How to check AI answers before they reach a client

AI tools can give false information that sounds completely credible. An invented regulation, last year’s price, a product specification that does not exist. This is not a bug the vendor will fix one day. It is a feature of the way these tools work - which is why checking their answers has to be part of the job, not the goodwill of individual people.

The problem is that “check everything” in practice ends with nobody checking anything. After a week the team decides the tool works well and stops reading. So instead of a general rule, I suggest a simple split and a short checklist.

Not every task needs the same level of checking

The most useful step is splitting tasks into three levels according to what happens if the answer is wrong.

Green - working material. A draft, brainstorming, tidying up your own notes, a summary for your own use. None of it goes outside. A quick read is enough - a mistake here costs a few minutes at most.

Yellow - content that will reach a client. An email, a reply to an enquiry, a service description, a post. Here there is one rule: a person always reads it before it goes out. Not “glances at it”, but reads it with the questions from the next section.

Red - numbers, deadlines, commitments, legal and accounting content. Here the tool’s answer is not the source. A person checks every number and every date against the source - in the price list, in the contract, in the system. Not from memory and not by comparing it with what “looks reasonable”.

It is worth agreeing on this split once, for specific tasks in the company, and writing it down. “Draft proposal - red, because it contains prices” is a sentence that saves a lot of misunderstandings.

Four questions before sending

For every piece of yellow and red content:

1. Do the numbers and dates match the source? Not your memory - the source. This is the most common place where a mistake goes unnoticed, because the number looks credible.
2. Are there claims here that I cannot confirm? Names, regulations, specifications, references to studies or testimonials. If you do not know where it comes from - remove it or check it.
3. Does this sound like us? The tone, the way we address the client, no promises the company cannot keep. The tools readily add “we guarantee” and “the best on the market”.
4. Did I paste in anything I should not have? Personal data, other companies’ trade secrets, content covered by a confidentiality agreement.

Answering “I don’t know” to any of these questions means one thing: check before you send.

What is not worth handing to the tool

However well it works, a few decisions should stay with a person:

• the final pricing decision and committing to a deadline,
• replying to a complaint or a conflict situation - at least in the first contact,
• content with legal consequences without a person reading it,
• sending directly to a client without approval, if the content concerns commitments,
• assessing people - recruitment, performance reviews, HR decisions.

In these matters the tool can prepare a draft, gather information or organise notes. The decision and the signature stay with a person.

What to paste into the tool, and what not to

Checking the answers is half of it. The other half is what goes into the tool.

• Do not paste clients’ personal data into tools that do not have this covered in a contract. It is worth checking what a given tool does with the content you enter - whether it stores it and whether it uses it for further training.
• Do not paste content covered by a confidentiality agreement or other companies’ trade secrets.
• Anonymise when you can. To prepare a draft proposal, “a client from the construction industry, 20 people” is usually enough, without a name or tax number.

Four limitations worth knowing up front

It is better to hear these before the first use than after the first mistake:

• The tool can give false information that sounds credible - and it will not signal that it is unsure.
• It does not know the company’s internal arrangements unless you give them to it. It does not know your price list, your deadlines or what you agreed with a client last week.
• The result depends on the quality of what it receives. A disorganised price list produces a disorganised proposal.
• The same request can give different answers. Consistency comes from templates and ready-made prompts prepared for the team, not from the tool itself.

When a mistake slips through anyway

It will - it is a matter of time. What matters is what happens next:

1. First the consequence, then the cause. You fix it with the client before you start analysing.
2. You record the case. What went wrong and at which stage the check failed.
3. You improve the instructions or the template, so the same mistake cannot happen again.
4. If the mistake was serious, you move the task up a level - from yellow to red.

A record of mistakes is worth more than a list of successes. It is what produces rules that actually work in a specific company, not in an article.

How to introduce this in the team

You do not need a rulebook for this. One page is enough:

• a list of tasks in which the team uses AI, each with an assigned risk level,
• the four control questions,
• a list of things that are not automated,
• who to ask when someone is not sure.

It is best to agree on this page together with the people who will use it - then it stays on the desk rather than in a folder of procedures. In team training I prepare it for specific roles and we practise it on the participants’ real tasks.