How to use AI in your business

For a couple of years, getting good work out of AI was treated like a craft. You learned the incantations: role, task, format, “act as a world-class marketer”, acronyms like CO-STAR, and packs of a thousand prompts for the price of a nice lunch. Then the models got good enough to understand what you actually meant, and the incantations stopped mattering. The advice didn’t get the memo.

The trap. Paying for prompt packs. Most of them are a wordier version of something you could just ask in plain English, and they break the moment the model updates. You’re buying a phrasebook for a language the model already speaks.

The people building this stuff have been saying so for a while. Andrej Karpathy, one of the people who built it, says the real skill is “context engineering”, which is a grand name for telling it what’s going on. Anthropic’s own guidance says the exact wording matters less as the models get more capable. Ethan Mollick, the Wharton professor who writes the most useful practical guide going, tells people to drop the tricks and just talk to it.

So this is how I use it. None of it needs a course, and all of it works on the $20 plan you might already be paying for.

Talk to it

The single biggest change I made was to stop typing. It sounds too simple to matter, but most people give AI far too little to go on, mostly because typing is slow and a bit boring. Talking isn’t.

Most of us type at around 40 words a minute and speak at about 150. And when you talk, you edit yourself less. You throw in the background, the caveats and the half-formed worry, all the stuff you’d never bother typing, and that’s exactly the stuff the model needs. The limit was never how clever the model is. It’s how much of what’s in your head makes it onto the screen.

Typing ~40 wpm

Speaking ~150 wpm

ChatGPT and Claude both have voice on their phone apps. Your laptop has dictation built in too (the microphone key on a Mac, Windows key plus H on Windows), and if you want it cleaner, Wispr Flow and Aqua Voice let you dictate into anything. Mollick’s first advice to new users is to start with voice mode, and I think he’s right.

Don’t perform, though. Nobody needs a tidy spoken paragraph. Ramble. Say the thing, then the other thing you just remembered, then what you’re worried about. Turning a mess into something useful is what these models are good at, so let them.

Brief it once, properly

Anthropic’s advice is to treat the model like a brilliant new colleague with no context on your business. Brilliant, yes. But on day one it doesn’t know where the kettle is.

So tell it, once, properly. What you do, who your customers are, what you charge, what you’re trying to get done this year, what you’re stuck with, how you like to sound. It takes about twenty minutes. Then make it stick, with custom instructions and memory in ChatGPT or a Project in Claude, so every conversation starts already knowing the score.

I keep one of these as a kind of general co-founder. It knows the business, the goals and the priorities, and I ramble at it about whatever’s in front of me that day. For the jobs I repeat, like development and SEO, I keep separate setups that know only what that job needs. One assistant that does everything sounds efficient. In practice it does each thing a little worse.

Know what you want out the other end

There’s a catch with rambling. Being loose about how you say something is fine. Being vague about what you want is not. Before you hit send, ask whether a capable person reading it would know what you’re after. If they wouldn’t, the AI won’t either, and it won’t tell you. It’ll guess, confidently, and hand you something slightly off in a way you can’t quite put your finger on.

“Help me with my pricing” is a ramble with no destination. “I think I’ve been undercharging for years. Here’s what I charge and roughly what it costs me. I’m worried about losing my regulars. Talk me through whether a rise makes sense, and how big” is every bit as loose, and it knows where it’s going.

Let it ask the questions

The most useful thing AI does for me isn’t answering questions. It’s asking them, especially the ones I didn’t know to ask.

When I’m working in something I don’t really understand, like tax, a new market or a corner of the tech I haven’t touched, I’ve stopped trying to sound clever. I say: “I don’t really understand this. What am I missing? What would someone who does this for a living flag straight away?” The answer to that teaches me more than any question I could have written myself.

On anything that matters, I also end with: “Before you answer, ask me whatever you need to know.” It comes back with three or four questions, and answering them fills in gaps I didn’t know I’d left. It’s a funny reversal when you think about it. The tool is at its most useful the moment you stop quizzing it and let it quiz you.

Make it argue with you

This is the part that worries me. Left to itself, AI agrees with you. It’ll back a shaky plan, nod along to a bad assumption and tell you the idea is great, in a warm, articulate voice that makes it very easy to believe.

In 2025 OpenAI had to roll back a ChatGPT update that had become so agreeable it was validating people’s doubts and fuelling bad impulses. Their own explanation was that they’d leaned too hard on short-term feedback, on whether people liked an answer in the moment. Which is a polite way of saying we taught it that agreeing with us gets rewarded, and it learned. Once a model starts agreeing with you, it also tends to keep going: one 2025 study saw that pattern hold around 78.5% of the time.

The obvious fix barely works. Telling it to “be brutally honest” hardly moves it. The UK’s AI Security Institute found that asking a model not to flatter you is one of the weakest fixes there is, and that the more confidently you state your own view, the harder it caves to it. What works is changing how you ask:

• Ask a neutral question instead of making a proud announcement. Not “here’s my plan, what do you think?” but “here’s a plan. What’s wrong with it? What would have to be true for it to fail?”
• Don’t tell it which answer you’re hoping for. The moment it knows, it leans that way.
• Make it argue the other side: “Give me the strongest case against this.”
• Check what it tells you, criticism included. It can invent a convincing objection as easily as a convincing agreement.

Don’t lead the witness. “I’ve decided to drop the cheaper tier, sound right?” has already told it the answer you want, and you’ll get it back with confidence. Ask it to attack the decision before you let on that you’ve made it.

When the old advice still wins

There’s one place the old prompt advice still holds. For short, mechanical jobs with one right shape of answer, like pulling phone numbers out of a hundred emails or getting the same five fields off every invoice, be exact. Say precisely what you want and show it one worked example, one input and the output you’d want from it, and it’ll copy the pattern across the rest.

Talking to it like a person is for thinking. It isn’t a religion.

That’s really all of it

You don’t need to buy anything. The subscription you’ve probably already got does everything here. What it costs is a habit: talk instead of typing, tell it the whole story once, let it ask you things, and make it argue before you trust it.

If you want somewhere to start, take one problem you’d normally chew on alone this week, dictate the whole messy story of it, and finish with “ask me what you need to know first.” Then see what it asks.