding window.
Does this help with the EU AI Act?
It provides the record-keeping and human-oversight controls the Act asks for. Whether a system is high-risk is assessed per engagement. See AI governance & verification.
See what this looks like in one of your workflows.
Related: AI governance & verification · AMEG
More explainers: The coordination tax · The verification premium · Agent orchestration vs RPA · Revenue per employee