The AI your accountant can sign off on
Every vendor now says 'AI'. The question that separates them is boring and decisive: what stops it from doing damage? Maya's answer is a governance system that was built before her capabilities were — and it is the reason she can be trusted with real work.
The eight controls
1) Propose → approve → execute: no action without a person. 2) Append-only audit trail: proposals, approvals and executions logged forever. 3) Undo: executed actions carry reversal where possible. 4) Value caps and rate limits per action. 5) Step-up authentication and two-person approval for high-risk actions. 6) Role-aware disclosure: a manager can be barred from margins and supplier costs, and Maya redacts accordingly — the AI respects your permission tree. 7) Consent gating: Maya cannot act at all until the owner accepts the agreements. 8) A kill switch that pauses everything, instantly.
Honesty as architecture
Maya refuses to fabricate: no data, no answer. The savings ledger only reports measured reductions, deliberately conservative. A data-quality guard flags problems rather than smoothing over them. These are design decisions, not model behaviours — they hold regardless of what the model would prefer to say.
What she remembers, you control
Her memory of your preferences is listed in Setup, item by item, deletable. Nothing about your business persists that you cannot see and remove.
Four properties, all of them required
An AI that can act on business data is only trustworthy if four things are true at once: it proposes rather than executes, a person approves, every action is recorded, and there is both an undo and a way to stop everything. Three out of four is not a safer version — it is an unsupervised employee with a gap in the record.
These are worth asking any vendor about, including us. Vague answers on any of the four are themselves the answer.
Limits are configuration, not trust
Capabilities are off by default and enabled individually. Actions can carry caps, high-risk ones can require re-authentication and a second approver, and the kill switch stops everything immediately rather than after the current queue drains.
None of this depends on the model behaving well. That is the point: safety that relies on the AI choosing correctly is not safety, it is optimism with logging.
An employee who asks permission.
Join the early-access list and meet her on your own numbers.