How the AI behaves
What the AI does, what it refuses, and where a person is required. Not a contract.
This page describes how the product actually works. It is not part of your agreement — the Acceptable Use Policy is the document that governs your use of the platform. This one exists so you can check our claims against the behaviour you observe.
An earlier version of this page said that deleting records, refunds and payments "stop and wait for a human to approve". That was not true, and it has been corrected below rather than quietly removed.
It says what it is
Every conversation opens by identifying the agent as an AI assistant for your business, and it confirms that if asked directly. This is enforced by code that runs after the model and before anything is sent, so it cannot be removed by a prompt, by a clever customer, or by a workspace setting. A reply that claims to be a person is blocked and replaced.
What it refuses
Medical, legal, tax and financial advice. Guaranteed outcomes. Prices and availability you have not given it. Anything about another customer. Any inference or action based on a protected characteristic.
A reply that breaks one of these is blocked before sending and the conversation goes to a person.
Where a person is required
Bulk messaging and messaging a recipient whose contact details have not been verified are classified as high risk. The platform will not execute either one: it records a request for approval and stops. This is enforced at the point of execution and a workspace cannot downgrade it.
The approval queue exists, at Approvals in your dashboard: a request that stops here is listed there with what was asked for and why it stopped, and an owner or admin can approve it, at which point the action actually runs. An earlier version of this page said the queue was still to be built, which was accurate when it was written and stopped being accurate on 11 September. A request that nobody answers expires rather than running.
Being accurate about the remaining limits: deleting records and anything involving payment are NOT classified as high risk — there is no payment tool at all, and cancelling an appointment runs without approval. Bulk messaging is refused outright rather than queued, because the sending machinery behind it has not been built, and a queue that accepts a request it could never carry out would be worse than a refusal.
Anyone who asks for a person gets one, immediately.
What it gets wrong
AI output is probabilistic. It will occasionally misunderstand, and the guardrails reduce but do not eliminate that. Every conversation is visible to you, every action is logged, and you can intervene at any point.
If an agent behaves badly, tell us at support@techmavs.com. We keep evaluation suites for exactly this, and a reported failure becomes a test case.