I sell artificial intelligence for a living, and for most of a year, I used it the way almost everyone I meet still uses it. I opened a chat window, typed a question of one or two lines, read the answer, and closed it. Sometimes the answer was useful. Usually, it was the kind of thing I could have found myself in ten minutes.

I concluded, reasonably, that the tools were overrated.

I was wrong, and the mistake was mine rather than the software’s. I was asking a colleague who had never seen my business to comment on my business.

The difference is what you give it, not what you ask it

The single change that made these tools useful to me had nothing to do with clever phrasing. I stopped asking questions in a vacuum and started giving the thing the documents.

A concrete example from my own week. I no longer ask what a good consulting proposal should contain. I hand over the client’s brief, three proposals I have written before, and the note I keep on how I price. Then I ask for a draft. What comes back is not generic advice about proposals. It is a draft of mine, in my structure, with my pricing logic, that I spend 20 minutes correcting rather than two hours writing.

The pattern is the same whatever your business. A hotelier comparing supplier contracts should be handing over the contracts. A finance director asking about last quarter should be handing over last quarter. Nobody would ask a new consultant for advice while refusing to show them the file, but that is precisely what a one-line prompt does.

Most of the current tools accept whole documents, and several will hold a folder of them so you do not start again each time. That capability is where the value sits, and it is the part almost nobody uses.

Which raises the question every executive should ask next, and most do not.

Where does that document go

If the way to get value is to hand over your contracts, your figures and your client files, then it matters a great deal what the vendor does with them. The answer is knowable, it is published, and it depends almost entirely on which plan you are on.

As things stand at the start of September 2026, and both companies publish this openly:

OpenAI uses conversations from its individual products, the free and personal paid ChatGPT plans, to train its models by default. You can switch that off in the settings, but it is on until you do. Its business products are the reverse: ChatGPT Business, Enterprise and the developer interface are not trained on unless a customer chooses to opt in.

Anthropic, which makes Claude, treats its consumer plans the other way round: your chats are used to improve the models only if you turn that setting on. Its commercial products sit outside that entirely.

Two things follow, and neither is obvious.

The first is that the defaults differ between vendors; so, a rule of thumb learned from one will mislead you about the other. The second, and more important: these are the vendor’s terms, not your policy, and they change. Both companies revised their consumer terms during the past year. Anything I write here has a date on it, and so should anything your team relies on.

The check that takes an afternoon

Here is what I would do before anything else, and it requires no technical knowledge at all.

Find out which plan your people are actually using. Not which plan the company bought. Which one they are typing into. In most organisations I walk into, someone has bought a proper business subscription for the department, and half the staff are still using a personal free account on their own phone because they had it first and it is familiar. Your carefully negotiated terms cover the account nobody is using.

Then look at the settings on those accounts, in front of the person, once. It takes minutes and it is the only way to know rather than assume.

Then decide what is not allowed in a chat window at all, and write it in one line that a busy person will remember. Not a policy document. One line. Client files, staff records, anything under a confidentiality clause, whatever your business considers the crown jewels.

None of that requires a consultant, a committee or a budget.

What comes after asking

Once you are handing over real documents and you know where they go, something else becomes possible, and it is where this is heading for most businesses over the next two years. You stop asking the tool questions and start giving it jobs. Not “summarise this contract” but “watch this inbox, and when a supplier contract arrives, check it against our standard terms and tell me what differs”.

That step, from asking to delegating, is a genuinely larger one than it looks, and it is where the interesting failures live. It is also where I will spend a good part of this column in the weeks ahead, because I am building exactly that at the moment and it has broken twice already.

But it is the second step, and there is no point taking it while your prompts are still one line long and nobody in the building knows which account they are typing into.

I spent a year on the first step. It need not take you that long.

Contact: [email protected]

Suggested for the bio box: a QR code linking to linkedin.com/in/marcisrael