Why ChatGPT and Copilot give different answers about your organisation

You ask ChatGPT a question about your own onboarding process. Your colleague asks Copilot the same question. Two different answers come back, both delivered with confidence. That is rarely down to the model. It is down to the knowledge layer behind it: what information each tool can access, which version it sees and whether anyone has established which document is authoritative.

Why do two AI tools give different answers about the same company?

Because they draw on different sources. An AI model has no knowledge of your organisation; for each question, it is handed pieces of text and forms an answer from them. If that text differs, the answer differs. Two tools with access to different folders are two colleagues who have read different files.

That also explains why the answer looks so convincing. The model presents whatever it was given in polished prose. It doesn't know that the proposal template it read was last year's version, or that the current agreements are in an email it never saw.

What does ChatGPT actually see of your organisation?

By default: nothing. A standalone account doesn't know your company. What it can see depends on what you connect or upload.

The business versions of ChatGPT include Company Knowledge, which lets ChatGPT consult sources such as a connected document store or email and cite the source it used in its answer. That works well, and that citation is exactly what you want to see. What is missing is curation up front: there is no step where someone checks whether the document that gets retrieved is actually the right one. The tool retrieves whatever is there, including the four outdated versions sitting alongside it.

There are also Projects, where you add files and instructions per project. Fine for an ongoing piece of work, but it is a personal or team workspace, not an organisation-wide agreement on what is true.

And what does Microsoft 365 Copilot see?

Copilot, on the other hand, does have an organisation-wide view: it draws on what is in your Microsoft environment and respects the permissions you have already set up. Someone who isn't allowed to see a folder won't get answers from it either. That is a strong feature, and it is also the answer to why Copilot and ChatGPT differ: Copilot reads your SharePoint, ChatGPT doesn't.

But that same feature exposes an old problem. Copilot surfaces whatever is there. If three folders each hold a document with "onboarding" in the title, those are three candidates. Microsoft provides serious tooling to clean up that mess, but it all happens after the fact: cleaning, enriching, keeping things current. There is no gate up front where knowledge is approved before the AI may use it.

Can you train ChatGPT on your own company data?

This is the most persistent misconception, so let's be precise: no, and that is just as well.

What people call "training" is almost always something else. When you add documents to an AI tool, they are split up and made searchable. When you ask a question, the system looks up the passages that match it and attaches them to your question as context. The model reads them at that moment, answers your question and forgets them again. The model itself doesn't change.

Fine-tuning a model does exist, but it mainly changes how the model writes, not reliably what it knows. Putting facts into model weights is expensive, slow and impossible to correct: if your price list changes, the old version is still in there.

The fact that it works differently from what you expected is actually good news:

• You can update something today and the answer is right tomorrow. No training round, no waiting.
• Permissions keep working. Anyone who isn't allowed to see a document won't get answers from it. With a fine-tuned model, that is impossible to guarantee.
• Answers can carry a source. You can check where something comes from, which by definition you can't do with a fine-tuned model.

So the real question isn't how to train a model. It is which knowledge the system is allowed to retrieve, and who decides whether that knowledge is correct.

Why the same document produces two answers

Even with one tool and one document, answers diverge. We see three causes at almost every organisation.

Multiple versions. The proposal is in a folder, an amended version is in an email attachment and the latest agreements are in a chat message. All three can be found; none of them has been designated as the right one.

Different permissions. You can see the finance folder, your colleague can't. You get different answers, and neither of you notices.

Fragments without context. Documents are retrieved in chunks. If the mandate is on page 1 and the exception on page 9, it is quite possible that only the exception gets retrieved. The answer may be technically correct but wrong in practice.

What you can do about it

The pattern is always the same: the problem isn't in the AI but in the knowledge it draws on. Three things really help.

Assign one source per topic. Not tidying up everything, but deciding which document is authoritative for the twenty questions that come up every day. That is editorial work, not an IT project.

Approve knowledge before the AI uses it, not after. Have someone who knows the subject approve what goes into the shared layer and when it needs reviewing. Cleaning up afterwards is a losing battle.

Demand a source with every answer. If an answer can't show where it came from, it is a suggestion. Citing the source also makes a mistake easier to trace and fix.

Do this and the difference between the tools disappears by itself. Not because you get rid of a tool, but because both draw on the same curated layer. That is exactly what a Second Brain is: not another chat window, but the overarching layer across your AI systems that gives every model the same up-to-date company knowledge.

When is this a real problem?

Not in a team of five sharing a table. There, everyone knows which proposal is current. If an answer is wrong, a quick conversation puts it right.

It starts to hurt somewhere between twenty-five and a hundred employees. By then there are several teams describing the same client differently, people who haven't lived through all the history, and enough documents that nobody can see which version is authoritative. That is exactly when colleagues start using AI to look things up quickly, and exactly when the difference in answers becomes a risk rather than an annoyance.

The test is simple. Ask two people and two AI tools the same question, one you know the answer to. Get the same answer four times, and your knowledge is in good shape. Get four variations, and you know where the problem lies: not with the AI.