What happens when the assistant gets something wrong?

Every vendor in this market will tell you their assistant is accurate. None of them will tell you it is perfect, because none of them are.

So the useful question is not “how often is it wrong” — a number nobody can give you honestly for your business before you have run it. It is: when it is wrong, what happens?

The three failure modes, in order of seriousness

It does not know, and says so. The benign case. The customer is told someone will follow up, and you pick it up. Slightly disappointing, entirely recoverable.

It does not know, and guesses. The dangerous one. A confidently wrong price or a promise you cannot keep costs you the argument afterwards, and you find out when the customer arrives expecting what they were told. A missed call loses you one customer; a wrong price loses you one customer and an unpleasant conversation.

It knows, but the knowledge is stale. You changed your prices in March and nobody updated the source. This is the most common one in practice, and it is an operations problem rather than an AI problem.

A product is largely defined by how hard it works to keep the first case common and the second rare.

What to look for

Does it answer from your material, or from the model? An assistant retrieving from your own knowledge base has a bounded set of things to be wrong about. One improvising from general knowledge has an unbounded set.

Can you see what it said? Every conversation should leave a transcript you can search. If you cannot audit it, you cannot correct it, and you will find out about mistakes only from customers.

Can you correct it in one place? If fixing a price means editing four channel-specific scripts, three of them will stay wrong.

Can you interrupt it? A live conversation you can step into is a materially different risk profile from one you can only read afterwards.

How ours is built

The assistant answers from a knowledge base you control, and when a question falls outside what it has been given, the design is to say a colleague will follow up rather than to improvise.

Every conversation is logged with a transcript and a structured summary, and calls are recorded. You can search all of it.

Corrections happen in one place: update the knowledge base and the phone, WhatsApp, Instagram and website chat all change at once — no redeploy, no waiting on us.

You can also take a conversation over mid-sentence on WhatsApp, Instagram or Messenger. The hand-over is atomic: if the assistant was composing a reply when you stepped in, that reply is dropped rather than sent alongside yours.

And on the sales side the assistant is deliberately bounded — it can move a deal on evidence in a column you have marked as its own, but it cannot close one, cannot act in a column you have marked yours, and cannot undo a move you made.

⚠️ The honest limits on all this. Website chat is read-only — you cannot step into one of those, and on Instagram and Messenger a thread past the 24-hour window cannot be answered by anyone. The history records that a human replied but not which teammate. And notifications reach you by email only — no WhatsApp, no SMS, no push — so “it escalates” means your inbox, not your phone screen.

The week-two question

When you evaluate any vendor, ask what happens in week two when it gets something wrong. You are listening for a specific, mechanical answer — where the transcript is, how you correct the source, how long the correction takes to reach customers.

“It’s very accurate” is not an answer to that question.