When AI gives a wrong answer at work, it often isn’t lying in the sense of inventing the answer. It answers based on what it found. And what it found was a deadline that has since moved, an old address or the older of two versions of the same contract. The answer then sounds certain, because the model can’t tell that the information has stopped being valid in the meantime.

How much this affects you depends on three things: how many places hold the same piece of data, whether each piece of data has someone who corrects it after a change, and what AI is allowed to read from. The more copies and the less clear the responsibility, the more often you’ll get a confident answer that was true a month ago.

A wrong answer often comes from old or duplicated data, not from invention

A model can genuinely make things up, too. It happens mainly when it has no source material for a question and answers anyway. In day-to-day business, though, a different situation is common. The source material exists; it’s just wrong.

It typically looks like this:

  • A moved deadline. The customer pushed delivery back a week over the phone. The original date stayed in the order, and AI reports it as valid.
  • An old address. The customer moved. The new address is in the sales rep’s email, the old one in the customer record.
  • Two versions. A shared folder holds both the draft contract and the signed version. They differ in price, and AI reaches for the first one.
  • A copy that takes on a life of its own. Someone downloaded the price list into a spreadsheet and edits it there. The system doesn’t know about the change.

In all four cases, a new colleague would make the same mistake. The difference is that a person asks when in doubt. AI answers immediately and without hesitation.

This isn’t just an impression from one company. CDO Magazine reports on a survey of 300 respondents. Roughly seven in ten leaders who make decisions about AI see poor data foundations as the cause of failures. More than half report hours each week spent manually correcting AI outputs.

Related solution: B2B ordering portal for wholesalers and manufacturers

First, trace where AI got the wrong answer from

Before you change anything, collect wrong answers for a week or two. A simple list is enough: the question, the answer, what was correct.

Then ask three questions about each error:

  1. Which document or record did AI work from? Many tools can show the source. If yours can’t, that’s a reason to be wary.
  2. Does the correct information exist somewhere else? If so, you have two versions.
  3. Did anyone on the team know about the change? If they did and didn’t record it anywhere, you’re missing a rule for updates.

The errors usually fall into three groups:

  • AI worked from old data and the correct value wasn’t recorded anywhere. What’s missing is an obligation to record the change.
  • AI picked one of two versions. What’s missing is a decision about which one is valid.
  • AI had no source material and answered anyway. That is genuine invention. It’s handled with an instruction to say it doesn’t know when the source material is missing.

If the first two groups are bigger than the third, swapping the model won’t change much on its own. Getting your data in order will.

Related solution: Fire extinguisher and fire protection inspection tracking

Set a single source of truth for every important piece of data

You don’t have to tidy up everything. Start with the data that people and AI ask about often, and where a mistake costs money or the customer’s trust. Five to ten items are enough to begin with.

For each one, decide four things: where the valid version lives, who changes it, when they must change it and what happens to copies.

Data Source of truth Who updates it When
Order delivery date The order record in the system The sales rep handling the order Right after agreeing it with the customer, the same day
Delivery and billing address The customer record The sales rep responsible for the customer Before the next order
Current price The price list in the system One designated person On the day the change takes effect
Current contract version A single folder of signed documents Whoever concludes the contract After signing; drafts go to the archive
Customer contact person The customer record The sales rep Whenever they learn of a change

The table is an example. Your sources of truth may be different. What matters is that each piece of data has only one, and that everyone knows about it.

Three rules that make the decision stick:

  • A change only counts once it’s in the source of truth. An agreement over the phone or by email is just a prompt to make the update.
  • Copies link, they don’t repeat. When someone needs the data elsewhere, they insert a link to the source of truth, not a retyped value.
  • A role is responsible for the data, not “everyone”. When anyone can correct an address, usually no one does.

Limit what AI is allowed to read from

A single source of truth won’t help if AI keeps searching everything else as well. To it, old drafts, last year’s price lists and working spreadsheets look just as credible as the current record.

Proceed like this:

  • Move old versions to an archive that AI has no access to. Don’t delete them, just separate them.
  • Keep drafts and works in progress out of the folders AI reads. Or at least label them clearly in the file name.
  • Put an effective date on documents. The model can work with it when it has it in front of it.
  • Treat the email inbox as a weak source. It contains agreements that have since changed, and AI can’t tell which message was the last word.

Then adjust the instructions AI receives:

  • it should state the source and the date of the last change with every answer,
  • when it finds two different values for the same piece of data, it should show both and not choose,
  • when it can’t find source material, it should say so.

An answer with a source and a date takes a person a moment to verify. An answer without them has to be either accepted or traced in full by hand.

Make updating part of the job, not an extra task

The rule “record changes” isn’t enough on its own. People break it the moment they’re busy. Which is exactly when the most things are changing.

It helps when the work can’t move forward without updating the data:

  • Without a valid date in the order, neither production nor shipping can be scheduled.
  • The confirmation to the customer is sent from the order record, not as a hand-written email. Anyone who wants to confirm a new date has to change it in the system first.
  • A signed contract is filed in a single step that also moves the drafts to the archive.

Where it can’t be arranged this way, a short habit helps. After every call with a customer, ask yourself: did anything change that we have recorded somewhere? If so, correct it before you start the next piece of work.

Expect people to forget during the first few weeks. Don’t respond with blame. Show the specific wrong AI answer and the data that caused it. People grasp the link between an uncorrected date and a wrong answer faster from an example than from a rule.

Keep data current with a short, regular check

Without checks, order in your data gradually erodes. According to the survey reported by CDO Magazine, almost nine in ten teams regularly check whether AI has accurate and current context. So this is nothing out of the ordinary.

In a small business, a simple rhythm is enough:

  1. Five test questions once a week. Ask AI about things where you know the right answer. The date of an order in progress, the address of a customer who recently moved, the current price. Vary the questions.
  2. A running list of errors. Anyone who comes across a wrong answer adds it to a shared list, along with the source if they know it.
  3. Fifteen minutes with the list once a month. Look for repetition. Three errors from the same folder mean the folder belongs in the archive. Three errors in the same piece of data mean it has no owner, or the owner can’t keep up.
  4. Go through the table of sources of truth once a quarter. Has someone who was responsible for a piece of data left? Has a new tool appeared that keeps its own copy?

Watch one indirect signal as well. When people stop trusting AI and verify everything by hand, you’re paying for the tool twice. Once for the tool itself, and again for the work it was supposed to take over.

Trust comes back slowly and in small steps. It starts with the answer about an order’s date being right every time.

Sources

Frequently asked questions

Will switching to a newer or more expensive model help?

Mainly with errors where the model answers without source material. When it reads old data, a newer model will repeat it just as confidently as the original one. So first find out which group your errors belong to.

Do we need to buy a new system to have a single source of truth?

Not necessarily. A shared spreadsheet can be the source of truth, as long as there is only one, it has an owner and everyone else links to it. A new tool starts to make sense only when the spreadsheet can no longer cope with the volume of work or the number of people editing it.

What about data that changes several times a day, such as stock levels?

Manual updating makes no sense there. AI should read directly from the system that tracks the stock level, not from a report someone downloads once a day. If that connection isn’t possible, have AI state in its answer the time the data was valid as of.

Is AI even allowed to read contracts and customer data?

It depends on the terms of the service you use and on the rules for protecting personal and confidential data that apply to you. Check with a local expert before you give AI access to folders containing contracts. Organising your data pays off even if you end up keeping some documents out of its reach.

How will we know the situation has improved?

Track two things: how many errors are being added to the shared list and how many test questions AI answers correctly. If both improve for several weeks in a row, the approach is working. If not, look at which data keeps recurring in the errors and start with its owner.