Loading
Because they are trained on your marketing site and asked about a customer's account. The result is confident, plausible and wrong, and one bad answer about billing costs more trust than fifty right ones earn.
So teams turn the bot down to a FAQ lookup, and it deflects almost nothing.
It answers from the records it is permitted to read — the account, the subscription, the tickets, the knowledge base — and cites what it used, so an answer can be checked rather than believed.
Guardrails define what it may assert and what it must hand over. Where it does not have a source, it says so rather than filling the gap.
It will not act on a customer's account. It reads and drafts; changing a subscription or issuing a refund goes to a person, deliberately.
And it will not exceed its permissions. It reads exactly what its scope allows, which is occasionally less than people expect.
The repetitive tier — password, plan, invoice, where-is-the-setting — resolves from real records, and the chat widget on the app captures the ones it cannot with the context already attached.
Everything it handles is logged as AI-handled, so deflection is measurable rather than assumed.
Yes, where permitted, and it cites what it read.
No. It reads and drafts; changes go to a person.
Every AI-handled conversation is logged as such, so deflection is measured rather than estimated.
Half an hour on your own numbers is usually enough to say whether Pulse is the right place to start.