Why your team gets different answers from the same AI

Here’s a situation we hear about a lot. A customer asks whether they can still return an order. Anna in support asks ChatGPT and gets “within 30 days”. Erik in sales asks Copilot and gets “within 14 days, unopened”. Both paste the answer into an email. The customer now has two answers from the same company.

Nobody did anything wrong. And it’s going to keep happening, because the cause isn’t the model.

Three reasons the answers drift

The model doesn’t know your company. Out of the box, ChatGPT knows a lot about returns in general and nothing about yours. Asked about your policy, it fills the gap with what’s typical. Typical is often close enough to sound right and still be wrong.

Everyone gave it different context. Anna once pasted the returns page into her chat. Erik has an old price list in his custom instructions. A colleague uses a project with last year’s handbook. Each assistant answers from what it was told, and they were all told different things, at different times, by different people.

The answers vary even with the same context. Language models don’t return the same text twice. With a clear source that barely matters: the wording changes, the facts don’t. Without one, the variation reaches the facts too.

Why better prompts don’t fix it

The usual reaction is a prompt guide: “Always say: please check our returns policy.” Or a shared document of good prompts. That helps people ask better. It doesn’t change what the AI knows.

A prompt guide also lives in yet another place. Now you have the returns policy on the website, an older one in a PDF, a third version in someone’s custom instructions, and a guide telling people how to ask about it. Four places, no owner.

What does fix it

The answers line up when three things are true.

There’s one source. The returns rule is written down once, in one place every AI tool reads. Not copied into each tool, connected to it. Today the common way to connect is MCP (Model Context Protocol), which ChatGPT, Claude and Copilot all support in their business plans.

Someone owns it. The head of support owns returns. If sales thinks the rule should be different, that’s a conversation between two people, not two chatbots. Until the Owner approves a change, the old rule stands.

You can see what the AI followed. When an answer looks wrong, you need to know which rule it used and which version. Then you fix that rule once, and every tool gets the fix. Without that trace you’re guessing which of the four documents the AI found.

A quick test for your company

Ask three colleagues to put the same question to the AI tool they use: something with a clear right answer in your business, like a payment term, a delivery time or a discount limit. Compare the answers.

If they match, you’re in good shape. If they don’t, count how many places that rule is written down and who would have to approve a change. That number is usually the problem.

Where to start

You don’t need software for the first step. Pick the question where wrong answers hurt most. Write the rule down once, with a reason and the name of the person who owns it. Then make sure everyone’s AI reads that version and not their own copy.

Doing that for one rule is easy. Doing it for a few hundred, across teams and tools, while people keep changing them, is where it gets hard. That’s the job of a company brain, and it’s what we build with CtxCore Brain.