How to start an AI services agency in the UK

Most guides to starting an AI agency skip the boring half and go straight to "find your first ten clients". The boring half is where people actually get stuck, so this one goes in order.

Nothing here is advice about whether you should do it. That depends on things nobody writing an article can know about you. What follows is the sequence, the decisions inside each step, and an honest note on which parts are hard.

Step one: decide whether you are a business or a job

This sounds abstract. It is not.

If you intend to do the setup work yourself, sell it yourself, and support it yourself, you have bought yourself a job. That is a perfectly reasonable thing to buy - it is flexible, it has low overheads, and it can be done in evenings while you keep your current work. But price it and plan it as a job, not as a company that will run without you.

If you intend to build something you eventually step out of, the decisions change immediately. You document everything from client one. You never configure anything in a way only you understand. You charge enough that paying somebody else to do the setup still leaves a margin. Very few people make that decision consciously at the start, and it costs them a year later.

Step two: the legal and financial setup

In the UK this is genuinely straightforward and people over-think it.

• Sole trader or limited company. Sole trader is free and immediate - you register for Self Assessment with HMRC. A limited company costs a small fee at Companies House and takes a day. Limited looks more established to some clients and gives you liability separation. Plenty of one-person agencies start as sole traders and incorporate later. Speak to an accountant rather than an article for anything to do with tax.
• A business bank account. Not optional in practice, even as a sole trader. Mixing personal and business money makes your first tax return miserable.
• VAT. You register when you cross the threshold, not before, unless you have a specific reason to register voluntarily. Again, accountant.
• Insurance. Professional indemnity is the relevant one. If you are handling client data - and you will be, because call transcripts contain customer names and numbers - you should also read up on your UK GDPR obligations. You will likely be a data processor acting for the client, which means a written agreement.

That last point is the one people ignore. A voice agent that records calls creates personal data. Who controls it, where it is stored, how long it is kept and what the caller is told about it are real questions with real answers, and whether recording is even available depends on the provider you use and how it is configured. Get it wrong quietly and it stays wrong until it does not.

Step three: choose what you sell before you choose your tools

The temptation is to sign up for six platforms, build something impressive, and then work out who wants it. This is backwards and expensive.

Pick a small, defensible set of services. Something like: call answering, website chat capture, review requests, missed-call follow-up. Four things. All of them explainable in one sentence. All of them things a local business already knows it is bad at.

Resist adding "custom AI consulting" to the list. It sounds impressive and it is unsellable to a small firm, because they cannot picture it and cannot price it against anything.

Once the list is fixed, the tools follow from it rather than the other way round. You will need, roughly: something to run the voice agent, something to run the chat widget, something to handle automations and messaging, somewhere to host client sites if you offer that, and a way to bill monthly. The full list of what a working setup needs is longer than most people expect, which is exactly why assembling it from scratch takes months.

Step four: build the demo before you build the pitch

This is the step that changes everything and the one most often done last.

You need a working phone number that a prospect can ring, from their own phone, in front of you, that answers as a plausible business and handles a normal enquiry. Not a video. Not a slide. A number.

The reason is that this technology is genuinely hard to believe until you hear it. Descriptions produce polite scepticism. A thirty-second call produces a specific reaction - usually a pause, then "hang on, that's not a person?" You cannot manufacture that with a brochure. Listening to a demo yourself is the fastest way to understand why the conversation turns on this single asset.

How you run that demo in front of a sceptical owner matters as much as having one. Build two or three variants for different trades. A demo that answers as a plumbing firm lands differently with a plumber than a generic one. It takes an hour and it is the highest-value hour you will spend.

Step five: pricing structure

Structure first, numbers second.

The structure that fits this work is a setup fee plus a monthly fee. The setup fee covers configuration, testing, the client's specific quirks and the inevitable first-week adjustments. The monthly fee covers hosting, usage, changes, and the fact that a system nobody maintains slowly becomes wrong.

Three things to get right:

1. Do not undercharge the setup. If setup is free or nearly free, every tyre-kicker in the county will have you configuring things for people who cancel in month two. A real setup fee is a filter.
2. Know your own costs per client. Voice minutes, messages and hosting all cost you something per client per month. If you do not know the number, you cannot know whether your monthly fee is a margin or a slow leak.
3. Write down what is included. "Reasonable changes" is a phrase you will regret. Say how many changes, or say what counts as a change.

What to actually charge is a local decision and it depends on the trade, the area and what else you are doing for them. Be sceptical of any figure presented as universal.

Step six: getting the first clients

Here is the honest part. This is the hard bit, it is entirely your work, and nobody can do it for you.

The methods that work for a new agency with no reputation are unglamorous:

• People who already know you. Not to buy - to introduce. The tradespeople who did your bathroom. Your neighbour's landscaping business. Warm beats cold by an enormous margin.
• Walking in. Genuinely. A salon, a garage, a dog groomer's. Ask for two minutes, show the demo on your phone, leave a card. It is uncomfortable and it works more often than email.
• Ringing during their quiet hours. Not 8am Monday. A garage at 3pm on a Wednesday will talk to you.
• Local networking. Slow, but it compounds, and a referral from someone in the room outperforms any advert you could buy.

What generally does not work at the start: paid ads with no landing page and no proof, cold email at volume, and posting on LinkedIn about AI. These may work eventually, once you have something to point at.

Expect a low conversion rate on cold approaches and a much better one on anything warm. Expect the first client to take longer than you think and the third to take much less time than the first.

Step seven: deliver the first one properly

Over-serve client one. Not forever - just at the start. Ring them in week one. Ring them again in week three. Ask what has been odd or annoying. Fix it the same day.

You are not doing this out of niceness. You are doing it because client one is your reference, your case study, your testimonial and quite possibly the source of clients two and three. A neglected first client is an expensive mistake.

Also: write down everything you did. Every setting, every phrase, every thing that went wrong. The second setup should take half as long as the first, and that only happens if you took notes.

What people underestimate

Three things, consistently.

The time to assemble the stack. Getting a voice agent, a chat widget, an automation layer, billing and a client site to work together - reliably, repeatably, for a client who will ring you when it does not - is weeks of work, not a weekend. That is the whole reason ready-made systems exist and why what they cost is worth comparing against your own time.

Support. Clients ring. Not constantly, but they ring. Someone's opening hours changed. Someone's number went to the wrong mobile. Build the expectation that you respond within a working day and hold it.

The emotional weight of selling. Being ignored is tiring in a way that reading about being ignored is not. The people who succeed at this are usually not the most technical - they are the ones who kept ringing in week six.

Who should not do this

Anyone who wants to avoid talking to strangers. Anyone who needs income by next month. Anyone who is interested in the technology but not in small businesses - because you will spend far more time listening to a salon owner complain about no-shows than you will tuning a prompt.

If none of that puts you off, the practical questions are covered in the FAQ, and if something is not, ask.