- Self-reported AI use among UK businesses with 10 or more staff has risen from around 12% to around 35% since late 2023, and the barriers firms name are use cases, cost and expertise: the gaps an agency should close.
- Gartner estimates only about 130 of the thousands of agentic AI vendors are real, so ask for evidence rather than a demo.
- Keep ownership of what's built, get a written processor contract and start with a small paid project.
An AI automation agency finds the repetitive, rules-heavy work in your business, builds AI and ordinary automation to take it over, connects it to the systems you already use and stays around to keep it running. Choosing one comes down to three things: evidence that it has done work like yours, a quote you can compare line by line, and terms that leave you owning what it builds. We run one, and the uncomfortable truth is that the category is crowded with firms that renamed a chatbot and learned to say agentic.
What a good agency actually does
The tell of a good agency is that it tries to talk you out of things. It pushes back on automating work that should stay human, because an automation that annoys your customers costs more than the time it saves. It measures results in cost per task rather than hours saved, since hours saved is easy to inflate and cost per task isn't. And it shows you running systems, not slides.
Most of its effort goes in after the build. Integration, testing, rollout and support are the bulk of a real project, and our guide to what automation services include walks through each stage. An agency that prices the build and waves a hand at the rest is quoting for the easy part.
Ask for evidence that matches your workflow
A polished demo shows design skill. It can't show how a system behaves with your data, your staff and your edge cases. Ask the agency to sort its portfolio into four piles: concepts, working demonstrations, deployed features, and integrations built but still waiting to be switched on. Then ask what it actually designed, built or optimised in each.
Match the evidence to your sector. For a property workflow, look for experience with tenant and landlord communication, records and who can see what. For a legal practice, ask how the team handled permissions, confidentiality and the change for fee earners. A famous client logo is context. The scope of the work is the evidence.
The red flags
- Guaranteed outcomes before anyone has looked at your process.
- Pricing with no visible link to the value delivered.
- A demo that turns out to be a rebadged chatbot wearing your logo.
- No plan for life after go-live, when keeping a system reliable becomes the real work.
- Vague answers about who owns the code, prompts and data, or lock-in you can't walk away from.
- Savings claims with no baseline, sample period or task boundary behind them.
- A promise that no person will ever need to be involved.
The pretending is widespread. Gartner estimates only about 130 of the thousands of agentic AI vendors are real, and describes "agent washing" as the rebranding of existing products such as AI assistants, RPA and chatbots without substantial agentic capabilities (Gartner). For technical agent builds, the guide to choosing an agent development company has more sorting questions.
AI consultancy or automation agency?
The two labels overlap, and plenty of firms use both. The useful split is what you get at the end. An AI consultancy hands you advice: which processes to automate, in what order, with what governance. An automation agency hands you a running system and keeps it running. Some consultancies build; many stop at the recommendation. Ask which you're buying before you compare quotes, because a strategy deck and a working integration aren't priced on the same basis.
Either way, one document should come out of the early work: a data protection impact assessment. The ICO's guidance on AI says that In the vast majority of cases, the use of AI will involve a type of processing likely to result in a high risk, which is what makes a DPIA a legal requirement under UK GDPR (ICO: AI accountability and governance). Where a project genuinely isn't high risk, the ICO still expects that assessment written down. A consultancy or agency can draft it with you. It can't own it: the accountability stays with your business as the controller. So a supplier that has never mentioned a DPIA by the second meeting hasn't thought about where your customers' data goes.
Agency, freelancer or in-house
We won't pretend an agency is always the answer. If you have the skills in-house and the need is simple, build it yourself and keep the money. For anything non-trivial the evidence leans the other way: MIT's 2025 research on enterprise AI found that tools built with external partners reached deployment about 67% of the time, against about 33% for tools built in-house.
| Route | Suits | Watch for |
|---|---|---|
| In-house | Simple needs and a team with delivery skills | Pilots that never reach daily use |
| Freelancer | A bounded build with one or two connections | Support, credentials and handover when they move on |
| Agency | Multi-system work, customer-facing systems and ongoing support | Resold tools, lock-in and quotes that stop at the build |
A freelancer can be the right call for a tidy, bounded job. Ask who supports it when they're on another contract, who holds the credentials, and whether the documentation would let someone else take over tomorrow.
Read the quote line by line
Compare like with like: discovery, build, integrations, data migration, testing, training and handover. List the recurring costs separately, including hosting, model usage, messaging charges, subscription tiers and maintenance. Ask how the price moves if task volume doubles or documents get longer.
There's no reliable universal price for "an AI agent". A narrow internal drafting tool and a customer-facing workflow wired into three systems are different jobs. Ask for the assumptions, exclusions and change-control process in writing, and consider paying for a short discovery phase before a bigger commitment. The pricing section of our services guide explains fixed-fee, day-rate and retainer models.
UK buyers: data, contracts and ownership
The ONS reports that self-reported AI use among UK businesses with 10 or more employees has risen from around 12% to around 35% since late 2023, and that the most commonly reported barriers are "difficulty identifying business use cases, cost, a lack of expertise" (ONS). Those are exactly the gaps an agency is paid to close, so ask how it will close each one. Then settle the paperwork.
- A processor contract. If the agency or its tools handle personal data for you, the ICO says a written contract is needed "every time a controller uses a processor to process personal data" (ICO: contracts).
- Where the data goes. An agency's UK address doesn't decide where its model and hosting providers process your data. Transfers to organisations outside the UK are restricted transfers, which the ICO says need adequacy regulations, appropriate safeguards or an exception (ICO: international transfers). Ask for the provider list and locations.
- Ownership and exit. Agree who owns the custom code, configuration, prompts, evaluation cases and documentation, and check you can export your records and move to another supplier. Some components will stay under third-party terms and licences; the contract should say which.
- Who controls the accounts. Model, hosting and messaging accounts should be in your name or transferable to you, with access you can revoke.
Start with a small paid project
A bounded first project shows how a supplier handles scope, feedback and surprises before you commit to anything large. Agree one useful outcome, a set of representative test cases and a review date. Judge the complete result, including the hours your team spends correcting or administering it. If the numbers only work while the agency is watching, they don't work.
Projects tend to stall between a working pilot and a system in daily use; the implementation section of our services guide has the IDC numbers. What decides it is whether testing, staged rollout and monitoring were in the plan from day one, which is how we run our own work, as set out in how we govern a workforce of AI workers.
What Orwell brings to the conversation
Orwell works across website development, custom software, AI agents and automation, with a London and Dubai presence. Our selected projects set out the scope of work for Eastmans, Walthams, Socialroom, Cranbrook Legal and Klientflo, and our delivery process runs through discovery to support.
A free 30-minute discovery call is a fair way to test fit. Bring one process, a few real examples and the result you want to improve, and ask us every question on this page. For ownership, timing and support arrangements, see the working with Orwell FAQs.
Common questions
What does an AI automation agency actually do?
A good one finds the repetitive, rules-heavy work in your business, builds AI and automation to take it over, connects it to your existing systems and stays to keep it running. The build is a fraction of the job; integration, testing, rollout and support make up the rest.
How do I tell a good AI automation agency from a bad one?
A good one tries to talk you out of automating things that should stay human, measures results in cost per task rather than hours saved, and shows running systems rather than slides. It can also separate its concepts and demos from deployed work, and say exactly what it built.
What are the red flags with an AI automation agency?
Guaranteed outcomes before anyone has studied your process, pricing with no link to value, a demo that is a rebadged chatbot, and no plan for life after go-live. Vague answers on who owns the code, prompts and data, or lock-in you cannot leave, are warnings too.
Is an agency better than building automation in-house?
Not always. With in-house skills and a simple need, build it yourself. For non-trivial work, MIT research found externally sourced AI tools succeeded far more often than internally built ones, so an agency usually pays, provided you keep ownership of what it builds and avoid lock-in.
What should a UK business check before signing with an AI agency?
A written processor contract if personal data is involved, the list of providers that will process your data and where they are, clear ownership of code, prompts and documentation, and an exit route. Then run a small paid project with agreed test cases before committing to anything larger.
Before your first agency call, write down one process, how many times a month it runs and what each run costs you now. An agency that asks for those three numbers before talking about tools deserves a second meeting. One that never asks has told you something as well.
Sources
Every figure in this article links back to the source below it was checked against.
Written for Orwell Lab’s practical guide collection. Figures come from the ONS, Gartner and MIT research linked or named beside them; data-protection points quote the ICO and are not legal advice.
Discuss your own requirements ↗