Handing a file to an AI agent: traceability, control and compliance

When we show Panora to a brokerage of fifty, a hundred or three hundred people, the demo is never the end of the conversation. It is the start. Then come the CIO, the compliance officer, sometimes the DPO. And their questions are always the same.

They are the right questions. Entrusting sensitive data to an agent that reads, writes and acts is a risk decision. It is made on facts, not on a promise.

What we call a trusted agent

A trusted agent is not an agent that never makes a mistake. It is an agent whose every action can be checked.

At Panora, every action of an agent is traced: who launched it, when, from which data, with which approval. It is auditable: the log can be read, per team member, per file. And it is replayable: it can be resumed, or rolled back, because the state before and the state after are known.

This is not a reporting layer added at the end. It is how the agent is built.

The five questions, and our answers

Where is our data? Hosted in Europe, on certified servers, isolated brokerage by brokerage.

Do you train your models on our files? No. No client data is used to train a model, ours or anyone else’s.

Who decides? A human. The agent prepares, executes, proposes. Every sensitive step waits for a team member’s approval, under their own account, and the log says so.

Do we need to change our tools? No. The agent works inside your email and your management software, which stays the reference. It reads there, and writes back there.

And if we leave? Your data is returned to you, then deleted. Nothing stays with us.

Why this is the right order

These answers could be seen as constraints imposed by large accounts. We see them as the definition of the product. A broker is accountable for every file. An agent acting on their behalf must be able to account for itself, in detail, at any time.

AI executes. Humans decide. Everything else follows.