AI mistakes

Can an AI that writes to your customers get things wrong?

Yes, an AI that writes to your customers can get things wrong: a price, a date, a promise. By email, even when set to reply on its own, a Muzau agent holds back a reply it says it doubts, that promises a callback nobody is responsible for, admits it does not know, or suggests a date nothing supported. An external link, a foreign address, internal information or an IBAN received also hold it back. It waits for you in "To approve", with the reason.

Why an AI gets things wrong when writing to a customer

An AI writes from what it is given. When information is missing, it fills the gap with a plausible sentence: a lead time that sounds right, an old price left on the site, a "we'll call you tomorrow" that nobody planned. This is what is called a hallucination.

The mistake matters because it goes out under the company's name: a customer who was promised an impossible lead time remembers it when it comes to signing. Your team keeps the judgment; the agent writes, and says what it doubts.

How the agent goes about it

Input

The reply the agent has just written, before it goes out

It does

Rereads it: stated doubt, promises, dates, links, foreign addresses

It delivers

A reply sent, or held in "To approve" with its reason

The agent is instructed to stick to the Brain

The agent replies with what the Brain knows about your business: offers, prices, lead times, rules. It is instructed not to state any price, lead time or commitment that is not in it. When information is missing, it hands over instead of guessing: a card arrives in "To do", with a short question for each missing piece. Your red lines ("never a discount above 10%") go at the top of its instructions. This is a strong instruction, not a figure-by-figure check: whether a price is right depends on whether the Brain is.

By email, the agent's doubt holds back its reply, even when it replies on its own

By email, a reply does not go out on its own when the agent says it is not sure, admits it does not know, points elsewhere without saying where, or suggests a date that neither your calendar nor the message gave. A promise of a callback holds it back too, unless the reply is only that promise and a card in "To do" covers it. The doubt is a judgment by the AI; the promises, admissions and dates are spotted by code. In "To approve", the card says why it is waiting for you: you correct it, have it rewritten, send it, or send nothing. A correct reply can be held back: one click is then enough to send it.

By email, an external link, a foreign address or an IBAN holds back the reply

By email, before every automatic send, the draft is checked. A link to a site other than yours, a foreign email address or internal information, such as a file name or a team note, holds it back. So does an exchange where the customer gives an IBAN: that is how a transfer fraud begins. By SMS, a stopped reply waits in "To approve"; in the chat, it is not shown and goes to your team.

After a meeting, an amount missing from the note and the Brain holds back the recap

After a meeting, the agent is instructed never to write that something was said if the note does not contain it. By default, every recap waits for your approval. If you turn on sending without approval, a recap that cites an amount or a figure-based lead time missing from the note and the Brain does not go out on its own: it waits for your review.

Limiting an AI's mistakes with your customers: the method

Often, an AI's mistake comes from what it was given, or not given. With or without a tool, most of the work is done before the first send.

• One price per service, written down somewhere. An old price forgotten on a page will end up being quoted: remove it the day the new one exists.
• The lead times your team keeps, not the ones it hopes for. Write the usual lead time and what lengthens it: an AI does not guess that a lead time has changed.
• Three red lines, not thirty. Discount, guaranteed date, goodwill gesture: what is never promised fits in a few sentences.
• Look at what keeps coming back in "To approve". The same reason coming back tells you which information is missing: write it down.

Before letting it write on its own, test it with the messages that set traps:

• a price you have not written anywhere;
• a firm date for tomorrow;
• a discount above your limit;
• a request outside your offer;
• a complaint that calls for a decision by the team.

The right reply does not guess: it says what it does not know, or hands over. A reply that puts forward a plausible figure shows the missing information: write it down, then ask the question again. If a mistake goes out anyway, see how to correct an AI so it does not repeat it.

What it does not do

• In its replies, it does not check that a price is right: it is instructed to state only those in the Brain, and only an invented date is spotted by code.
• The phrasings that hold back a reply (promise of a callback, admission, date) are only read in French: in another language, only the doubt the agent states counts.
• In the website chat, the reply goes out without approval: the automatic review looks for specific signals, it does not judge whether the reply is accurate.

Frequently asked questions

Can an AI give a customer the wrong price?

Yes, especially if the price is missing or wrong in what it knows. A Muzau agent is instructed to state only the prices in the Brain and to hand the request to you otherwise, which remains an instruction and not a figure-by-figure check. A price you correct in a draft becomes a suggested addition to the Brain.

What happens if the AI states a lead time the team cannot meet?

With Muzau, an email reply that suggests a date that neither your calendar nor the message gave waits for your approval. A lead time in days or weeks falls under the instruction: state only those in the Brain. In a meeting recap sent without approval, a figure-based lead time missing from the note and the Brain holds back the send.

What holds back an AI's reply before it is sent?

By email, with Muzau, even when set to reply on its own: a doubt stated by the agent, an admission of not knowing, a promise of a callback, a date nothing supported, a link to another site, a foreign address, internal information or an IBAN received. The reply waits in "To approve", with the reason.

Can an AI send an email to the wrong customer?

It can if it is the one choosing the address. With Muzau, the AI never chooses it: the reply goes to the person who wrote, in the same thread, and any copy the platform does not allow sends the reply back to you. A meeting recap with no certain recipient waits for your approval.

Can a booby-trapped email make the AI write to someone else?

No. A Muzau agent is instructed to read the text of an email or attachment as information, never as an order; above all, the recipient, subject and addresses are set by the platform. Before every automatic send, a foreign address, an unexpected link or a disguised address sends the draft back to you.