---
title: "What to ask before you let AI near a customer"
url: "https://biznesiq.com/blog/before-you-let-ai-near-a-customer"
description: "A small firm has one reputation and no communications department. Five questions that matter more than any demo when software starts writing to your customers."
---

# What to ask before you let AI near a customer

July 9, 2026·2 min read·BiznesIQ

A large company that sends a customer something wrong has a process for it: an apology template, a person whose job includes handling it, and enough other customers that the incident is a rounding error. A five-person firm has one reputation, a town that talks, and a referral pipeline that is also its marketing budget.

That asymmetry should change how a small firm evaluates any software that writes to customers, and it is not the way these tools are usually demonstrated.

## Does anything go out without a person seeing it?

The only question that really matters, and it should be answered with a plain no.

Drafting is where nearly all the value is: the quote written, the invoice raised, the follow-up composed. Sending is where nearly all the risk is, and it saves seconds. A product that insists on automating the sending is optimising the wrong end and should be asked why.

## What happens when it is wrong about a price?

Not whether — when. A model reading an email thread will occasionally misread scope, miss a condition, or price against the wrong comparable.

The right answer involves showing its working: which previous jobs it priced from, what it assumed about scope, and what it was unsure of. A number with no reasoning cannot be checked in ten seconds, and anything that cannot be checked in ten seconds will eventually be approved without being read.

## Where does the customer data go?

Every message it reads is customer data, and quite a lot of it is confidential — addresses, schedules, what somebody paid, sometimes what they are going through.

Worth asking directly: which provider processes it, whether it leaves the country, whether it is retained, and whether it is used to train anything. Vague answers here are informative, and under GDPR the firm using the software is the one answerable for the arrangement.

## Can you tell what it did last week?

Small firms discover problems late, because nobody is watching in real time. A log of what was drafted, what was approved, what was changed before sending and what was rejected turns a bad surprise into something you can see building.

It is also the only way to know whether the thing is helping. If most drafts get heavily edited, it is costing time rather than saving it, and you want that visible rather than felt.

## What happens if you stop paying?

The uncomfortable one. If the system holds your customer history, quotes and invoices, then stopping means either exporting into something worse or staying because leaving is too expensive.

Ask what an export contains, in what format, and whether it includes the reasoning and the documents or only the rows. The answer tells you what the relationship actually is.

## The version we hold ourselves to

Everything outbound is a draft. The reasoning is shown next to the result. The data is in a database the firm controls. Every action is logged where the owner can read it. And the export is complete, because the software is open source and the database is theirs.

None of that is generous. It is the minimum for software that writes in your name to people who pay you.

-   AI
-   approval
-   risk
-   customer communication

## Keep reading

-   [Your business software is a filing cabinet with a login](https://biznesiq.com/blog/your-business-software-is-a-filing-cabinet)
-   [The unsent invoice is the most expensive thing in a small firm](https://biznesiq.com/blog/the-unsent-invoice)