Straight answers.

The loudest voices are vendors, investors and commentators talking to each other. For a business with 5 to 50 staff the useful question is much smaller: which hours of repetitive work each week could a machine do, and is the data it would need in a fit state?

Start there. Ignore the rest until it turns up in software you already use.

No. The chat box is the demo. The value is AI wired into your inbox, accounts and CRM, doing the steps you’d otherwise do by hand. That’s an agent.

Judging AI by ChatGPT is like judging the internet by a web browser’s home page. See the difference.

The big tech companies are spending enormous sums on data centres, and economists do argue about what that does to bond yields and rates. That’s one for your financial adviser, not your IT guy.

What I can tell you is the part you control: the price of using these models has fallen sharply over the past few years, and a small business doesn’t need capital investment to get started.

For drafting and thinking, off the shelf is fine. The risks are around it:

• Staff pasting client data into personal, free accounts
• No record of what was shared, or with whom
• Nobody checking the output before it goes to a customer

Fix: business accounts with the data controls checked, a one-page usage policy, then look at where custom automation earns its keep.

Yes, both. Done badly, you leak data or automate a broken process. Not done at all, your competitors quote faster, answer quicker and carry less admin.

The low-risk path is dull on purpose: tidy the foundations, pick one workflow, measure it, then expand.

Less dramatic than that, but pointing the same way. Nobody went broke in a year for not using email. They went broke slowly, quoting by fax while someone else didn’t.

It’s a non-negotiable on a timescale of years, not weeks. That’s enough time to do it properly.

The leading models are closer than the marketing suggests. Choose by where your business already lives:

• Microsoft 365 shop: Copilot is already in the building
• Google Workspace: Gemini is built in
• Writing, analysis, documents: ChatGPT and Claude are both strong

Pick one, pay for the business tier, and learn it properly. Tool-hopping costs more than any difference between them.

No. They all take plain English. The default tone differs a little between them, and you can tell any of them to be brief, casual or blunt.

The skill that matters is the same across all of them: giving clear context about the job.

Stop asking it questions and give it jobs. Open last week’s sent folder and calendar. Anything you did three times is a candidate.

Paste in the real task with the real document, say who it’s for, and describe what good looks like. Treat it like briefing a new starter, not searching Google.

Both, in that order. Give full context up front: who you are, what it’s for, what good looks like, an example if you have one. Then refine with follow-ups.

One-line questions get one-line quality.

It predicts plausible text, and plausible isn’t the same as true. For anything that matters:

• Give it your own documents to work from, not its memory
• Ask for sources, then click them
• Keep a human sign-off on numbers, legal and anything a customer sees

Running the same question through two models catches some errors. Agreement between them still isn’t proof.

Five:

• Model: the engine (GPT, Claude, Gemini)
• Prompt: your instructions
• Token: the unit you’re billed in, roughly three-quarters of a word
• Hallucination: confidently wrong
• Agent: AI with a goal and access to your tools

The rest you can pick up as you go.

Closed models (ChatGPT, Claude, Gemini) run on the vendor’s servers. Business tiers generally don’t train on your data, but it does leave the building, so read the settings.

Open-weight models (Llama, Mistral, Qwen and others) can run on your own hardware, so nothing leaves. The cost is setup, hardware and usually some capability.

My usual recommendation for a small business: a business tier of a closed model, with the data settings checked.

Mixed, and fair enough. Plain English gets you a working prototype fast. Making it secure, maintainable and not fall over at 2am is still engineering.

I build with AI every day. Decades of knowing what breaks is the part it doesn’t replace.

Thinner margins for experiments, stricter expectations about privacy, and a healthy suspicion of anything that sounds like a sales pitch.

That last one is a strength. The answer is proof on a small scale, on your own work, rather than bigger promises.

It replaces tasks well before it replaces jobs. The roles most exposed are the ones that are mostly moving data between systems.

The better play is to give your good people agents and let them spend more time on the work you hired them for. Tell staff early what you’re doing and why. Rumours do more damage than software.

Partly, yes. Once everyone has speed, speed stops being an advantage. What’s left is judgment, relationships and quality.

Automate the table stakes so your people have the time for those.

Less James Bond, more very reliable junior. An agent is AI given a goal, a set of rules and access to specific tools, such as your inbox or accounts, so it can carry out a multi-step job without you typing each step.

You decide what it can touch and where it has to stop and ask.

A whiteboard and a couple of hours. What comes in (emails, calls, forms), who touches it, which systems, where it waits, and where it goes wrong.

That map shows which jobs are worth automating, and where the foundations need work before anything gets automated.

• Board meetings: packs summarised overnight, actions tracked to owners
• Lending and credit: each application checked against policy, exceptions flagged for a human
• Customer comms: updates drafted from the job status in your system
• Sport: match stats turned into half-time notes for the coaches’ box

Same pattern each time: data in, rules applied, a person decides.

Learn it on your own work, not in a course. Short hands-on sessions using your real inbox and documents, then one small automation built alongside you so you can see how it works and own it.

That’s most of what I do. Get in touch and we’ll work out where to start.