Skip to content
Gathena
da

Learn · All posts

Grounding is the antidote to hallucinations

AI models guess when they lack the facts, and say it with confidence. Grounding makes them answer from your own data and show the source.

Lasse Feldt-Rasmussen· 30 October 2026

Being confidently wrong is not just a human trait – AIs do it all the time, and it’s a real problem.

When you are finished reading this post, you will know the cure, but before we get to that, you need to learn a new AI term called grounding.

Why AI is confidently wrong

You ask an AI tool what your gross margin was last quarter. It answers at once, calmly and with certainty. The number is invented.

This is called a hallucination: an answer that is wrong or made up, delivered as if it were fact. It is not a rare glitch. Researchers at OpenAI compare language models to students facing a hard exam: when they are unsure, they guess, because the way models are trained and tested rewards a guess over “I don’t know” (Kalai et al., 2025).

Questions about your own company make it worse. A model learned from public text up to a fixed date. Your ledger, your stock and your contracts were never part of that. When you ask about them, the model has nothing to go on, so it fills the gap with something that sounds right.

The new term: grounding

Grounding means the AI answers only from what it has just looked up in your sources, and shows you where it found it.

Picture two employees. One answers from memory. The other opens the handbook and the spreadsheet, finds the line, and gives you the answer with the page number. Grounding turns AI into the second employee.

A grounded answer comes with its source: the file and the page, or the table and the row. “Notice period: 3 months. Staff policy, page 12, section 3.2.” You can open the source and check it yourself.

When there is no source, a grounded agent says so. It answers “I can’t find this in your data”, or it points to where the answer may be. That is the behaviour that replaces the confident guess.

Research backs this up. Models that look things up in a source before they answer hallucinate substantially less than models that answer from memory (Shuster et al., 2021). The model providers build it in: Claude, for example, can return citations that point to the exact passage behind each statement (Anthropic, Citations).

The cure: ground your agents in your own data

Grounding needs something to stand on. If your agents are to answer questions about your business, they must be grounded in your business: your accounting data, your stock, your orders, your contracts, and what each number means.

That is what a data foundation is. Your systems are connected to one warehouse in your own cloud, the data is cleaned, and every figure can be traced to its record. Getting your data in order grounds your agents in your business context and makes them more capable.

If the data is messy, the answers are messy too, just with a source attached. That is why the data foundation comes first. How to start is in Get your data in order before you buy another AI tool.

How fresh the facts are

“Real-time” is a word that gets used loosely. Most business questions need data that is current, not data that is live to the second. A warehouse updated every night answers “How did we do yesterday?” well. A grounded answer tells you the date of the data it used, so you know what you are looking at.

What it looks like at work

An employee asks: “Why did costs rise in March?” The agent looks in the accounting data, finds the entries that moved, and answers with the documents behind them. The employee opens two of them, sees that they match, and takes the answer into the meeting.

That is how it works at DuffBeauty: employees ask questions in Claude Team and get the answer with the documents behind it (the case).

What the antidote does not promise

Grounding does not make a model infallible. It can still pick the wrong row or read a document the wrong way. What changes is that every answer can be checked, and a missing source is said out loud instead of covered with a guess.

So the operator stays in charge. Where an answer feeds a decision or a booking, a person looks at the source and signs off. Grounding makes that check possible, because the source comes with the answer.

One step to take this week

Take one question your team answers every month, for example “What did we spend on freight last quarter?” Ask your AI tool, and then ask it where the number came from. If it cannot point to a record you can open, treat the answer as a guess.

If you want to know what it takes to ground your agents in your own data, that is what a conversation with us is for.

Sources