Most AI automation pitches sell hours saved and promise nobody’s job will change. Both undersell it. Here’s what an agency does, what it costs, and where we think AI should and shouldn’t go.
AI + data
Written by
Adam Lyth
Date
Read time
11 minutes

Photo by Mapbox on Unsplash
An AI automation agency designs, builds and runs software that does repeatable work for your business. You’ll also see it sold as AI automation services, or by an AI automation company. Today that mostly means large language models (LLMs, the technology behind ChatGPT). It also means agents: AI tools that take actions, not just write text. Paying one is worth it when the work touches customers or revenue, happens often and runs on data you can trust. It isn’t worth it when a cheap tool would do the job, or when the agency keeps the keys.
You’re not early. In September 2026, 30% of UK businesses used at least one AI technology, up from 9% in September 2023, according to the ONS. The most common use among businesses with 10 or more staff was text generation with LLMs, at 21%. The ONS flags these as official statistics in development, so treat them as a direction rather than a precise count. The useful question is where AI earns its keep, and who should build it. We build this for clients and run our own business on it, so this guide takes sides.
Our project tracker hands work to an AI agent that writes the code. Our research and drafting engine uses AI and holds everything for a person to approve. That gives us four views most agencies won’t put in writing.
1. Some roles will shrink. Anyone telling you otherwise is selling something. Routing, chasing, re-keying and first drafts are going away. Judgement, taste and owning the customer relationship are growing. What you do with the time you get back is your decision, so make it on purpose.
2. “Augmentation” is often a polite word for half a job. A copilot that suggests and waits for someone to act saves minutes. Hand the AI a whole step, start to finish, and put a person at the point where a mistake would cost you. That’s where the hours come back.
3. Put AI in front of customers only where a wrong answer is cheap. Quotes, bookings, order status and opening hours, yes. Complaints, refunds, anything legal or medical, and your biggest accounts get a person, with AI working behind them.
4. The model is the easy part. The data is the project. Models are rented, improve every few months and cost pennies a task. Clean data with a clear source is what makes the output right. Skip that work and you get a system that is confidently wrong.
Most automation pitches lead with hours saved. That’s the smaller prize. The bigger one is customers getting what they came for, inside your website or product, without waiting for a person.
Picture a customer who wants a quote at 9pm, or needs to move a booking on a Sunday. If they get a useful answer straight away, they can finish there and then. If they get a contact form and a reply on Tuesday, some will have bought elsewhere.
Most businesses are automating the wrong end. They start with back-office paperwork because it feels safe. Your customers already get instant, automated service in other software they use, and they now expect it from you. A fast answer at the moment someone wants to buy is worth more than a tidier back office. The questions that matter: will it make you more competitive, will it make you more money, and how far are you from offering it?
Measure it the way you’d measure any sales change:
If an agency can’t tell you which of these it expects to move, it’s selling you activity. Our worked example of AI customer service goes through one case in more detail.
Time savings come second, but they’re real and easier to prove. Among businesses with 10 or more staff using AI, 38% said administrative and clerical roles were the most affected, says the ONS. Only 6% reported a smaller headcount as a result. Don’t take too much comfort from that. Most of those businesses use AI to help people write, not to do whole jobs, and the number will move as that changes. Decide now whether the time you free up goes into serving more customers or comes off the payroll. Both are legitimate. Drifting between them isn’t.
Two of our own results show both sides. For The Home Help Company, we automated billing and payroll, which gave the CEO 30 hours a month back. For Everyone’s Invited, AI anonymisation of submissions saved about 12,500 hours of review. Some 52,794 testimonies have been published.
Strip away the branding and a good agency does five things. You don’t need the technical detail of any of them. You do need to know what to ask at each stage.
| Stage | What you should get | What to ask |
|---|---|---|
| Review the work and the data | A short list of tasks ranked by value, and an honest view of your data | Which of these would you not automate, and why? |
| Design the workflow | What the customer or staff member sees, and where a person checks | Where does a human approve before anything goes out? |
| Build and connect | Links into your CRM (customer records system), inbox, booking system or accounts software | Whose accounts does it run in, and do we get admin access? |
| Test on real cases | Results from past examples before any customer sees it | What did it get wrong in testing, and how often? |
| Run and improve | Monitoring, fixes and updates as the AI models change | Who gets the alert when it breaks, and what does that cost? |
The last row is the one buyers forget to budget for. Connected systems change, AI providers update their models, and your process moves on. Something has to keep up.
If the review shows your data is scattered, fix that first, and be wary of any agency happy to build on top of it. Our guide to getting your data ready for AI explains what “ready” means. The AI readiness scorecard gives you a quick self-check.
An agency is one of five routes. Each suits a different situation.
| Route | Suits you when | Watch out for |
|---|---|---|
| A no-code tool you set up yourself | One simple, low-risk task, and someone has time to own it | It becomes nobody’s job when that person leaves |
| An AI feature in software you already pay for | Your existing supplier already does what you need | Bending your process to fit the feature |
| A freelancer | One well-defined workflow with a clear finish line | One person holds all the knowledge, and sometimes the logins |
| An agency or specialist partner | Several workflows, customer-facing work or sensitive data | Lock-in, vague running costs and juniors doing the work |
| A full-time hire | AI is continuous, core work at the heart of what you sell | Hard to judge, and hard to replace |
If you’re weighing a ready-made tool against something built for you, read when to buy, configure or build. For the freelancer route, see this answer on cheaper teams.
Our view on hiring is blunt. LLMs and agents are new, and the tools change every month. Hiring one person to lead AI puts all the knowledge in one head. A business new to AI also can’t easily judge whether that person is any good.
For most businesses, start with a partner who does this all the time. They review the data, set strong foundations and get it working. Then you hire someone to own it. Hiring first still makes sense when AI is continuous, core work at the heart of what you sell. There’s more in our guide to hiring an AI developer.
Budget for three things, not one: building it, the platforms it runs on, and keeping it running.
In our experience, £5,000 is the realistic lower bound for a well-managed, organised project. You’ll find quotes for a few hundred pounds. They usually buy a template wired to your inbox, with nobody accountable when it goes wrong. Where yours lands depends mostly on the state of your data and what you want the agent to do. Our AI agent cost guide breaks this down further.
If your automations run on a workflow platform such as n8n, that platform charges its own subscription. Its published prices start at €20 a month, billed annually, and climb with the number of workflow runs. Those sums are small next to a build. What matters is whose name and whose card they’re on. If the agency pays and rebills you, leaving gets complicated.
Running costs cover AI model usage, the platform subscription and someone fixing things when they break. Ask for them line by line, per month, before you sign. For AI model prices in pounds, see what an AI chatbot costs to run.
Pick one workflow and run the numbers before anyone builds anything. The figures below are made up to show the method. Use your own.
The revenue side. Say you get 200 quote requests a month and win one in five, so 40 jobs. If instant out-of-hours answers lifted that to one in four, you’d win 50. That’s ten extra jobs a month, multiplied by your margin per job.
The time side. Say three people each spend five hours a week re-keying orders. That’s about 65 hours a month. At an example cost of £25 an hour, including overheads, it comes to around £1,600 a month.
The payback. Say the build costs £10,000 and running it costs £300 a month. On time savings alone, you’re £1,300 a month better off. The build pays back in just under eight months. Any extra jobs won shorten that.
Two rules of thumb come out of that sum:
Our AI for UK SMEs guide covers the split between agents for customers and tools for staff in more depth.
Settle this before you settle the price. It decides whether you’re buying an asset or renting a dependency.
If the workflows run in the agency’s accounts, leaving can mean starting again. Before you sign, get these in writing:
Our guide to who owns your code goes deeper on contracts and accounts.
If an agency processes your customers’ personal data on your behalf, the ICO says you need a written contract with it. It must say the agency acts only on your documented instructions and keeps the data secure. It can only bring in sub-processors, meaning other firms that help it, with your prior authorisation. It must delete or return the data at the end, at your choice, and allow audits.
You stay primarily responsible for compliance, even when the agency does the work. The ICO notes this guidance is under review after the Data (Use and Access) Act. For the wider picture, read our AI and GDPR guide.
If a supplier has let you down before, this answer is worth a read before your first call.
We do AI work for existing businesses, and it starts with a free chat. You get guidance on what we’d do, from people who do this all the time. A paid discovery doesn’t have to follow. When there is one, it ends in a strategy you take away and run yourself, or a scoped plan and a price. Either costs far less than getting it wrong.
We run our own business the way we’d tell you to run yours. In October 2026 we replaced our project tracker with one we built, called Helm. A person writes a story and hands it to an AI agent, which writes the code, runs the checks and opens a pull request. In its first day of use, the agent delivered 14 stories through 12 pull requests. None went live until one of our engineers chose to merge it.
What you can hold us to:
Our founder was the founding engineer at AI Music, which Apple acquired. Results like The Home Help Company’s 30 hours a month are the standard we work to.
See our AI + data work, or talk to us. The first chat is free. Bring one workflow and the number you want it to move.
It finds the repeatable work in your business, then designs, builds and runs software that does it, usually with LLMs and agents. A good one starts with a customer or revenue number, not a tool. It also tells you which tasks to leave alone. Then it keeps things working as your systems and the AI models change.
In our experience, £5,000 is the realistic lower bound for a well-managed, organised project. Where yours lands depends mostly on the state of your data and what you want the agent to do. Platform subscriptions come on top. n8n, for example, lists plans from €20 to €667 a month, billed annually, before its custom Enterprise tier. Ask for monthly running costs line by line.
Usually not first. The tools change every month, and one hire carries all the knowledge. A business new to AI also struggles to judge whether that person is any good. Start with a partner who does this all the time to review the data and get it working. Then hire someone to own it. Hire first only if AI is continuous, core work at the heart of what you sell.
It can, and it’s better to decide that on purpose than drift into it. So far only 6% of UK businesses with 10 or more staff using AI report a smaller headcount, says the ONS. That mostly reflects AI helping people write, not doing whole jobs. Hand over whole steps and some roles will shrink. Choose whether the time goes into serving more customers or comes off the payroll, and measure the result in customers served and revenue.
Whoever the contract says, so settle it before you sign. Get the accounts in your company’s name, or at least with your admin access. Keep the workflows and code somewhere you own, and get the IP assigned to you in writing. At Contic, you own the code and IP, and our contract assigns it to your company.
It can be, with the right contract and controls. If an agency processes personal data on your behalf, the ICO says you need a written contract. It must cover acting only on your instructions, security, authorised sub-processors, deleting or returning data at the end, and audits. You stay primarily responsible for compliance.
You’re ready when a task repeats often, the data behind it is reasonably tidy, and you can name the number it should move. If your data lives in scattered spreadsheets, sort that first. The AI readiness scorecard is a quick self-check.
Start with a free chat. We’ll tell you honestly whether AI is worth doing yet, and if it is, which process we’d start with.
Our newsletter
Occasional emails about what we’re building and what we’ve learned the hard way
You may unsubscribe from these communications at any time. For information please review our privacy policy.