What happens when AI uses the wrong business data?

When AI uses the wrong business data, it gives answers that sound just as convincing as correct ones, but are based on an outdated price list, a duplicate customer or a CRM field nobody maintains. A person reading a report sometimes has doubts. An AI system does not doubt by itself, and an AI agent acts on the error immediately and repeatedly. The solution does not lie in a smarter model, but in knowing which source is leading, checking the data before AI acts on it, and making every outcome traceable to its source.

Why is wrong data worse with AI?

The old principle of "garbage in, garbage out" applies to AI in amplified form. An Excel formula on wrong data gives a wrong number. A language model on wrong data gives a wrong number with a well-written explanation attached. The presentation makes the error more believable.

There is more. A language model that cannot find the right data sometimes fills the gap itself. Ask for the contract price of a customer with no linked contract, and you may get back a plausible amount that comes from nowhere. That is called hallucination. Strictly speaking it is not wrong data, but the effect is the same: an answer without a source.

In what ways does AI use the wrong data?

1. The wrong source

The CRM says the contract value is EUR 120,000. The accounting system invoices EUR 108,000. Which is right? An AI system that treats the CRM as the truth reports a EUR 12,000 leak that may not exist, or misses a leak that does. The question of which data source leads for revenue has to be answered before AI gets to work.

2. Outdated data

Last year's price list, a contract without its latest amendment, a customer status not updated after cancellation. AI working from a snapshot that is weeks old makes decisions based on a reality that no longer exists.

3. Duplicate and inconsistent records

"Example Installations Ltd" and "Example Installations" as two customers. Revenue is split, the discount arrangement sits with one, the invoices with the other. A per-customer analysis sees two half-truths.

4. Fields nobody maintains

A CRM field "contract end date" that is empty for half the customers and holds a date from 2019 for a quarter of them. AI that predicts renewals from it is predicting noise. Why this happens so often is explained in why CRM data is not the same as financial data.

5. The wrong customer, the wrong line

An agent looks up "Johnson" and gets the first of four. A query links invoices to orders on a field that is not unique. The result is technically executed correctly and substantively wrong.

Why does this weigh more heavily with agents?

With an AI assistant that writes a report, a person reads the result. With an AI agent that acts, there is sometimes no longer a person between the error and its consequence.

A few scenarios:

• An agent prepares renewal quotes for 80 customers based on an outdated price list. If nobody checks them, they go out at the old rate.
• An agent sends payment reminders based on a receivables list in which credit notes have not yet been processed. Customers receive a demand for amounts they do not owe.
• An agent updates CRM fields based on a contract filed under the wrong customer. The error spreads into forecasts and reports.

People make such mistakes too. But a person makes them one at a time, and often notices halfway through that something is odd. An agent makes all of them, quickly and consistently.

Worked example: an outdated price list in an indexation round

Worked example: suppose an agent prepares the annual price adjustment for 150 maintenance contracts. It uses a price list stored in the CRM, but the actual rates were already adjusted in the accounting system halfway through last year. For 40 contracts the base is therefore too low, by EUR 30 a month on average.

• New price too low: 40 x EUR 30 x 12 = EUR 14,400 a year.
• Because future indexations are calculated on this base, the error carries forward every year.
• If customers have already received the new price, reversing it is a commercial conversation, not an administrative correction.

In the opposite case, a base that is too high, you send 40 customers an unjustified price increase. That costs no revenue, but it does cost trust and time.

The figures are assumptions. The mechanism is ordinary: two systems with a different truth, and an AI that picked the wrong one.

How do you recognise that AI is using the wrong data?

Signs that an AI system is working from the wrong data:

• Outcomes that do not match the accounts. A revenue total from the AI that differs from the general ledger account.
• Findings that often turn out to be unfounded on checking. If six of the ten reported leaks are not leaks, something is wrong in the data or the connection.
• Answers without a source. An amount that cannot be traced to a record.
• Different answers to the same question.
• Sudden shifts after a system change, migration or new connection.

How do you prevent it?

Decide which source leads for each data item

Prices from the accounting system or price management, not from the CRM. Contract terms from the signed contract. Invoiced revenue from the accounting system. Pipeline from the CRM. Record this, and let AI use only the leading source for each item.

Check the data before AI acts on it

Simple checks catch a lot: are mandatory fields filled in, are dates plausible, are customer numbers unique, do totals reconcile with the accounts? See how do you check CRM data automatically.

Make every outcome traceable

A finding without a source is not a finding. Every amount must be traceable to records in a system: this contract, these invoice lines, this date. Then a person can check it in a minute.

Do not let AI guess

Where numbers are concerned, AI retrieves them through a query or function instead of estimating them itself. See RAG vs database queries for business data.

Do not let uncertain findings through

A system that is in doubt must say so, and with too much doubt publish nothing. That principle is called fail-closed. How to judge confidence is covered in how do you know whether an AI recommendation is reliable.

A person between error and consequence

Anything that goes to a customer or affects money is seen by a person before it leaves the building.

Checklist: is your revenue data ready for AI?

1. Is it recorded, for each data item, which system is leading?
2. Are customer numbers unique, and used identically in every system?
3. When was the data last refreshed?
4. Does the revenue total the AI uses reconcile with the accounts?
5. Can every finding be traced to a record?
6. What happens when the AI cannot find something: does it report that, or does it fill the gap?
7. Who sees an action before it goes to a customer?

What is the reverse lesson?

There is a positive side too. An AI system that systematically compares sources exposes exactly the data errors that also cost revenue outside AI. The gap between CRM and accounting that confuses the AI is often the same gap that lets additional work go uninvoiced or an expansion never reach billing. Whoever gets the data in order for AI closes leaks along the way. The wider context is covered in how does AI work within Revenue Intelligence.

Frequently asked questions

Can AI recognise by itself that data is wrong?

Partly. It can spot anomalies, such as a price far lower than for comparable customers. But it cannot know which of two conflicting systems is right, unless you have recorded that.

Who is responsible when AI acts on the wrong data?

The company that deploys the system. An AI system carries no responsibility. That is why it matters that the choice of sources and the approval of actions rest with named people.

Does my data have to be perfect before I use AI?

No. But you need to know where the weak spots are and not let AI act on its own there. Reading, comparing and reporting can be done on imperfect data. Executing cannot.

How quickly will I notice that it is going wrong?

That depends on your monitoring. Without checks you only notice when a customer calls or the month will not close. With checks on outcomes and deviations you see it much sooner.