Why enterprise AI pilots die in security review

Every few weeks we meet a team with a brilliant AI pilot that has been stuck for months. The demo works. The business loves it. And it still cannot ship, because somewhere between the demo and production sits a security review that nobody planned for.

This is the most common failure mode in enterprise AI. Not hallucinations. Not cost. Paperwork, in the best sense of the word.

What security teams actually check

Having sat through more of these reviews than we can count, the questions are surprisingly consistent:

• Where does our data go? Which model provider, which region, what gets logged, and who can read those logs.
• Who approved this action? If an agent sends an email or updates a record, the reviewer wants a named human or an explicit policy behind it.
• Can we turn it off? A kill switch, per agent, that a non-engineer can operate.
• What happens when it's wrong? Escalation paths, rollback, and how errors surface to the team that owns the process.
• Can we prove all of the above to an auditor? Not in a slide. In an export.

None of these questions are about model quality. A pilot built in a weekend has answers to none of them, and that is why it stalls.

Build the boring parts first

The fix is unglamorous: treat the governance layer as part of the product, not a wrapper you add later.

In practice that means every agent we ship starts with the same skeleton. Inputs are declared, not open-ended. Every model call, tool call, and decision lands in an audit log with a timestamp and a version. Actions that touch external systems go through an approval gate until the error rate earns automation. And the whole thing deploys inside the customer's VPC when data residency demands it.

When the security review starts, the answers already exist. Reviews that took a quarter now take two weeks, and the pilot survives them.

A checklist you can steal

If you are building internally, put these in your definition of done before the first demo:

• Audit log with per-run detail, exportable to CSV
• Role-based access mapped to your existing groups
• A human approval step on any write action
• A one-click disable per agent
• A data flow diagram your DPO can read

It is an afternoon of planning that saves a quarter of stalling. And if you would rather not build the skeleton yourself, that is exactly the part we bring to every engagement.