Where Custom GPTs work for team knowledge, and where they break
Espen Oddvik · Founder
A client showed me their Custom GPT collection this spring. Nine of them. Proposal Writer, Support Replies, LinkedIn Posts, Blog Drafts, and five more with names I've forgotten. The person who built them was proud, and honestly, rightly so. The team's output had visibly improved.
Then their pricing changed.
That story contains the whole argument, so here it is stated plainly. Custom GPTs work well for team knowledge that belongs to a repeated task, like a proposal structure or a support tone. They break when company knowledge like pricing, product facts and policies gets copied into them, because each GPT keeps its own copy and copies go stale. Company Brain holds company facts in one reviewed place and delivers them through MCP to ChatGPT, Claude, Copilot and Cursor.
I'll come back to the pricing change. First, credit where it's due.
What Custom GPTs genuinely do well
A Custom GPT is a saved setup inside ChatGPT: instructions, a few reference files, a name, a link you share with the team. For a repeated task, this works. Everyone who opens Proposal Writer gets the same structure and the same rules, instead of whatever they'd have improvised. The builder does the thinking once and the whole team inherits it.
If your team lives in ChatGPT and the task is well-defined, I have no quarrel with this. Build the GPT. Enjoy it.
Where Custom GPTs break: company knowledge in the instructions
The trouble starts quietly, and it starts with a fact that belongs to the company rather than to the task.
The pricing goes into Proposal Writer, because proposals mention pricing. Then into Support Replies, because customers ask. Then into two more. Nobody decided to duplicate the company's pricing into four places. It just accumulated, one reasonable copy at a time.
So when my client's pricing changed, someone updated Proposal Writer, and the other three GPTs kept quoting the old numbers for six weeks. A customer noticed before the team did. That's the moment the collection turns from an asset into a maintenance surface, and nine GPTs is nine things that can quietly go stale.
Three more cracks show up on the same schedule. Everything lives inside ChatGPT, so the colleague who prefers Claude, and the developer in Cursor, get none of it. Everything depends on the builder, and builders change jobs. And the instructions have no review step, so whatever got typed in is now, functionally, company policy.
A rule of thumb: task knowledge in the GPT, company knowledge out
Task knowledge belongs in the GPT. Company knowledge doesn't.
"Proposals follow this structure, in this order" is task knowledge. Put it in Proposal Writer and never think about it again. "Our price is X", "our product is called Y", "we never promise Z": that's company knowledge, and every copy of it you place inside a GPT is a small debt with a due date you don't get to pick.
Company knowledge wants one home, one reviewer, and automatic delivery to every tool, which is the job we built Company Brain for. The facts live in one reviewed place and load into ChatGPT, Claude, Copilot and Cursor through MCP, the protocol that connects AI tools to shared context. Your Custom GPTs keep their task instructions and pull the company's truth from underneath, so a pricing change is one edit, everywhere, including all nine GPTs.
Company Brain's free plan includes 50 company facts and one curator, with the whole team connected through MCP, free forever. If you're weighing the two approaches side by side, our comparison of Custom GPTs and a shared company context covers where each one wins, including when a Custom GPT alone is enough.
My client still runs most of their collection, by the way. They just emptied the company facts out of them.
The GPTs got better, and nobody has quoted the old pricing since. Both things at once. That's a good trade.