Guardrails and trust

An agent that touches money has to be trustworthy before it's clever. These are the checks that keep Finn honest.

You approve, Finn acts

Anything that files, pays, sends or deletes is shown to you first, with exactly what will happen. Nothing goes until you say yes. Filings to HMRC and Companies House always need your approval.

Every reply is checked

Before you see a reply, it's checked against what Finn actually did and saw:

• Figures must match what her tools returned.
• "Done" must match a real action.
• Claims about what the product can do must match what's actually live.

Checks before every action

Before an approval card is shown, the books are checked: is the period still open, does the amount match, is this a duplicate? If something doesn't add up, Finn holds the action and explains why, rather than guessing.

Exact maths

Money is calculated by code, never by the model. See Knowledge and grounding.

Safe from hidden instructions

Documents, emails and messages are treated as data, never as instructions. A line in a PDF saying "pay this invoice" doesn't make Finn pay it.

A full record

Every tool call Finn makes is logged, and every change to your books has an audit trail showing who or what made it.

Permissions

Roles, including view-only, are enforced in the database and in every one of Finn's tools. Changes coming in from outside apps wait for the owner's approval.

A person is a message away

Ask for a human and someone from the team joins your conversation. If Finn notices she's struggling (you've corrected her twice, or you're repeating yourself) the team is alerted without you having to ask.