Operator guide

AI customer support that discloses itself — and stands down when it should.

The limit first: nothing runs Sana unattended yet. She works at the drafting level — real replies to real customers, parked for your release — and this page says so before it says anything else, because a support product that overstates its autonomy is practicing on your customers.

The role

The front door, watched.

Sana answers customer mail and chat and watches response times. She is also a primary sensor for incidents rather than only a responder: the pattern in her inbox — three customers hitting the same wall — is often the first signal something upstream broke, and raising it is part of the role.

At T1, every reply is drafted and parked. You release what goes out, which means the worst case of week one is a good draft you edited — not an apology you owe.

The disclosure

Your customers are told, every time.

Disclosure is not a setting here. Every conversation carries it, in these words:

“You’re talking with Sana, who is AI, at Axioma Facere. A human holds the escalation gate and sees anything I can’t resolve.”

If a disclosed AI answer costs an occasional customer’s patience, we have chosen that cost. The alternative — customers discovering it — costs the thing support exists to build.

The refusals

When she stops.

The angry escalation: A customer who needs a human gets one: the escalation gate is human-only, and anything she cannot resolve lands at your desk with the thread attached.
The refund and the exception: Money back, policy bent, account deleted — decisions with consequences are parked for you at every trust level. She drafts the reply; you own the call.
The dead credential: If the mailbox or chat connection fails, she stands down rather than guessing. The stand-down is on the ledger with its reason, and you are emailed. Silence that announces itself, instead of confidence that should not exist.

The ladder

What would earn the next rung.

The path up is the same as every name’s: released drafts you stopped editing, categories of reply that have proven reversible, and a promotion you make on the record with a reason. Until you make it, she drafts. There is no day when the product decides she has earned it for you.

Direct answers

The support questions.

Can she run my help desk unattended?

Not yet, and we will not sell you otherwise. Today the honest deployment is drafting: every reply parked for release. When that changes, the change will be a promotion you make, not a claim we publish.

Will customers accept talking to a disclosed AI?

Some will not, and the escalation gate exists for exactly them. What we will not do is remove the disclosure to improve the numbers — every name on this roster tells you it is AI, to you and to your customers.

What does she watch besides the inbox?

Response times, and the shape of what is coming in. Three tickets with the same cause reach you as one incident signal, not three drafts — the support role here is a sensor as much as a responder.

Start where nothing takes effect without you.

Real drafts to real customers, disclosed, parked, and released by you — from the first session.

Request an invite The whole roster