Guide · updated July 2026
How to choose an AI automation agency.
We are one, so treat this as a checklist written from the inside. Ask these seven questions of anyone pitching you AI, including us. The good agencies will enjoy answering them. The demo shops will change the subject.
Why the filter matters
Most business AI never makes it past the demo. Not because the technology fails, but because a demo and a production system are different products that happen to look alike in a pitch meeting. A demo needs to work once, in front of you. A production system needs to work on Tuesday at 2am, on the messy real data, unattended, at a cost someone signed off.
The UK market prices both the same: anywhere from £5k to £60k+ per system. So the price won't tell you which one you're buying. The questions below will. If you're still deciding whether an agency is even the right route, start with our comparison of agency vs in-house vs no-code first.
The due-diligence checklist
Seven questions to ask before you sign anything.
- 01
What happens after launch?
The single best filter. Demo shops go quiet at handover because the demo was the product. A production builder has a concrete answer: who monitors the system, what gets logged, what the support arrangement is, and what happens when an API the system depends on changes. If the answer is vague, the system will be too.
- 02
Who owns the code, the prompts and the data?
Insist on all three in writing. Some agencies build on their own platform and rent it back to you, which is fine if you know that going in and disastrous if you find out at the end. Ask what leaving looks like: if the honest answer is "you'd rebuild from scratch", you're buying a subscription, not a system.
- 03
Show me a production number, not a screenshot.
Anyone can screenshot a chat window. Systems that really run leave evidence: leads per run, bounce rates, rows processed, cost per thousand calls. Ask for one number from a live system and how it's measured. The agencies that have one will tell you fast. The ones that pivot to talking about their process don't have one.
- 04
How do you stop the AI making things up?
This is the question that exposes the most. A serious answer mentions grounding the system in your data, validation checks on outputs, and teaching the system to refuse when it can't support an answer. If the reply is some version of "the latest models are very accurate", walk. Model quality is not a hallucination strategy.
- 05
What does a run actually cost?
AI systems have metered costs: every model call is billed. A builder who runs real systems can tell you unit costs and what caps stop a bug becoming an invoice. If they've never thought about cost ceilings, their systems have never run unattended.
- 06
What stays behind a human decision?
Ask which actions the system can take on its own and which wait for a person. The right answer draws a hard line at anything irreversible: sending outreach, moving money, publishing publicly. An agency that automates everything by default is automating your mistakes too.
- 07
What would make you turn this project down?
An honest agency has real answers: the problem is too vague, the data doesn't exist, a £40/month off-the-shelf tool already does it. An agency with no answer will take any project, including the ones that fail. You want the people who say no to bad fits, because that means the yes means something.
Walk-away signals
Six red flags that predict a failed project.
A guaranteed ROI figure quoted before anyone has looked at your data.
Every answer routes back to a chatbot, whatever the problem was.
No unit costs. If they can't say what a run costs, nothing has ever run.
The contract is silent on who owns the code and prompts.
The case studies have adjectives where the numbers should be.
No answer to "what happens when it breaks at 2am?"
The jargon, translated
Six terms you'll hear in every pitch, in plain English.
LLM pipeline
A system that runs a language model over your data at scale with fixed steps and checked outputs: reading documents, classifying items, scoring records. The opposite of someone pasting things into a chat window.
Grounded agent
An AI assistant restricted to answering from a defined source of truth, like your live database, rather than from the model's memory. Properly built, it refuses to answer when the data can't support an answer.
Hallucination
When an AI states something false with confidence. Not a rare glitch but a default behaviour that production systems must be engineered against, with grounding, validation and refusal rules.
Evals
Repeatable tests that measure whether an AI system's outputs are actually good, run before launch and every time something changes. If an agency can't describe their evals, they're shipping on vibes.
Human-in-the-loop
A checkpoint where a person reviews or approves the system's work before it takes effect. Where the checkpoint sits, and what can bypass it, is a design decision you should be shown.
Handover
The point where the system, its code and its documentation become yours to run, with or without the agency's ongoing help. Handover-first builds are designed for this day one; platform builds never really hand over.
Put us through the checklist.
Genuinely. Bring these seven questions to a call and see how we do, or start with the production numbers in the case studies.