Two teammates asked AI the same question, got different answers

The question was ordinary. "What's our position on using client data in AI tools?" Two people at a fourteen-person consultancy needed the answer in the same week. One asked ChatGPT, which held months of her pasted documents and chat history. The other asked Claude, working from a project he'd set up around an old services agreement. Both got confident, well-written answers within seconds. The answers were incompatible.

Her RFP response said the firm never lets client material touch third-party AI tools. His statement of work, sent the same week, described an AI-assisted review step that did exactly that, with consent language lifted from a policy draft the partners had discussed and never adopted. The same client received both documents.

Neither teammate did anything wrong, and that is the whole problem. The short version: when two teammates get different answers from AI, the AI is rarely at fault, because each model answered from the private context that person happened to feed it. Company Brain removes the divergence at the source. It holds one reviewed set of company facts and delivers it to every teammate's AI through one shared MCP address, so every assistant starts from the same approved answers.

The rest of this post is what actually diverged, why nobody saw it coming, and what the fix looks like in practice.

Why does AI give different answers to the same question?

Because no two people give it the same question. The words match. The context doesn't.

Each person's AI context is a private accumulation built over months. The onboarding deck from whenever they joined. A policy draft pasted in during some long-forgotten thread. Half of last year's positioning document. A competitor comparison a colleague shared once. Whatever the tool's memory feature decided to keep. No two of these accumulations match, none of them is versioned, and nobody can say which parts are still current, because nobody can see the whole pile at once.

The model doesn't rank any of it by freshness. It answers from what's in front of it, with the same fluent confidence whether the underlying document is this quarter's policy or a draft from two owners ago. Her ChatGPT had absorbed the older, stricter policy from a proposal she'd pasted in the spring. His Claude project contained the never-adopted draft. Both models did their job perfectly.

I wrote earlier about how a team's AI output drifts in style, where the same mechanics make one person's drafts formal and another's chatty. This post is about the more expensive half of the same problem. Style drift makes you look unpolished. Fact drift makes you wrong, in writing, on both sides of the same question, in front of the person deciding whether to hire you.

Why the contradiction stays invisible until a client sees it

AI sessions are private. Two teammates never see each other's context, each other's prompts, or each other's answers. Before AI, an unsure person asked a colleague or went hunting for the policy document, and disagreement surfaced in conversation, inside the company, where it was cheap to fix. Now each person gets a fluent answer alone, and fluency reads as settled. Nobody double-checks an answer that arrives sounding like policy.

Volume makes it worse. When answering took effort, a company produced a handful of position statements a year and senior people touched most of them. AI turned every teammate into someone who can produce a confident company position in thirty seconds. The number of stated positions went up tenfold. The number of places those statements get cross-checked stayed at zero.

So the first place two private answers meet is the client's inbox. In the consultancy's case, the client's procurement lead put the two paragraphs side by side and asked, reasonably, which one was true. The partners spent a call repairing trust that two paragraphs had spent. The actual policy took ten minutes to settle once someone finally asked the question out loud. It had simply never been written down anywhere the AI tools could see.

What changes when facts live in one governed source

The fix has two parts, and both are unglamorous.

First, the company writes its facts down as short, reviewable statements. "We do not use client material in third-party AI tools without written consent from the client." One sentence, one owner, one place. If you want a concrete starting point, my guide to writing your first 50 company facts walks through which facts to capture first and how to phrase them.

Second, and this is the part that changes the story above, new facts don't go straight in. In Company Brain, every new fact is a proposal first. Before it can be approved, it's checked against the facts already approved, and when it contradicts one, the conflict is flagged and a person decides which statement is true. The never-adopted consent draft would have collided with the existing policy the moment someone tried to add it, in front of a reviewer, months before any RFP. The contradiction still happens. It happens at write time, inside the company, where resolving it costs one conversation.

Once a fact is approved, the whole team reads from the same place. Company Brain holds your company's facts in one reviewed place and delivers them to Claude, ChatGPT, Copilot and Cursor through MCP. The team connects through one shared MCP address per company, so the March hire's assistant and the founding partner's assistant pull the same approved position, worded the same way, current as of the last review. There is a free plan, and it stays free.

The whole fix operates on the input side: every teammate's assistant starts from the same reviewed facts, which was the actual problem all along.

Where would two of your teammates get different answers today?

That's the question worth acting on this week. Pick something clients actually ask about: your refund terms, your data handling position, your current response time commitment, what happens to deliverables after a project ends. Ask two teammates to get the answer from their own AI, separately, without coordinating. Then compare.

If the answers match, good. Check another question next month. If they don't, you've found a client-facing incident before a client did, and you've also found the first fact worth writing down.

The disagreements were always there, sitting quietly in old decks and half-adopted drafts. AI made them fast, fluent, and easy to ship.