How to choose an AI agency: the checks that hold up

This guide does not rank anyone, and it does not rank us. A ranking tells you what its author valued. What you need is a way to tell, on your own project, which agency will put something into production that works and that you can keep. Everything below can be checked before you sign, and we expect to be checked the same way.

Why a ranking cannot choose for you

A ranking sorts agencies on criteria someone else picked. Your project has its own task, data, systems and constraints, and only evidence on those tells you who fits.

There is also a labeling problem. In a June 2025 press release, Gartner described “agent washing”, the rebranding of assistants, robotic process automation and chatbots “without substantial agentic capabilities”, and estimated that only about 130 of the thousands of agentic AI vendors are real. The same release predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, “due to escalating costs, unclear business value or inadequate risk controls”. A good selection process tests for those three reasons.

How to choose an AI agency: ten checks before signing

Each check below asks for something you can see, read or try. A serious agency can show it within days.

Use the same table for every agency you shortlist. Two or three candidates tested properly will tell you more than a long list.

Reviews and references help with checks 1 and 2, if you read them closely. Clutch, for example, states that its pricing data comes from first-party reviews submitted by verified clients. Look at the project size, the date and what was actually delivered. Then ask the agency for a reference whose project looks like yours, and ask that reference one question: is the system still running today?

Test it on your own work before you commit

The strongest evidence is a result on your own cases. A demo shows what the agency can build in general, and a test on your data shows what it will deliver for you.

Pick a sample of real past cases: emails already answered, invoices already entered, questions already resolved. Include the awkward cases: the scanned invoice, the angry email, the question with no answer in the documents. Keep the expected answers on your side so the result is not tuned to them. Agree on the pass mark before the test, such as the share of cases handled correctly and the errors you cannot accept, then let the agency run it. Every shortlisted agency gets the same sample and the same mark.

If the agency proposes a paid pilot, ask how it ends. A pilot should state its success criteria, its duration, and what happens next if it passes or fails. A pilot with no route to production is a demo you paid for. You can try our own knowledge assistant demo to see what a test on a document set looks like. To pick the task to test first, our free AI readiness check gives a first score in about two minutes, and an AI readiness assessment ranks the candidates in depth.

What the data contract must say

If the agency processes personal data for you, the GDPR requires a written contract, and its minimum content is set by law. Ask for it early; the answer tells you a lot.

Article 28 of the GDPR, whose full text the CNIL, France’s data protection authority, publishes online, first requires you to use only processors that provide sufficient guarantees of appropriate technical and organizational measures. Its paragraph 3 then lists what the contract must include: the subject matter and duration of the processing, its nature and purpose, the type of personal data and categories of people concerned, and your obligations and rights. It must also bind the agency to act only on your documented instructions, keep the data confidential, secure it, use sub-processors only with your prior written authorization, help you answer people’s requests, delete or return the data at the end, and allow audits.

Beyond the contract, ask three practical questions:

• Where is the data hosted, and can it stay in the EU or on your own servers?
• Which model provider sees it, and do that provider’s business terms exclude training on your data?
• What is logged, for how long, and who can read the logs?

If your data must stay in Europe, our private LLM service is built for that case.

The EU AI Act dates your agency should know

An agency that builds AI for European companies should know the AI Act timeline without looking it up. The European Commission’s AI Act page, last updated on August 3, 2026, gives these dates.

These dates have moved once already. The Commission states that the simplification proposal known as the “AI Omnibus” was adopted on November 19, 2025, reached political agreement on May 7, 2026 and entered into force on July 27, 2026. Ask the agency how it tracks such changes. Our EU AI Act guide goes further, and our guide to AI governance covers who decides, documents and checks once a system is live.

Who owns what when the project ends

Ownership decides whether you can change provider, bring the work in-house or simply keep running what you paid for. Settle it in the contract, before work starts.

List the pieces: source code written for you, prompts and configuration, test sets and their results, documentation, cloud accounts, model API keys, and the data the system learned from or stored. Each should be yours, or its terms of use should be written down. When accounts and API keys are in your company’s name, leaving is a matter of handing over the documentation.

Ask for the running cost at the same time. Tekrevol, one of the agencies that publishes price grids, notes that most disputes trace back to three things nobody wrote down: who pays for API usage during the build, who owns the test data, and what “done” means. Our AI agent cost guide lists the published monthly ranges.

Questions to ask every agency, us included

Ask the same questions to every candidate, and ask them to us too. The good answers are specific, and they come with something you can check.

1. Can we try it on our own cases before we sign? A good answer names the sample size, the pass mark and the date.
2. What will you refuse to build for us? An agency that refuses nothing has not understood the risks.
3. What happens if the agent misses the pass mark? Look for a consequence written into the contract.
4. Show us a project you stopped, and why. Honest agencies have one and can explain it.
5. Who does the work, and do you subcontract any of it? You want roles, locations and the subcontractors listed in the contract.
6. Where exactly does our data go? Expect a region, a list of sub-processors and a retention period.
7. What does it cost per month at our volume, in year two? Expect a figure with its assumptions.
8. What do we own at the end, and how do we leave? Expect a list, and a handover procedure.

For an agent that acts inside your tools, our page on AI agents for business shows how we handle testing, access and ownership on our own projects.

You are welcome to put us through the same test. Book your free 30-minute assessment and bring one task with a few real cases: we answer the eight questions above for your project, say what we would refuse, and tell you what a test on your data would look like. We reply within one business day, and the scope and price of any project are fixed in writing before we start.