How to use AI in your business

The best way to use ChatGPT or Claude in your business in 2026 is to stop hunting for clever prompts and start treating it like a sharp new co-founder. Talk to it, hand it everything you know, and make it argue with you. Almost everything you read about “the perfect prompt” was written for weaker models and is now a waste of your time.

Here’s what changed. Two years ago the models needed careful handling, so a whole craft grew up around wording: role-task-format templates, “act as a world-class marketer” openers, acronyms like CO-STAR and RACE, and marketplaces selling prompt packs by the thousand. Then the models got good enough to read what you actually mean. The people who build these tools now say so plainly. Andrej Karpathy, one of the field’s most respected engineers, argued the real skill is “context engineering” rather than prompt engineering: filling the model in, not finding magic words. Anthropic’s own guidance notes the exact formatting of a prompt matters less as models get more capable. Ethan Mollick of Wharton, who writes the most-read practical guide going, tells people to drop the tricks and just talk to it conversationally.

So the skill moved. It’s not knowing the right words anymore. It’s giving good context, knowing what you don’t know, and getting honest pushback instead of flattery. This is how to do all three. None of it needs a course, and most small businesses are already paying for the tools: a 2025 SBE Council survey of small-business employers put AI use near 88%. The gap isn’t access. It’s habit.

1. Stop hunting for the perfect prompt

There is no perfect prompt, and chasing one is the most common way people waste their first month with AI. The model isn’t waiting for a secret phrase. It’s waiting for enough information to be useful, and you give it that by talking, not by formatting.

The instinct most people have is to type a tight one-liner, the way you’d type into Google, then judge the AI on the thin answer it gives back. That’s the wrong shape entirely. You’ll get far more out of a loose, rambling paragraph that says what you’re trying to do, why, what you’ve tried, and what “good” looks like, than out of a clever sentence. The research backs the instinct: a 2026 study found that over-engineered “improved” prompts can actually make output worse on some tasks. More wording, less result.

Our own way of working is blunt about this. We don’t think about briefing the AI well. We seed it with enough context about what we’re doing, then go completely freeform and loose. The model is good at sorting a mess into something useful. That’s its job, not yours.

The trap. Paying for prompt packs. There are marketplaces advertising 270,000-plus prompts for sale. Most are a wordy version of something you could ask in plain English, and they break the moment the model updates. You’re buying a phrasebook for a language the model already understands.

2. When you do need a prompt, get the AI to write it

When you genuinely do need a tight, reusable prompt, don’t sit there trying to craft it yourself. Tell the AI what you’re trying to achieve and get it to write the prompt for you. It’s better at it than you are, because it knows how it reads instructions better than you ever will.

There are real cases where a polished prompt earns its keep: something you’ll reuse on a repeatable task, the instructions behind a custom GPT or a Claude Project, an image prompt, the brief for an agent that runs without you watching. For those, you want something sharp. So ask for it directly: “I’m trying to do X. Write me the best prompt to get the result I’ve described, and ask me anything you need before you write it.” You’ll get something tighter than you’d have written, and you tweak it from there and keep it.

This is the real answer to step 1. The fix for “I need a good prompt” was never to go and learn prompt engineering. It’s to ask the one thing that already speaks the language to write it for you.

3. Talk to it, don’t type

Turn on voice and speak to it. This is the single change that does the most for the average person, because the reason people give the AI too little context is that typing is slow and tiring. Speaking isn’t.

The numbers are stark. Most people type around 40 words a minute. Most people speak at about 150. That’s nearly four times the context for the same effort, and context is the whole game. When you talk, you naturally include the background, the caveats and the half-thoughts you’d never bother to type, and those are exactly the bits the model needs. Mollick flat-out recommends new users start with voice mode.

Typing ~40 wpm

Speaking ~150 wpm

Every tool you’d use already has this built in. ChatGPT and Claude both have voice on their phone apps. Your laptop has dictation in the operating system (the microphone key on Mac, Windows key plus H on Windows). If you want it sharper, Wispr Flow and Aqua Voice let you dictate cleanly into any app on your machine. This is proven, easy, and the fastest upgrade on this list.

The trick within the trick is to stop performing. Don’t compose a neat spoken paragraph. Ramble. Say the thing, then the other thing you just remembered, then the worry you have about it. A stream of consciousness with all your real context beats a tidy three-sentence brief every time.

4. Say it like you’d say it to a person

Before you fire off a ramble, ask yourself one thing: if you handed this exact brief to a capable person, would they know what you actually want? If the answer’s no, the AI doesn’t either. It just won’t tell you. It’ll guess, confidently, and hand you something slightly off that you can’t quite explain.

This is the guardrail on everything above, and people get it wrong by mixing up two different things. Rambling is fine. Being vague about the outcome is not. You can be as loose and messy as you like about how you say it, as long as you’re clear in your own head about what you’re trying to get out the other end. This isn’t a contradiction of step 1. Loose wording is fine. Fuzzy intent is not. Step 1 says don’t polish the phrasing. This says know the target.

Here’s the difference in practice. “Help me with my pricing” is a ramble with no target: a person would have to ask you what you even mean. “I’m trying to work out whether to raise my prices. I’m worried I’ll lose my regulars, but I reckon I’ve been undercharging for years. Here’s what I charge now and roughly what it costs me. Talk me through whether a rise makes sense and how much.” Same loose tone, completely different odds of a useful answer, because the second one knows what done looks like. If you can’t say what done looks like, that’s the real gap, and it’s the thing to sort out first. You can even sort it out with the AI, which is exactly what the next two steps are for.

5. Brief it like a new co-founder, not a search engine

Context is the lever that moves everything, so give the AI a standing brief on your business the way you’d brief a sharp new hire on day one. Anthropic’s guidance literally frames it that way: treat the model like a brilliant new colleague who has zero context on your business. Brilliant, but it knows nothing about you until you tell it.

So tell it, once, properly. What the business does. Who your customers are and what they’re like. What you sell and at what price. Your goals for the year. The constraints you’re stuck with. How you like to sound. Spend twenty minutes on this and you stop re-explaining yourself in every chat.

The part most people miss is that you can make this permanent. In ChatGPT, custom instructions and memory hold a standing picture of you across every conversation. In Claude, you set up a Project and drop your business context into its knowledge, so every chat in that Project starts already knowing the score. We keep one of these as a general co-founder: it understands the business, the goals and the priorities, and we ramble at it about whatever’s in front of us that day. The quality jump from “cold model” to “model that knows your business” is bigger than any prompt trick you’ll ever learn.

6. Make it find your blind spots

The highest-value thing AI does for a non-expert isn’t answering your questions. It’s finding the questions you didn’t know to ask. This is the part that changes how much you get out of it, especially in areas you don’t understand well, and it’s the opposite of how most people use the thing.

Here’s the mindset that makes it work. You don’t need to know the right questions. You only need to know that you have gaps, and then make the AI surface them. Whenever we’re working in a domain we don’t own, software, tax, a new market, the first move isn’t to ask a clever question. It’s to say: “I don’t fully understand this. What am I missing? What should I be asking that I’m not? What would someone who does this for a living flag straight away?” That one move pulls up the stuff that would’ve bitten you later.

The simplest version is to make it interview you before it answers. Put this on the end of any real request: “Before you answer, ask me whatever you need to give me the best result.” It comes back with three or four questions, and answering them surfaces everything you left out of your first ramble. There’s a name for this going round, “reverse prompting”, but ignore the label. It just means letting the AI ask you things instead of only the other way round, and it works because you’re not the one who can see what’s missing from your own brief. You’re too close. It isn’t.

This isn’t a fringe idea anymore. Stanford runs a class on working with AI as a thinking partner that’s explicitly about using it to surface your assumptions and expose blind spots. The institutions teaching it have caught up with what the good operators were already doing.

7. Force it to argue back

AI agrees with you far too much, and that’s the most dangerous thing about it for a business owner making real decisions. Left alone, it will validate a shaky plan, nod along to a bad assumption, and tell you your idea is great. You have to actively force it to be a critic, because it won’t be one by default.

This is a known, measured problem, not a vibe. OpenAI had to roll back a ChatGPT update in 2025 because it had become so agreeable it was validating people’s doubts and fuelling bad impulses, behaviour the company itself called sycophantic. And once a model starts agreeing with you, it tends to keep agreeing: one 2025 study measured that pattern persisting around 78.5% of the time. The word researchers use is sycophancy. You don’t need the word. You just need to know the machine wants to please you, and that pleasing you is not the same as helping you.

Now the part most guides get wrong. Telling it “be brutally honest” barely helps. The UK’s AI Security Institute found that instructing a model not to flatter you is one of the weakest fixes there is, and worse, the more confidently you state your own view, the harder the model caves to it. So the fixes that actually work are about how you ask:

• Frame your idea as a neutral question, not a proud announcement. Not “here’s my plan to open a second location, what do you think.” Instead: “here’s a plan to open a second location. What’s wrong with it? What would have to be true for this 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.” “Steelman the opposite decision.”
• Check the facts it gives you, including its criticisms. It can invent a convincing objection as easily as a convincing agreement.

The trap. Stating your conclusion, then asking what it thinks. “I’ve decided to drop the cheaper tier, sound right?” You’ve told it the answer you want, and it will hand it back to you with confidence. Ask it to attack the decision before you’ve revealed you’ve already made it.

8. One broad context for thinking, scoped agents for the repeatable jobs

Use one rich, always-on context for thinking and strategy, and separate scoped setups for specialised, repeatable work. This is the honest split, and it’s worth getting right because cramming everything into one place quietly makes all of it worse.

The broad one is the co-founder from step 5. It knows the whole business, and it’s where you ramble, make decisions, stress-test ideas and think out loud. It’s deliberately general because thinking is general.

The scoped ones are different. When a job is narrow and you do it over and over, give it its own setup with only the context that job needs. We run a separate agent for development work and another for SEO, each loaded with just its own material, so they stay sharp and don’t drag in noise from everything else. You’d do the same with a Project called “Proposals” or “Customer replies”: one job, one clean context. The rule of thumb is simple. Broad context for thinking, narrow context for repeating. A single mega-assistant that’s meant to do everything ends up doing each thing slightly worse than a focused one would.

9. Know when the old advice still wins

For short, mechanical, repeatable tasks, a tight and explicit instruction still beats a ramble. The “just talk to it loosely” advice is for thinking and creating. It’s the wrong tool for narrow jobs with one correct shape of answer.

If you’re pulling phone numbers out of a hundred emails, sorting feedback into categories, reformatting a list, or extracting the same five fields from every invoice, be precise. Say exactly what you want and, this is the old trick that still earns its keep, show it one worked example. Give it one input and the exact output you’d want, and it’ll copy the pattern across the rest. That’s not prompt engineering theatre, it’s just being clear about a clear task. Don’t turn “talk to it like a person” into a religion. Match the approach to the job.

The honest call

You don’t need to buy anything to start. No prompt packs, no $150 bundle of “1,000 ChatGPT prompts”, no course promising to teach you the secret formula. The $20-a-month subscription you may already have does every single thing on this page. The whole industry of people selling shortcuts is selling a problem that the model updates already solved.

The only real cost is the habit. Talk to it instead of typing at it. Give it the full picture of your business, once, and keep it. Make it ask you things, and make it argue with you before you trust it. Pick one actual problem in your business this week, the kind you’d normally chew on alone, dictate the whole messy story of it, and finish with “ask me what you need to know first.” That’s the entire method. The rest is just doing it often enough that it becomes how you work.