Which AI Tool Should I Use?
Start With the Job, Not the Brand

"Which AI tool should I use?" is the question I get asked more than any other, and the honest answer starts with a question back: what are you actually trying to do?

That sounds like a dodge. It isn't. There are now more AI tools than anybody can sensibly evaluate and every one claims to be the best, so picking a tool before you've picked a job is the reason most people try AI once, get something bland back, and quietly decide it's overrated.

The numbers say a lot of people are stuck exactly there. Which? surveyed 4,189 UK adults in September 2025 and found 51% had used an AI search tool personally. Among those who hadn't, the commonest reason given wasn't cost, or privacy, or scepticism. It was that they didn't know enough about it. One in five simply didn't have a way in.

This is the way in, written for people who aren't technical and don't want to be. How to get started with AI, which tool suits which job, what the learning curve honestly looks like, how to write a good AI prompt, and how to give an AI enough context about your business that it stops writing like a brochure. It's written for sole traders and owners weighing up AI tools for small business use, and for anyone who just wants to work out what everyone else is on about.

Why "which AI should I use?" is the wrong first question

Ask ten people which AI tool is best and you'll get ten confident answers, most of them describing the one they happened to try first. That's not expertise. It's exposure.

Here's the thing that took me a while to accept: for the ordinary jobs most people want doing, the top four or five assistants are close enough that the difference in the finished work comes almost entirely from what you tell them. Give ChatGPT, Claude and Gemini the same lazy one-line request and you'll get three similar, similarly useless paragraphs. Give any one of them a proper brief and it will outperform all three of the lazy attempts.

So the difference that decides your answer usually isn't the model. It's where your information lives. If the thing you need is buried in Outlook, that points one way. If it's in Drive, another. If it's in a stack of PDFs on your desktop, a third. That's a routing question, and it has a clear answer.

I read that the week we'd spent an afternoon arguing about benchmark tables, and it landed like a mild telling-off. We'd been comparing scores from tests nobody in this office will ever sit. Mollick's point, drawn from a study where professionals graded models on realistic work, is that ChatGPT came out the better sales manager and Claude the better financial advisor. Averages hide that completely. The only benchmark that tells you anything useful is your own Tuesday afternoon, run through two tools, compared honestly.

Which AI tool for which job: writing, research, documents and images

Start here. Find your job in the left column, try the tool in the middle, and ignore everything else for now.

That last row matters more than the other nine. AI is very good at helping you arrive at a professional's office already understanding the problem. It is not a substitute for the person who carries the liability.

ChatGPT vs Claude vs Gemini vs Copilot: what each is good at

Short version, written the way I'd tell a client over the phone. Prices move constantly, so treat any figure here as "checked 14 August 2026" and see our full comparison of AI subscription prices for the buying decision.

One correction worth making early, because it trips up nearly everybody: the app is not the model. Perplexity runs other companies' models underneath. Copilot runs OpenAI's models inside Microsoft's data handling. Gemini Notebook is a product built on Gemini. When someone tells you "Perplexity is better than ChatGPT", they're often comparing a search interface with a chat interface, not two different intelligences.

Why we use ChatGPT, Claude and Gemini together

There's an assumption buried in "which AI tool should I use" that you have to marry one. You don't, and we don't.

Here's our own stack, honestly described. We use Claude for nearly all technical work, because it's the strongest at code and at following a long list of constraints without losing one halfway down. In June, Claude Fable 5 built the free AI Crawler Checker on our main site from a single brief. That tool tests 13 different AI crawlers against a site's robots.txt and against live server responses, and it worked first time, design included. I've never had that from any model before or since.

But Claude cannot draw. Not "draws badly": it has no image generation at all. So every photograph on this article came from OpenAI's image model, picked because it renders text inside an image without garbling it and doesn't default to the gloomy neon look. And when we needed eighteen consistent hero images across a locksmith's area pages, we used Gemini's image model with reference images to lock the style, and got twelve usable variants on the first pass with no regenerations.

Three vendors, three jobs, no loyalty. If you're worrying that using more than one tool means you haven't decided yet, stop. Using more than one tool is the decision.

AI desktop apps: install it, do not live in a browser tab

Almost every beginner starts on the website, and most never leave. ChatGPT, Claude and Gemini all ship desktop apps for Windows and Mac as well as phone apps, and the difference is bigger than it sounds.

What you gain is friction removed, which is a duller benefit than it deserves to be. A keyboard shortcut brings the assistant up over whatever you're already doing, so you don't lose your place. You can drag a file straight onto it. Dictation is right there, and talking is roughly three times faster than typing for most people. The conversation follows you to your phone, so a question you started at your desk can be finished in the van.

Be clear about what it doesn't change, though. Same account, same usage limits, same privacy settings, same model. Nothing gets smarter because you installed it. You just use it more, which is the only thing that actually makes you better at it.

How to write a good AI prompt: the five-part brief

Forget the word "prompt" for a minute. You're writing a brief, the same as you would for a designer or a new starter, and every rule you already know about briefing people applies.

Five parts, in this order:

1. The job. The finished thing you want, not the topic. "Draft a reply to a customer whose installation was delayed", not "help with customer service".
2. Who it's for. One line. Who they are, what they know already, how they're feeling.
3. The facts it can't know. Your prices, the order email, the policy, the dates. Anything you leave out gets omitted or invented.
4. The boundaries. What must not happen. This is the part everyone skips and the part that saves you rewriting the answer.
5. The shape. Length, format, tone, spelling. Then ask it to flag anything you should check.

Here's what that looks like written out. First the version most people send:

Write a Facebook post about my new packing service.

And here's the same request briefed properly. It takes about ninety seconds longer to write, and it's the difference between something you'd post and something you'd delete:

Draft a Facebook post for my removals company in Brewood, Staffordshire.

THE SERVICE
We now pack the whole house the day before the move. Starts at £180
for a three-bedroom house, materials included.

WHO IT'S FOR
People moving house who have run out of time. Usually families with
young children who have underestimated the packing.

BOUNDARIES
Don't promise a fixed price. Say "from £180" and that we confirm
after a video survey. Don't claim we're the cheapest.

THE SHAPE
About 80 words. Plain UK English. No exclamation marks, no emoji.
End with "Message us for a date."

Then list anything in the post I should double-check before publishing.

Notice how little of that is technique. There's no magic wording, no role-play, no "you are a world-class copywriter". It's just the information a competent person would need, plus the things you'd tell them not to do.

Prompt engineering advice that stopped being true

A lot of what's still being taught as prompt engineering was worked out on models from 2023 and has since been tested properly and found wanting. Three pieces of very popular advice are now either useless or actively harmful.

"Tell it to act as an expert"

This one is in every listicle. In December 2025, researchers at Wharton's Generative AI Labs ran expert personas across six models on graduate-level questions in physics, chemistry, biology, engineering and law. Their finding: "Expert personas don't improve factual accuracy", even when the expertise matched the question.

The more useful half of that study is what happens in the other direction. Personas implying limited knowledge, like "Toddler" or "Layperson", "consistently decreased performance across multiple models, with the degree of harm correlating to the level of implied ignorance". So "explain it to me like I'm five" is not a free simplification. You're asking a model to be worse at the thing before it answers. Ask for plain English and short sentences instead.

"Offer it a tip, or threaten it"

Sergey Brin said publicly that models do better if you threaten them. Wharton tested it in August 2025 across five models, and their conclusion was blunt: "Threats and rewards don't work." Individual questions swung wildly, some by more than thirty points in either direction, but nothing consistent came out of it. Their recommendation was to "focus on simple, clear instructions that avoid the risk of confusing the model".

"Always say think step by step"

This one used to work. It doesn't any more on current reasoning models, because they already do it internally. OpenAI's own guidance now says plainly: "Since these models perform reasoning internally, prompting them to 'think step by step' or 'explain your reasoning' is unnecessary."

The same inversion has happened with "double-check your work". Anthropic now tells developers to delete that instruction, because Claude Opus 5 verifies without being asked and the extra line just burns tokens. We covered that reversal when the model landed, in Claude Opus 5 checks its own work.

I find this whole area quietly reassuring, and I'll say why. If the winning move were a secret incantation, the people who'd win would be the ones who read the most Reddit threads. It isn't. The winning move is describing your own job clearly, which is a skill a removals foreman or a practice manager already has and a prompt-hack collector often doesn't.

How to give AI context about your business and stop generic answers

When somebody tells me AI writes rubbish for their business, the output is nearly always technically fine and completely anonymous. It could belong to any firm in the country. That isn't a model failure. It's a briefing failure, and it's the single most fixable thing on this page.

What struck me reading that is how completely it inverts the advice of the previous two years. We were all told to hunt for the perfect phrasing. Anthropic's own engineers are now saying the phrasing was never the bottleneck, and the same piece notes that "smarter models require less prescriptive engineering". The work moved. It's no longer about how you ask; it's about what you put in front of it before you ask. That's a much less glamorous skill and a far more transferable one, because it's just the discipline of keeping your own business facts written down and current.

Write your context brief once

Open a document. Spend twenty minutes. Never write it again:

ABOUT US
[Business name], a [what you do] in [town]. We serve [who] across [area].
Trading since [year]. [Number] staff.

WHAT WE SELL AND WHAT IT COSTS
[Service 1] - from £X
[Service 2] - from £Y
All prices exclude VAT. We confirm after a survey.

HOW WE SOUND
Plain English, short sentences, British spelling. No jargon, no hype,
no exclamation marks. We say "we", not "the company".

WHAT WE NEVER SAY
We don't do [thing you don't do].
We never quote a fixed price without seeing the job.
We never claim to be the cheapest or the biggest.

WHO WE'RE TALKING TO
[Customer type], usually [situation], who care most about [X].

HOW TO ANSWER ME
If you don't know something about our business, ask. Don't guess.
Flag any claim you can't support from what I've given you.

Last reviewed: [date]

Then put it somewhere permanent rather than pasting it every time:

The evidence that this works is better than you'd expect. Researchers testing GPT-4o against thirteen neurology guidelines found that on its own it cited only real sources in 56% of answers and invented sources in 42%. Give it the actual guideline documents to work from and correct citations rose to 68% while fabrications fell to 14%. That's a controlled clinical task rather than a marketing email, so don't carry those exact numbers across to your own work. Carry the direction: handing it your own material cut invented sources by two thirds.

There's a habit worth stealing here too. When Claude Fable 5 built that crawler tool for us in one go, the instruction wasn't a specification. It was a description of what the finished thing should do and how it should feel to use, and then it went and read the existing code on the site before writing any. Brief by outcome, and let it gather the context itself where it can. It works far better than trying to dictate every step.

The AI learning curve for non-technical beginners

Honestly? Shorter than people fear and longer than the marketing suggests.

If there's one line in this article I'd want a nervous 55-year-old business owner to read, it's that one. The people who struggle most with these tools in my experience aren't the least technical. They're the ones who assume it's a computer thing and that being bad with computers means being bad at this. It doesn't. The person who has spent thirty years explaining a job to a new lad on site already has the exact skill required. They just don't know it's the skill, and nobody has told them, because the field insists on calling it engineering.

Can you trust AI answers? Check before you use them

Now the part the enthusiasts skip.

In 2025, the Tow Center at Columbia gave eight AI search tools 1,600 queries asking them to identify the source of real news excerpts. Collectively they got it wrong more than 60% of the time. Perplexity did best at 37% wrong; Grok 3 got 94% wrong. The detail that should stay with you is the researchers' observation that the tools "presented inaccurate answers with alarming confidence, rarely using qualifying phrases such as 'it appears', 'it's possible', 'might'".

That was a deliberately hard, narrow test of citation retrieval, not a measure of general usefulness, and it's only fair to say so. But it tells you the failure mode precisely: you will not be able to tell a wrong answer from a right one by how sure it sounds.

So use a rule of thumb instead of a feeling. Three questions, applied to anything that leaves your desk:

1. Would it cost me money, credibility or a customer if this were wrong? If yes, check it against a source you'd defend.
2. Is there a number, a name, a date, a price or a law in it? Those are where fabrication concentrates. Verify every one.
3. Did it cite something? Open the link. A citation that doesn't support the sentence is worse than no citation, because it buys unearned confidence.

Everything else, the tone-of-voice rewrites, the brainstorms, the first drafts, needs no ceremony at all. Read it, use your judgement, move on.

Is my data safe? What AI tools do with your conversations

The most common mistaken belief I hear is that paying for a plan makes your conversations private. It usually doesn't.

On personal ChatGPT plans, including Plus and Pro, your conversations may be used to improve the models unless you switch off the setting called Improve the model for everyone, which lives under Settings then Data Controls and is on by default. Gemini keeps activity on by default, and with it on, conversations can be reviewed by humans. Perplexity has AI data retention on by default across free and paid. Business and enterprise contracts are a different matter with different defaults, but a consumer subscription is not one.

None of that means don't use these tools. It means know which door you're walking through. For a UK business, the practical rules are short:

• Don't paste customer or staff personal data because it's convenient. Redact names and account numbers first.
• Keep special-category data, passwords and payment details out entirely.
• Check the training setting on day one, on every tool, and again after a big product update.
• For anything systematic, use a business product with proper processing terms rather than a personal account.

The ICO's AI and data protection risk toolkit is the sensible UK starting point if you need to be able to show your working.

One myth worth putting down while we're here: the widely repeated claim that OpenAI must permanently keep every deleted conversation because of a court order. That broad requirement ended in September 2025, and OpenAI says UK and EEA users were excluded from the special retention that preceded it.

What AI will not tell you it left out of your website

This is where our own work comes in, and it's the thing I most want people to take away.

In March, a client came to us wanting a new website. He'd already had a go himself using AI, and here's the uncomfortable part: on the surface it didn't look bad at all. Tidy, clean, sensible pages. He was pleased with it, and I could see why.

What was missing was everything a visitor can't see. Structure. Technical SEO. Schema markup. Internal linking. Location targeting. Trust signals. AI Discovery Files, starting with the llms.txt file that states in plain text who a business is and what it does. He had no way of knowing those existed, because nothing in the process he'd used had mentioned them. The site he'd built had almost no chance of ranking for anything competitive, and no chance at all of being described accurately by an AI assistant.

We saw the same pattern from the other end last July. A hosting customer, a non-technical chap running a domestic energy assessment firm, opened a support ticket because he couldn't update his site. The site turned out to be a single HTML file with every image encoded inline, which is what one of the site-generation tools produces. He'd attached his new images to the ticket, but uploading them would have achieved nothing, because his site referenced no image files at all. It also had to be renamed to index.html before it would serve, and nothing had warned him. We wrote back explaining the architecture, sent a screencast of the upload, and he was live the next morning able to repeat it himself.

Neither of those is an argument against using AI. Both men got further on their own than they would have a few years ago. The point is narrower and more useful: AI tools finish confidently, and they don't tell you what they left out. A website that looks right can still be unreadable to the systems that decide whether anyone hears about you.

There's a third version of this that costs more slowly. A client of ours published AI-written blog posts for about a year with no context brief, no guidelines and no training. The posts were grammatical and completely generic, and most were never indexed at all. That failure was ours before it was his: we handed over a system that depended on writing discipline nobody had shown him. Which is exactly what this article is trying to prevent.

If you've built or refreshed anything with AI recently, the free check takes two minutes and tells you what a machine can actually see. Our definition of AI Visibility Checking explains what's being measured, why AI visibility matters covers what it costs to ignore, and the quick-start guide walks through fixing it if the answer is "not much". If you only ever add one file, make it identity.json, which is the machine-readable version of "this is who we are".

How to get started with AI: your first week

Five short sessions. Nothing here takes more than half an hour, and by Friday you'll know more than most people who've been dabbling for a year.

After that, the only thing that improves you is using it on real work. If you want to go further on the business side, our guide to preparing a website for AI agents covers what happens when the AI isn't a person's assistant but a buyer acting on their behalf, and the AI Discovery Files directory lists sites that have already done the groundwork. If you'd rather someone else handled the technical layer, our done-for-you setup pack researches, writes and deploys the whole set of files for your business.

Frequently asked questions