The questions you're actually carrying.

Every figure arrives with its source system and its as-of date, so a wrong number is a line you can trace to the record it came from — before anyone signs it. And when the data cannot support an answer, RIAI says which part it cannot answer, instead of averaging a guess in.

You are. Every outward act carries a person’s signature, and that person is at your firm — the same as today, with any tool and any junior hire. What we owe you is that you can always see where a number came from, so checking it means reading one line instead of trusting one sentence.

Exactly what the person asking may see, enforced in the request that goes to the database rather than in a prompt the model is trusted to obey. A household outside someone’s book is never fetched, so no phrasing reaches it. The mechanism.

No. Your records stay in your firm’s database. The model is Claude, reached through the Anthropic API under commercial terms that do not use the traffic for training — terms your firm can read directly. Whose model, running where.

Only what a person signed. Outgoing mail is drafted for a person to approve; filing a document and attaching one to a calendar event wait for sign-off as shipped; licensed work — money movement, trades, advice of record — is refused outright. The approval binds to the exact words: change one character and it will not run.

One append-only record holding every act, every refusal, and every setting change, reconstructable in order under the settings live at the time — plus your AI policy printed from the running configuration, so the document and the system cannot drift apart. Read a sample.

The database is already in your firm’s Azure tenant, in a documented schema. Stop paying and it stays where it always was — it just stops updating. No exit fee, no export request to file, no cooperation required from us. The exit terms, next to the money.

That connection is ours to fix — maintenance is what the subscription buys. While it is down, the consolidated database keeps answering from the last sync, and because every figure carries its as-of date, staleness is visible instead of silent. A source that is behind schedule says so on the card.

No. The split screen pairs every answer with the records it was built from; the deterministic parts — data, reports, checks — keep running with the AI switched off completely; and everything, including refusals, lands on the append-only record. The product, piece by piece.

Less than you’d think. The evaluation learns your firm by looking at its systems, not by questionnaire. After install, your people review settings rather than compose them, and the recurring complaints they raise unprompted set what gets built next — loudest first.