Starting with AI is not a technology decision. It is a problem decision. Pick one task that costs you real money or hours every week, test a cheap tool against it for a fortnight, and only spend real budget once that test shows a genuine saving. Most UK businesses should begin on free tools, not a build.
We say this as the people who get paid to build the systems. The order below is the one that protects your money, and it happens to be the one most likely to work.
How do I start using AI in my business?
Start with the problem, not the tool. Write down the single most repetitive, expensive, or error-prone task in your week. Put a number on it — hours, missed deals, rework. Then run the cheapest possible experiment against that number before you commit a penny to software or a build.
Here is the whole method in five steps. The rest of this guide expands each one.
- Name one costed problem. Not "we should use AI" — "we lose six hours a week manually chasing quotes."
- Check you have the raw material. Is the data or content the task needs actually written down somewhere a tool could read?
- Run a two-week experiment on a free or cheap tool. Measure the before and after.
- Decide the spend tier based on what the experiment showed — keep the tool, buy an automation, or scope a build.
- Give it an owner. One human is accountable for the process and the results, or it drifts.
The businesses that waste money skip straight to step four. They buy the platform, then hunt for a problem it might solve. Reverse it.
Why start with the problem, not the technology?
Because the technology is cheap and abundant, and the problem is the scarce thing. There are hundreds of capable AI tools. There is only one version of your business, with its own bottlenecks, data, and margins. The value comes entirely from matching a specific tool to a specific costed pain — and that match is impossible to judge until the pain is written down as a number.
A problem-first brief looks like this:
- The task: "Sales replies to inbound enquiries."
- The cost: "First reply takes 40 minutes and often slips past a day. We think we lose 2–3 deals a month to slow response."
- The measure: "Time-to-first-reply, and number of enquiries that go cold."
- The bar: "Cut first-reply time to under 5 minutes without sounding like a robot."
Now any tool can be tested against that. Without it, you are shopping in the dark, and every vendor's demo will look impressive because you have nothing to hold it to.
There is a second reason. Starting with the problem tells you early when AI is the wrong answer — when a simple rule, a form change, or a process fix would do the job cheaper and more reliably. That honesty saves the biggest money of all. We wrote a whole guide on the signals: when AI is the wrong tool.
What does a realistic first 90 days look like?
A first 90 days should end with one working improvement in production and a clear read on whether a bigger investment is worth it. Not a strategy deck, not a platform rollout — one real thing that saves measured time, plus the evidence to decide what comes next.
| Phase | Weeks | What you actually do | What you have at the end |
|---|---|---|---|
| Find the problem | 1–2 | List every repetitive task; cost the top three; pick one | A one-line costed brief |
| Cheap experiment | 3–4 | Test a free/low-cost tool against the brief; measure | A yes/no on whether AI moves the number |
| Make it real | 5–9 | Put the winning approach into the daily workflow; train the owner | One improvement live, being used |
| Measure and decide | 10–13 | Compare before/after; decide keep, expand, or build | Evidence for the next spend decision |
Two rules keep this honest. First, only one problem at a time — parallel pilots dilute attention and none get properly measured. Second, the owner runs it, not a consultant. If the person who lives with the process isn't driving, you learn nothing durable.
Most businesses find that the first improvement is smaller than they imagined and pays back faster than they feared. That is the point. You are buying evidence cheaply before you buy anything expensive.
How much does it cost to start using AI?
Start at zero and climb only when the evidence justifies it. The spend ladder below runs from free tools to a full production platform. The mistake is jumping tiers — commissioning a build to solve a problem a £30/month tool would have handled.
| Tier | Typical cost | What it buys | When it's right |
|---|---|---|---|
| Free / cheap tools | £0–£50/month | ChatGPT, Claude, off-the-shelf assistants | Proving a problem is real; light, low-stakes tasks |
| Paid tool + config | £50–£500/month | Team seats, a proper subscription, light setup | A tool clearly works and you want it in daily use |
| Fixed-price automation | ~£3,000 | A built workflow — forms, routing, drafting, data extraction | Repetitive glue work across tools; a clear, bounded job |
| Fixed-fee AI System Audit | £8,000 | 2 weeks, a 30-page report scoping a bigger opportunity | You suspect a real system is worth building and want it de-risked first |
| Custom build | £12,500 → £100k | A web app or production platform you own outright | A proven, valuable process worth owning end to end |
A few things worth knowing about how we price, because it changes how you should plan. Our tiers are fixed-price — you know the number before we start, and if the work overruns, we absorb it, not you. The AI System Audit is a £8,000 fixed fee, split 50/50, and it exists precisely so you don't spend build money on the wrong build. And the AI Readiness Assessment is free — it is the sensible first conversation before any of the paid tiers.
If you want the full breakdown of what drives a build price, we keep an honest pricing guide with worked examples.
How do I run a cheap experiment before committing?
Run a two-week test that costs almost nothing and answers one question: does AI move the number in my brief? You do not need a project plan. You need a task, a tool, a baseline measurement, and a fortnight.
Here is a concrete example. Say the problem is "our team spends too long summarising long client emails and documents before acting on them."
- Baseline (day 1). Time how long five real documents take to read and summarise by hand. Say it averages 12 minutes each.
- Set up (day 1, ~20 minutes). Open a free or low-cost assistant. Paste in one document with a clear instruction: "Summarise this in five bullet points, flag any deadline, and list any action for us."
- Run for two weeks. Use it on real documents as they come in. Keep the owner checking the output, not trusting it blindly.
- Measure (day 14). Time-per-document now, and how often the summary was wrong or needed heavy correction.
You are looking for two things: a clear time saving and an error rate low enough to trust with a human check. If summaries drop to 3 minutes each and are reliably right, you have your signal. If they are fast but frequently wrong on the thing that matters — the deadline, the number — that tells you AI needs more structure here, which is exactly the kind of finding an audit is for.
The whole test costs a few pounds and a few hours of attention. That is the correct price for de-risking a decision. For a menu of tests like this, see our AI quick wins guide.
How do I avoid buying the wrong thing?
Refuse to buy anything you can't tie to a costed problem and a measured test. Nearly every expensive AI mistake comes from reversing that order — buying capability first and looking for a use later. A few guardrails keep you honest.
- No brief, no budget. If you can't write the one-line costed problem, you are not ready to spend. Go back to step one.
- Buy the smallest thing that could work. Try the £30 tool before the £3,000 automation, and the £3,000 automation before the £30,000 build. Climb the ladder; don't leap it.
- Own what you pay to build. If you commission a build, you should own the code and the infrastructure outright. Rented black boxes trap you. (We hand over everything — code and infrastructure — as standard.)
- Fixed scope, fixed price. Open-ended "AI transformation" engagements are where budgets disappear. Insist on a defined deliverable and a number.
- Watch for the demo trap. A polished vendor demo runs on the vendor's tidy data, not your messy reality. Judge tools on your own documents, in a test, not on a sales call.
- Mind the regulations early. If the work touches personal data or high-stakes decisions, UK GDPR and the EU AI Act shape what you can build. Better to know before you spend, not after.
One more: beware buying a build to fix a process problem. If the real issue is that nobody owns the enquiry inbox, no amount of AI fixes it — you need an owner and a rule first. Sorting the process, then automating the sorted process, is almost always the right sequence.
What should I not automate first?
Leave anything expensive-to-get-wrong, judgement-heavy, or data-poor until later. Early wins should be forgiving. The point of a first project is to learn cheaply, and you can't learn cheaply from a task where a single wrong answer costs you a customer or a fine.
Good first candidates — repetitive, high-volume, cheap to get slightly wrong, and something a human still checks:
- Drafting first-pass email or enquiry replies
- Summarising documents, calls, or long threads
- Tagging and routing incoming enquiries
- Extracting structured data from forms or invoices
- First-draft content that a human edits before it ships
Hold off on these until you have proof and structure:
- Anything that makes a final decision unsupervised (pricing, credit, hiring)
- Tasks where the data doesn't exist in writing yet
- High-stakes advice — legal, medical, financial — sent straight to a customer
- Work where a wrong answer is expensive, hard to spot, or hard to undo
If your instinct is to start with one of the second group, that is a strong sign you should book a proper look at it rather than a quick tool test. It might be a great project — it is just not a first project.
Where does this leave you?
Starting with AI well is unglamorous. One costed problem, one cheap test, one owner, and the discipline to climb the spend ladder instead of leaping it. Do that and you will either land a real saving or learn — for a few pounds — that AI wasn't the answer here. Both outcomes are wins.
If you want a second pair of eyes on where AI would genuinely pay off in your business, our AI Readiness Assessment is free, and it is deliberately built to send you away with free tools if that is the honest answer. Before that, the 10-point readiness checklist will tell you whether now is even the right time.
FAQ
How do I start using AI in my business? Pick one expensive, repetitive problem you can measure — hours spent, deals missed, errors made. Test it for a week or two with a free or cheap tool before committing budget. If the test saves real time, scale it: a paid tool, a small automation, or a proper build. Start with the problem, never the technology.
How much does it cost to start using AI? Nothing, to begin with. Most businesses should start on free or low-cost tools (£0–£50/month) to prove the problem is real. A first automation typically runs a few thousand pounds; a fixed-fee audit of a bigger opportunity is around £8,000; a full custom build starts around £12,500 and climbs with scope.
What is the first thing I should automate with AI? The task that is repetitive, rules-light, high-volume, and cheap to get slightly wrong — drafting first-pass replies, summarising documents, tagging enquiries, extracting data from forms. Avoid anything where a wrong answer is expensive or where you have no clean data to work from.
Do I need a data scientist to start using AI? No. For most UK businesses the useful work is applying existing models through off-the-shelf tools or a small custom layer. You need a person who owns the process being changed, not a research team. Specialist hiring only makes sense once you have proven value and volume.
How long before AI saves my business money? A well-chosen first experiment shows a signal within a week or two — you can feel the time saved. A production automation usually pays back within a few months if it targets a genuine cost. If you can't see the saving in the first month of a pilot, you probably picked the wrong problem.
Is my business too small to use AI? No. Small businesses often see faster returns because one person wears many hats, so removing an hour of admin a day is immediately felt. The barrier is rarely size — it is having a specific, measurable problem and someone with time to run a short experiment.
What's the biggest mistake when starting with AI? Buying the technology before defining the problem. People sign up for a platform because a competitor did, then look for a use for it. Reverse that: name the costed problem first, then choose the smallest tool that could solve it.