Guide10 min readOctober 2026
How to choose an AI agency: a checklist for companies
Six questions for every candidate, three red flags, a table to score them and what has to be in writing before you sign.
To choose an AI agency, ask every candidate the same six questions: who owns the code and the agent, where your data is processed, how the price is set, what happens after go-live, what they run in production and what their agent does when it is unsure. Pick the one that shows you a working demo on your own data and gives you a fixed price before you sign the build.
Part 01
Why it matters which AI agency you choose
Most companies that consider AI have nobody in-house who can build it. Eurostat asked EU enterprises that had considered using AI but did not why they held back. 70.89% named a lack of relevant expertise, far ahead of any other reason (Eurostat, 2025 data).
The gap is widest in small firms. In 2025, 17% of enterprises with 10 to 49 employees used AI, against 55.03% of those with 250 or more. Across all enterprises the EU figure was 19.95% (Eurostat, isoc_eb_ai). For an accounting firm in Lyon or a property manager in Rotterdam, the first AI agent is almost always built by someone outside.
Hiring outside help is fine. The trouble is the noise in the market. In June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. In the same release it estimated that only about 130 of the thousands of agentic AI vendors are real. It calls the rest "agent washing": existing assistants, RPA and chatbots rebranded without real agentic capability (Gartner, June 2025).
Agentic AI projects predicted to be cancelled by the end of 2027. Gartner, 2025
Real agentic AI vendors among the thousands selling it. Gartner, 2025
Enterprises that will abandon agentic AI built by vendor engineers by 2028. Gartner, 2026
The third number is the most useful one for a buyer. In September 2026 Gartner predicted that by 2028, 70% of enterprises will abandon agentic AI built by a vendor's forward-deployed engineers, trapped by soaring costs and unable to evolve it on their own. Its advice is to put deliverables, knowledge transfer, intellectual property rights and exit responsibilities into the contract (Gartner, September 2026). The questions in the next part go straight at that.
Part 02
What to ask an AI agency: six questions
Ask all six to every candidate, in the same order, and write down the answers word for word. A serious agency answers each in two sentences. If the answer is "it depends" and the "on what" never comes, that tells you something too.
- Who owns the code and the agent?Good answer: the repository lives in your account. The prompts, evaluation set, configuration and documentation are yours. The model provider and hosting accounts are in your company's name, and the invoice comes to you.
- Where is my data processed?Good answer: a named region, the list of sub-processors (model provider, cloud, database) and a data processing agreement. Better still if there are options: hosted by the agency in the EU, in your cloud or on your own servers.
- How is the price set?Good answer: a fixed price for a written scope, agreed after seeing your process and a sample of your data. Separately, a calculation of the monthly model usage and hosting cost, which you will pay.
- What happens after go-live?Good answer: clear options with a known price. Full handover to your team, support for a fixed fee, or an ongoing partnership. Who watches for errors, who updates the system when the provider retires a model version, and how you leave if you want to.
- What do you have running in production?Good answer: systems that real people use today, at real volume, and a client you can talk to. A pilot, a prototype or a video does not count.
- What does your agent do when it is unsure?Good answer: a confidence threshold, a human review queue for whatever falls below it, a log of every decision, and an evaluation set built from your own real cases, measured before go-live.
The sixth question separates people who have put agents into production from people who have only built demos. An agent that sorts a property manager's tenant emails will get most of them right. The design shows in the other cases. The email that mixes a broken boiler with a complaint about the neighbours has to reach a person, not the wrong category.
For high-risk systems, Article 14 of the EU AI Act requires that people can oversee them effectively while they are in use. Even if your use case is not high-risk, the design in Figure 2 is the practical way to work in that spirit from day one.
Part 03
Red flags when hiring an AI agency
None of these signs proves bad faith. Each one moves risk from the agency to you. Three cost the most.
- An hourly estimate with no ceiling"We expect 300 to 500 hours" means you pay for the overrun. The builder is best placed to control the scope, so the builder should carry the risk of going over it. If they only work by the hour, ask at least for a written cap.
- Slides instead of a demo on your dataA presentation proves that someone can make presentations. A demo on sample data proves that the sample works. The only thing that predicts your result is watching the system work on your invoices, your emails or your pitch decks.
- Dependence on the agencyModel provider accounts in the agency's name, a proprietary platform with no export, prompts you cannot see, or a licence that lapses when you stop paying for support. Any one of these turns support into a toll.
A few smaller ones are worth noting:
- They promise an accuracy figure before seeing a single case of yours.
- They name a delivery date before they know where the data comes from.
- What they call an agent is a chatbot that answers and never acts in your systems.
- Nobody can say where the data is processed or which sub-processors are involved.
- There is no way to measure the result because nobody mentions a baseline.
Part 04
A table to score every agency
Copy this table into a spreadsheet. Score each criterion from 1 to 5 using the answers you wrote down, multiply by the weight and add up. The maximum is 50. Proof on your data counts three times because it is the only criterion that predicts how your project will go, and not someone else's.
| Criterion | Weight | 1 point if | 5 points if | Your score |
|---|---|---|---|---|
| Proof on your data | ×3 | Only slides or other clients' cases | A working demo on your data before the build | __ / 5 |
| Ownership and exit | ×2 | Agency's accounts and code, no exit plan | Everything in your name, documented handover | __ / 5 |
| Price certainty | ×2 | Hours with no ceiling | A fixed price for a written scope | __ / 5 |
| Data and hosting | ×1 | Nobody knows where it is processed | Region, sub-processors and agreement in writing | __ / 5 |
| Systems in production | ×1 | Only pilots and videos | Systems in use and a client who takes your call | __ / 5 |
| Handling uncertainty | ×1 | "The model does not make mistakes" | Threshold, human review, log and evaluation set | __ / 5 |
| Total | 40 to 50: next round. 30 to 39: ask in writing for what is missing. Below 30: drop. | __ / 50 | ||
Here is the table applied to three kinds of proposal that come up often. They are illustrative examples, not real agencies.
| Proposal | Proof ×3 | Ownership ×2 | Price ×2 | Data ×1 | Production ×1 | Uncertainty ×1 | Total |
|---|---|---|---|---|---|---|---|
| A. A 40-page deck and an hourly estimate | 1 | 3 | 1 | 4 | 4 | 2 | 21 |
| B. Demo on sample data, fixed price, proprietary platform | 3 | 1 | 5 | 3 | 3 | 3 | 30 |
| C. Demo on your data, fixed price, code in your name | 5 | 5 | 5 | 4 | 3 | 4 | 46 |
Proposal A has the better-known clients and still comes last. B looks cheap, but the fixed price buys a system you cannot take with you. C is the only one you can compare with your baseline before you commit the build budget.
A demo on your data is worth more than anyone else's success story. It is the only proof that speaks about your process.
Alpgency
Part 05
What to put in writing before you sign
Answers in a meeting are worth nothing if the contract says something else. Before you sign, check that it covers these five points.
| Point | What the contract should say |
|---|---|
| Scope and acceptance | What the system does, on which data, and the acceptance test: a set of your own real cases and the minimum accuracy it has to reach on them. |
| Ownership | That the code, prompts, evaluation set, configuration and documentation become yours, and whose name the accounts are in. |
| Data processing | A data processing agreement under Article 28 of the GDPR: only your documented instructions, no sub-processor without your written authorisation and, at the end, return or deletion of the personal data. |
| Running costs | Who pays for model usage and hosting, what to expect per month at your volume, and what happens if it spikes. |
| Exit | How the handover works, what documentation you receive, how many training sessions are included, and when access and data are handed over if you end the contract. |
A quick test
Ask: "If we stopped working together tomorrow, what would we receive and who could maintain the system?" If the answer takes more than a minute or needs a separate meeting, the exit plan does not exist yet.
Part 06
How we answer these questions
If you are comparing AI agencies, here are our answers to the six questions. Score them with the same table as everyone else.
- Price: the Audit is free and starts with a 30-minute call. At the end of it we agree one fixed price for the first use case. We never bill by the hour.
- Proof: before the build you see a working demo on your real data. Then we build to production, inside the same fixed price.
- After go-live you choose: full handover to your team, a fixed monthly support fee, or we stay on as your tech partner on a fixed schedule. With the handover, your team keeps the code, the prompts and the evaluation set.
- Data: GDPR-compliant deployments by default, hosted by us, in your cloud (AWS, GCP or Azure) or on your own servers.
- Production: a deal-flow scoring system for a European pre-seed fund, live in production 9 days after kickoff, in 2026. For Memorable, a session transcript becomes scores and a written report that their facilitators review before it reaches the client.
The roadmap and findings from the Audit are yours to keep, whether you build with us or not. If another agency scores higher on your table, hire them.
FAQ
Frequently asked questions
How do I choose an AI agency?
Ask every candidate the same six questions: who owns the code and the agent, where the data is processed, how the price is set, what happens after go-live, what they run in production and what their agent does when it is unsure. Ask for a demo on your own data before you sign the build, and score the answers with the weighted table in this article.
What are the red flags when hiring an AI agency?
The three most expensive are an hourly estimate with no ceiling, a slide deck instead of a demo on your data, and a setup that ties you to the agency: accounts in its name, a proprietary platform with no export, or no exit plan. Be wary, too, of anyone who promises a specific accuracy before seeing your data.
Should I pay an AI agency by the hour or a fixed price?
A fixed price for a defined scope moves the risk of overruns to the builder, who is best placed to control it. To be fair, the price should be agreed after the agency has seen your process and your data, and the contract should say what it includes and what happens if the scope changes.
Who owns the AI agent when the project ends?
Whoever the contract says, so write it down. The sensible default is that the code, prompts, evaluation set, configuration and documentation become yours, and that the model provider and hosting accounts are in your company's name.
Does the agency have to process data inside the EU?
The GDPR does not ban processing outside the EU, but international transfers have to meet its conditions, and a processor may only act on your documented instructions. Ask in writing where each part of the system runs, which sub-processors are involved, and whether there is an option in the EU, in your cloud or on your own servers.
Test us on your own data.
In the free Audit we map your process, check whether AI is the right answer and put a fixed price on the first use case before you commit. It starts with a 30-minute call.
Book a free AuditKeep reading
AI consulting for companies: where to start with AI
How to choose your first AI use case, score it with a simple weighted matrix, decide whether to build or buy, and what a good AI Audit should hand you at the end.
AI agents for business: what they are, what they do and what they cost
What an AI agent does inside the tools your company already uses, what sets its price, how far European companies have got with AI, and the questions to ask a provider before you sign.
The EU AI Act for companies: what changes and what to do
What applies to a company that uses AI today, which dates the Digital Omnibus moved, and a checklist you can run this month.