GUIDE

AI Hallucination at Work: How to Catch an Unsupported Date

Why AI writes a date nobody gave it

Generative AI tends to fill gaps rather than leave them empty. If a document seems to need a delivery date, the AI produces one. The date looks fluent and convincing, yet no source supports it.

This is called an AI hallucination. In business documents, its riskiest forms are dates, figures and promises. Those lines go straight to a client, a manager or a supplier.

The answer is not to stop using AI. The answer is a short habit of asking where every line comes from. This guide shows that habit step by step, with a realistic example.

This happens most often when the input documents are incomplete. The more gaps there are, the more the AI is tempted to fill them. So the check matters even more when your inputs are incomplete.

Two incomplete papers, one delivery date: a worked case

Elif, a procurement specialist, has two documents on her desk. One is a short email from the supplier, the other is an internal site note. Her manager wants a delivery date to send to the client.

Elif gives both documents to the AI and asks for a one-paragraph delivery note. The draft arrives in seconds and looks very professional. The example below shows its critical line and how to catch it.

The company and person names are invented. The situation is a familiar one: two incomplete papers and a manager who expects a complete answer.

The supplier email gives no firm shipping date. The site note only states how long installation takes after the material arrives. With this information, a firm delivery date cannot be written.

Example (fictional): Example: a delivery note for Kaya Build (invented company and people)

Document 1, supplier email: "The pumps are in production. We will share the shipping week next week."
Document 2, site note: "Installation takes 5 working days after the material arrives."
AI draft: "Delivery date is 14 November 2026; the supplier has confirmed this date."
Check: 14 November appears in neither document. There is no written basis for "has confirmed" either.
Corrected line: "Shipping week pending from the supplier. Installation takes 5 working days after arrival. The project manager will confirm the date."

Neither the date nor the confirmation appears in any document. The AI invented both to fill the gap. The corrected line states only what the documents say and leaves the decision with the project manager.

Checking where each date comes from

Next to every date in the draft, write one question: which line of which document states this date? The answer must be a document name and a line.

If the date is not in any document, there are three possibilities. The AI calculated it, guessed it or made it up entirely. If it was calculated, ask for the calculation in writing and redo it yourself.

The AI can help with this step too. Ask it to show the source of every date in a separate column. Any row with an empty source column is the first row you check.

Check by opening the document, not from memory. Search for the sentence in the source and compare it word for word. A similar phrase is not the same phrase.

Spotting the promise nobody approved

Made-up content is not limited to dates. Phrases like "confirmed by the supplier", "firm delivery" or "no delays expected" often slip into drafts. Each one makes a promise that nobody actually made.

While reading, watch for verbs that commit someone. Mark words such as confirmed, committed and finalised one by one. Then check whether a written approval stands behind each of them.

If there is no approval, rewrite the sentence to match the facts. "The supplier has not yet confirmed the shipping week" is a clear and strong sentence. It is always firmer ground than a wrong promise to a client.

Keeping the decision with the responsible person

The AI drafts, you check, and the responsible person decides. Telling the client a delivery date is a decision for the project or procurement lead. When this order is clear, the document moves forward with confidence.

Add a clear decision line at the end of the draft. It states who will confirm the delivery date and after which piece of information. The open point is not hidden; it is handed to its owner.

When your manager reads the note, they see what is known and what is pending, separately. Decisions get faster, because the missing information is now a visible action item.

This approach protects you as well. Every sentence in the document has either a source or a decision owner behind it. When someone asks, your answer is ready: which document, which line, whose decision.

Checklist

Use the list below after every AI draft. It takes a few minutes and lets you catch the made-up line before the document leaves your desk.

Checklist

I marked every date, figure and promise in the draft.
I noted which line of which document each marked item comes from.
I redid every calculated date myself.
I deleted any line without a source or rewrote it as "not yet known".
I compared words like confirmed or finalised with a written approval.
I wrote a clear decision line and named the decision owner.

Try this check yourself in free Module 1 at /academy/en/free-preview, with no sign-up. You see a finished example, do the same task yourself and compare it with the explained solution. For your team, see the plans at /academy/en/pricing.

Go to the module: Free Module 1: A delivery date from two incomplete papers Pricing