AI drafts. A licensed human approves.
Certova's agents draft. Licensed humans approve. Here is why that boundary is built into the database instead of into a policy document.
An agency cannot delegate a licence to a model
When a certificate goes out wrong, the exposure belongs to the agency and to the licensed person whose judgement stands behind it. No amount of model accuracy transfers that. So the question is not whether the AI is good enough to be trusted — it is what the system does with work the AI produced.
In Certova the answer is structural: an agent's output is a draft in an approval queue. There is no configuration screen where that can be turned off, no admin mode that skips it, and no integration path — API, MCP server or in-app copilot — that reaches a policy or a certificate without going through it first.
Cited, not asserted
Every field an agent fills carries a provenance chip: where the value came from, when it was fetched, and what produced it. Machine-produced values are a different colour from everything else in the product, so a reviewer can see in one pass what still needs a human eye.
That is the difference between a system that tells you the roof was replaced in 2019 and a system that tells you the county permit record says the roof was replaced in 2019, pulled on the fourteenth of August, with a link to the permit.
Review has to be fast or it becomes a rubber stamp
An approval queue that takes as long as doing the work by hand does not get used; it gets clicked through. So the queue shows the change field by field, old value against new, with the checker's findings and the derived impact of accepting it, and the reviewer can edit before approving instead of rejecting and starting over.
The measure we hold ourselves to is not how much the AI did. It is how much of what it did a licensed human could confirm in seconds.