Custom GPTs for business: when they're worth building

What a custom GPT actually is

Three things in a wrapper: a set of instructions that tell it how to behave, a small library of files it can look things up in, and optionally a connection out to another system so it can fetch or send data. You build it through a form inside ChatGPT, mostly by describing what you want in plain language. No code required for the first two parts.

What it isn't, and this trips people up constantly: it isn't training. The model doesn't absorb your documents or get smarter about your business over time. Every conversation starts fresh and looks things up in the files you gave it. That's why the quality of a custom GPT is almost entirely the quality of its instructions and its reference material, not anything clever in the setup.

Most businesses that ask me about custom GPTs belong one row higher than they think. Work down the table, not up it.

Business uses where they earn their keep

• The SOP answerer. Load your procedures and let staff ask it how something is done instead of interrupting the one person who knows. It pays for itself in a business where the same five questions get asked every week.
• The brand-voice drafter. Instructions covering tone and structure, plus a handful of examples of writing you're happy with. Output stops sounding like everyone else's AI copy because it has your examples to imitate.
• The onboarding assistant. New starters ask it the awkward small questions they'd rather not ask a person in week one. Cheap to build, immediately useful, low risk if it gets something slightly wrong.
• The quote or proposal drafter. Your price list and three winning examples as knowledge, your rules as instructions. The person still checks the numbers, every time.
• The document reviewer. A GPT that reads a draft against your own checklist and lists what's missing. Checklists are exactly the kind of stable reference material this format handles well.

The limits worth knowing before you start

• The knowledge goes stale. Files are a snapshot. Change your price list and the GPT keeps quoting the old one until somebody re-uploads it. Decide up front who owns that job.
• There's no permission model inside it. Everyone who can open the GPT can effectively reach everything you loaded into it. Don't put salary data or anything client-confidential in a GPT shared with the whole company.
• Retrieval isn't perfect. Give it forty documents and it will sometimes answer from the wrong one. A tight, curated set beats everything you own.
• Sharing depends on your plan. Building needs a paid plan, and keeping a GPT private to your team means a business or enterprise workspace. On a personal plan your options are a link or the public store.
• It can't enforce a process. Nothing stops a user ignoring it. If a step must happen every time and be provable afterwards, that's software, not a GPT.
• It isn't a moat. Anyone can build the same thing in an afternoon. The advantage is your reference material and your process, never the GPT itself.

Build it yourself, or get advice first

If the task is clear and the reference material already exists, build it yourself this week. Write the instructions as though you're briefing a new employee, upload four or five documents rather than forty, then have two people use it for real work and rewrite the instructions based on where it went wrong. That loop matters more than anything you do on day one.

Get advice first if you're about to build six of them, if the material is confidential, or if someone has quoted you real money to build one. That last case comes up more than it should: a custom GPT is often sold as a bespoke AI system when it's a form with instructions in it. Knowing which side of that line you're standing on is worth an hour of someone's time.

My work here is advice: what to use AI for, in what order, and what to avoid, before you spend on the wrong thing. I don't sell the software I recommend. If you decide you need a real build afterwards, that's a separate engagement with its own scope, price and contract. Before any of it, make sure the underlying use case is proven manually, which is what the use cases guide is for.