How to Use AI in Your Business: The Four Rungs From Asking to Running

You already pay for AI. It is sitting in the browser tab you keep open, and quite possibly in the app launcher your Microsoft 365 subscription came with. And yet your Tuesday looks almost exactly like your Tuesday looked last year.

I meet this constantly. An owner has done everything the articles told them to do, has a chatbot open most of the day, and still cannot point to a single hour it gave back. They assume the problem is them, or the tool, or that the hype was always nonsense.

It is none of those. The problem is that almost every guide to this topic treats using AI as one thing, when it is really four things stacked on top of each other, and only the top two hand you hours. So here is how to use AI in your business past the chatbot stage: what each rung actually is, what it costs, what the ICO expects of you, and which rung your business is realistically ready to climb to this month.

What does using AI in your business actually mean?

Using AI in your business means handing a defined piece of work to a model and getting back something you can use. That covers a huge range, from typing a question into a chat window to a script that reconciles supplier statements at six in the morning while you are asleep.

The word “using” hides that range, and the range is the whole story. One question sorts it cleanly: were you in the room when the work happened?

If you had to be there, you bought yourself a faster typist. Useful, genuinely, but it scales with your attention and disappears the first busy week. If the work happened without you, you bought back time you can spend on something else. Everything below is organised by that difference.

The four rungs, from asking a chatbot to work that runs without you

Rung one is asking. You open a chat window, describe what you want, and copy the answer somewhere. It takes no setup and it is genuinely useful for thinking out loud or unsticking a blank page. It is also completely dependent on you remembering to do it.

Rung two is drafting in place, where the AI lives inside software you already pay for. Copilot Chat summarising a forty message email thread in Outlook. Gemini writing the first pass of a reply in Gmail. The Teams recap that turns a call into actions. The gain here is that the context comes along automatically, so you stop pasting your own documents into a browser.

Rung three is where the shape of the thing changes. You write the instructions down once, in a file, and the machine follows them the same way next month. For a non technical owner this is where Claude Code belongs: you describe the job in plain English, it writes and runs the actual script, and the brief stays on disk. The second time you need it, it is one line.

Rung four is when something other than you decides to start it. A file lands in a folder. A form gets submitted. Monday morning arrives. You only look when something goes wrong.

Now the number that explains why this ladder matters. The ONS Business Insights and Conditions Survey, run in late June 2026, found that around 35 percent of UK businesses with ten or more employees use at least one AI technology. At 250 staff and above it is 49 percent. But among businesses already using AI, only about 10 percent describe that use as extensive.

Read that again. Adoption is broad and shallow. Nearly all of it is rung one.

Which job should you hand over first?

Hand over the job you have already explained to someone else. If you have trained a new starter on it, written it in a checklist, or caught yourself saying “it is the same every time” out loud, it is a candidate. If you have never articulated the process, AI cannot infer it, and you will spend three frustrated evenings discovering that.

That single filter kills most of the exciting ideas and saves you a fortnight.

The jobs that fit tend to be unglamorous. Reconciling supplier statements against purchase invoices. Turning an enquiry email into a first draft quote using your actual price list. Converting a site survey voice note into a structured job sheet. Categorising the bank feed. Pulling four spreadsheets into the report you rebuild every Monday because the numbers live in different places.

What to leave alone at the start: anything that reaches a customer without a human reading it first, and anything where being quietly wrong costs real money before you notice. Those jobs are automatable, but they are not your first one. You want the failure mode of your first build to be mildly annoying, not expensive.

If you want a quick sense of where the hours actually sit in your own week before you pick, the AI Opportunity Scorecard takes about two minutes and gives you an estimate rather than a sales call.

What does it cost to use AI in a UK business?

Less than owners expect at rungs one to three, and the subscription is rarely the expensive part. Here is what the tools most UK businesses actually touch are listed at right now.

So one owner getting to rung three costs roughly the price of a couple of client lunches a month. That is why “we cannot afford to try” almost never holds up.

The real cost is the hour you spend writing the brief properly the first time. That hour is the whole investment, and it is also the part nobody sells you, because nobody can do it for you as well as you can.

One warning on Copilot. It is an add on, not a product. If you are not already on Business Standard, Business Premium or an E plan, the headline figure is not what you will pay.

What does the ICO expect if you point AI at customer data?

There is no permission to apply for and no register to join. What you do need is a lawful basis for the processing and a privacy notice that actually describes what you are doing. On top of that, a Data Protection Impact Assessment becomes mandatory where the use is high risk, which in practice means automated decisions that have a significant effect on a person, or sensitive data processed at scale.

The ICO publishes detailed guidance on artificial intelligence and data protection, plus a free AI and Data Protection Risk Toolkit that walks a small team through the assessment without a lawyer. Use it before you pipe your customer list anywhere.

Two practical points that catch owners out. First, check the training setting on whatever account you are using, because consumer and business tiers of the same product often differ on whether your inputs can be used to improve the model. Second, the responsibility sits with you, not the vendor. If your enquiry triage bot mishandles someone’s data, the ICO comes to your door.

None of this is a reason to hold off. It is a reason to start on internal work, where the data is yours and the stakes are lower, and to reach customer facing use once you know what the tool actually does.

Why does so much of this never survive past a trial?

Because it stays on rung one, where every result depends on a human remembering to ask. MIT’s GenAI Divide study, which analysed 300 public deployments alongside surveys of business leaders, found that 95 percent of pilots produced no measurable profit and loss impact. More than 80 percent of organisations had piloted something like ChatGPT or Copilot. Almost none had changed how the work actually got done.

The authors were blunt about the cause. It was not model quality and it was not regulation. It was approach.

I see the small business version of this often. Someone spends a fortnight producing a brilliant prompt, gets a great result, shows the team, and then the quarter gets busy and the prompt lives in a Notes app nobody opens again. Nothing was ever written down as a process, so nothing survived the first busy Monday.

The same study found something more encouraging. Builds that combined internal people who knew the business with outside expertise succeeded around 67 percent of the time, against 22 percent for projects handed entirely to an IT function. Knowing the work matters more than knowing the technology, which is good news if you are the owner.

What should your first month look like?

1. Week one, pick the job. One task you have already explained to someone. Write the explanation down as if you were training a new starter, including the exceptions and what “done properly” looks like.
2. Week one, do it by hand with AI watching. Run the task once in a chat window with the real file attached. You are checking whether the output is usable at all before you invest in making it repeatable.
3. Week two, turn the explanation into a brief. Move your written instructions into a file and have Claude Code work from it. Run it, correct what is wrong, and have it update the brief rather than fixing the output by hand. That distinction is the entire skill.
4. Week three, run it twice on real work. Different inputs, same brief. If it survives two genuinely different weeks without you rewriting it, it is real.
5. Week four, decide about rung four. Only automate the trigger once the recipe is boring. Scheduling something unreliable just means it fails while you are not watching.

Most owners try to jump to step five in week one. That is how you end up in the 95 percent.

Should you learn to build this, or pay someone to?

Learn the brief yourself, and pay for the wiring. The part that compounds is knowing how to describe a piece of your own business precisely enough that a machine can repeat it, and no consultant can hand you that, because it is made of knowledge only you have.

I say this as someone who sells the help. It is still true.

Where paying makes sense is the reliability engineering. Making something survive an API being down at two in the morning, or connecting two systems that were never designed to talk to each other, is a real skill, and buying a few hours of it is cheaper than losing a month to it. What you should refuse is the arrangement where the thing works, you do not understand it, and you are renting the understanding forever.

If you want the first build done alongside you rather than handed over finished, that is exactly what the AI Opportunity Audit is: a fixed fee session that finds the highest impact task in your business, builds one live in Claude Code so you can see how it is done, and leaves you with a prioritised roadmap.

Where I would start if this were your business

I would ignore the tool question entirely for the first fortnight. Whichever chatbot you already pay for is fine, and switching costs you a week while teaching you nothing. Pick the dullest repetitive job on your desk, the one you could describe in your sleep, and get it onto rung three. Just that one.

The reason I push this so hard is that I watched it happen to my own week before I ever sold it to anyone. The first thing I automated properly was not clever. It was a reconciliation job I had been doing by hand on a Sunday evening for longer than I want to admit. What changed was not the hour it saved. It was that I stopped seeing my week as a fixed set of obligations and started seeing which parts of it were genuinely mine.

That is the shift worth chasing, and it starts with one boring task and a written brief. If you get stuck partway up the ladder, or you want a second opinion on which job to pick first, my background and how I work will tell you whether I am the right person to ask.