Buying guide
How to choose an AI automation agency, and how to spot a bad one
How SMEs should choose an AI automation agency: supplier types, pricing models, the questions that expose weak firms, and contract terms.
The short answer
Choose an AI automation agency on three things: whether they diagnose before they build, whether they can integrate with the systems you already run, and whether they will still own the workflow six months after launch. Ask for a costed use-case ranking, a named engineer, an evaluation method, and full ownership of code and prompts. Avoid anyone who quotes a build before understanding the process, or who cannot explain what happens when the agent is wrong.
Key takeaways
- Four supplier types exist. Only two suit a 10 to 250 person business.
- A build quote given before a process review is a sales tactic, not an estimate.
- Insist on ownership of code, prompts, evaluations and documentation.
- Ask how they measure quality. No evaluation method means no production discipline.
- Running cost, not build cost, is what surprises SMEs in year one.
The phrase AI automation agency now covers everything from a two-person no-code shop to a former software consultancy that rebranded. The label tells you nothing. This guide is how to tell them apart when you are the one signing.
The four types of supplier
| Type | Typical strength | Typical weakness | Fits SMEs? |
|---|---|---|---|
| No-code automation shop | Fast, cheap, good at connecting SaaS tools | Breaks at real complexity, thin on evaluation and monitoring | For simple workflows only |
| AI product reseller | Quick to deploy a known platform | Fits your process to the product, licence costs compound | Sometimes, if the product genuinely fits |
| Enterprise consultancy | Method, governance, scale | Priced and staffed for organisations with in-house data teams | Rarely |
| Independent AI firm | Diagnoses, builds and operates with senior people | Capacity is limited, so availability varies | Usually the best fit |
The distinction that matters is not size. It is whether the firm does the diagnosis and the engineering. Suppliers who only build will happily build the wrong thing. Suppliers who only advise leave you holding a roadmap nobody can execute.
Signals of a strong firm
- They insist on reviewing the process before quoting the build.
- They talk about the work, not the model. Model choice is an implementation detail that changes twice a year.
- They can name the systems they have integrated with, including the awkward ones.
- They have an evaluation method: a test set of real cases the agent is scored against before launch.
- They discuss monitoring, drift and what happens when the agent is wrong, unprompted.
- They will tell you a use case is not worth doing.
- They train your team as part of delivery rather than selling it as an upsell.
Signals to walk away from
- A fixed price for an unspecified build, quoted on the first call.
- Demos on their data rather than a sample of yours.
- No answer on where your data goes or which providers process it.
- Ownership of prompts, code or the automation platform account stays with them.
- Headcount reduction used as the primary business case.
- A senior team in the pitch and a junior team in delivery. Ask who does the work and write the names into the contract.
- Claims of full autonomy with no human in the loop for a process with legal, financial or safety consequences.
Pricing models, and what each one hides
| Model | How it works | Watch for |
|---|---|---|
| Fixed-scope project | Agreed deliverable, agreed price | Scope written vaguely so change requests carry the margin |
| Monthly retainer | Ongoing build and operate capacity | No definition of what a month buys, so output drifts |
| Per-workflow | Priced per automation deployed | Incentive to ship many small workflows over one valuable one |
| Outcome-based | Fee linked to a measured result | Attribution arguments, and the measurement usually favours the supplier |
Whatever the model, ask for the running cost separately: model inference, telephony minutes for voice, hosting, monitoring and the supplier's maintenance fee. Build cost is a one-off you can plan for. Running cost is what makes a workflow uneconomic in month nine.
What to insist on in the contract
- Ownership of code, prompts, evaluation sets and documentation, in writing.
- Accounts and infrastructure in your name where practical, with the supplier granted access rather than holding it.
- A named delivery lead and a change process if that person leaves.
- Data processing terms naming every sub-processor and where data is held. UK and EU clients should confirm transfer arrangements explicitly.
- An exit clause with a defined handover package and a maximum notice period.
- Acceptance criteria tied to evaluation scores on real cases, not a demo.
A sensible way to run the selection
Shortlist three. Give each the same one-page brief describing a single real workflow, with volumes and the systems involved. Ask for a written response covering how they would approach it, what they would need from you, what could go wrong and a cost range. You are not buying the answer. You are buying the quality of thinking, and a paid discovery is a reasonable next step with whoever thinks best.
Then check one reference in a business of similar size, and ask that reference a single question: what is still running today that they built for you.
Frequently asked questions
- How much does an AI automation agency cost for a small business?
- For a business of 10 to 250 people, a scoped discovery typically runs in the low five figures, a first production workflow commonly lands in the mid five figures depending on integration depth, and ongoing operation is usually a monthly fee plus usage. Treat any quote given before a process review as a placeholder.
- Is an AI automation agency worth it versus using no-code tools ourselves?
- No-code tools are genuinely good for simple, low-risk connections between SaaS products. Bring in a firm when the workflow touches a system without a clean API, when being wrong has a cost, or when it needs monitoring and evaluation to stay reliable.
- Should we hire an AI engineer instead of using an agency?
- Below roughly 250 people it is hard to hire and retain one, because a single engineer has no peer review and no cover. A common pattern is to use a firm for the first year, have them train an internal owner, then take delivery in-house once there is enough work to justify the role.
- How do I know if an AI agency is any good before signing?
- Give three suppliers the same one-page brief on a real workflow and compare their written responses. Look for whether they ask about volumes, exceptions and escalation, whether they name integration risks, and whether they are willing to say part of it should not be built.
- Who owns the AI agents an agency builds for us?
- You should, and it needs to be in the contract: code, prompts, evaluation sets, documentation and the accounts the system runs on. If the supplier retains the automation platform account or the prompt library, switching later means rebuilding.
Work through this with us
brep runs AI operating reviews, workforce training and agent deployment for businesses of 10 to 250 people in the UK and US. One diagnosis, one workflow, someone accountable for keeping it working.