You asked three providers for a quote on an AI agent and the three proposals look alike. The same models are named, the same demos work, and prices vary by a factor of three. What separates a provider who can deliver is not visible in the sales deck. It shows in the questions they ask, in what they refuse to promise, and in the way they talk about failure. Here are twelve signals you can observe before signing, ordered by stage of the project.
Key Takeaways
- A provider who announces a gain in percentage before looking at your data is not estimating, they are selling.
- A demo that never shows a failure case is not a demo, it is an advert.
- A quote that does not separate build from running costs sets up a surprise invoice.
- Missing written acceptance criteria is the most expensive signal: you will have no way to recognise delivery.
- Three signals are enough to rule a candidate out. The twelve help you compare the ones who remain.
The market grew faster than the skills
In 2025, 20.0% of EU enterprises with 10 or more employees used artificial intelligence, up from 13.5% in 2024, according to Eurostat (2025). A 6.5 percentage point rise in one year produced more providers than it trained practitioners.
The breakdown of uses is even more telling. The most common use remains analysing written language, reported by 11.8% of enterprises, ahead of generating images, sound or video (9.5%) and generating written or spoken language (8.8%). Put plainly, much of what is sold today as an autonomous agent is text processing wired to a set of files. That is not a flaw in itself. It becomes one when the pitch suggests something else.
A serious provider does not hide that gap. They state what the system does, on which data, with what error rate, and what it will not do. The twelve signals below are gaps between that account and the actual project.
Three signals in the first conversations
1. The answer to a technical question is a tool name
You ask how the agent will know that it does not know. The answer is "we use the best model on the market". The model is a component, not an answer. A provider who has shipped before talks about document retrieval, enforced output format, guardrails and logging. If they can only describe the model, they have not yet put an agent into production.
2. The gain is announced before the audit
"40% less time on invoice processing." The number lands in the first meeting, before a single one of your files has been opened. Real gains depend on your volume, the quality of your documents and the error rate you are prepared to accept. Announcing a percentage before measuring means selling a market average as if it were your project.
3. Nobody asks to see your data
An agent project starts with a sample: ten invoices, fifty emails, one ticket export. If nobody asks for that sample before quoting, the scope is invented. Silence on the legal side belongs to the same family. Since 2 August 2026, the European framework on AI (European Commission, 2026) has made transparency obligations applicable, including telling a person that they are interacting with a machine. A provider who never raises this point leaves you to handle it alone, after go-live.
Three signals in the demo
4. The demo contains no failure case
A well-prepared demo shows the agent giving correct answers. Ask what happens with an unreadable document, a request outside scope, two contradictory instructions. An agent with no fallback behaviour, one that cannot say it does not know, hand over to a human or ask for a clarification, is still a prototype. The fallback is not a detail: it is the part of the system that protects you in production.
5. The provider cannot measure quality
A simple question: how will you know the agent works? If the answer is "we will look at the results", there is no method. A real project defines a test set before building, an acceptance threshold and a tolerated error rate. Without that frame, sign-off comes down to impression, and the conversation about quality has nothing to stand on.
6. The agent does everything, on every one of your tools
A first project covering customer relations, accounting and HR in the same batch is sales, not design. Every connected tool adds an interface to maintain and one more failure mode. A provider who refuses to split the scope is doing you a favour: they force you to pick the first use case that pays.
Three signals in the quote and the contract
7. The quote does not separate build from running costs
An agent costs money to build and money to run: model calls, hosting, monitoring, fixes. A quote showing only a build price leaves the second line outside the frame, and that is not necessarily an error. Ask for a monthly running estimate, even a rough one. If it cannot be produced, the provider has never operated what they sell.
8. No acceptance criteria are written
"Delivery of a request-handling agent." That sentence gives you no way to recognise delivery. A serious quote describes testable scenarios: out of twenty standard requests, the agent handles fifteen correctly, passes five on and leaves a trace. That level of detail is not bureaucracy. It is what protects you from the final disagreement, when one side talks about features and the other about results.
9. Ownership and exit are not written
Who owns the prompts, the code, the connectors? What happens if you change provider, model or hosting company? A contract silent on those two questions ties you to the provider for the entire life of the agent. Reversibility is cheap to write at the start. It becomes impossible to negotiate three years later.
Three signals after delivery
10. No maintenance, no handover
An agent that works today can degrade in three months: a vendor changes an interface, your documents evolve, the model is updated. A provider who mentions neither monitoring nor knowledge transfer is selling a photograph, not a system. Ask who reads the error logs in week 12.
11. The technical contact changes at every exchange
Sales in pre-sales, a project manager next, an unknown developer at delivery. If you cannot name the person who will write the code and talk to them before signing, you are buying a promise. This is not about company size: a solo independent can be an excellent choice for a narrow scope, provided they are the one who builds.
12. No record of decisions
Three months after go-live, nobody remembers why the agent answers the way it does. A serious project leaves short documentation: flow diagram, business rules, prompt versions, rollback procedure. Without it, the smallest fix turns into billable archaeology.
What the sales promise leaves out
| What you are told | What it often hides | The reflex |
|---|---|---|
| "Our solution adapts to anything" | No defined scope | Ask for a trial on your own documents |
| "The AI learns on its own" | No test set, no measurement | Ask for the acceptance criteria |
| "We use the most advanced model" | No work on data and guardrails | Ask for the full flow diagram |
| "One all-inclusive fixed price" | Running costs and maintenance out of scope | Ask for the monthly running cost |
| "You will be autonomous" | No handover in the contract | Ask who maintains it in six months |
The thirty-minute test
One meeting is enough to sort candidates if you ask each of them the same five questions. Ask for an example of a project they stopped, and why. Ask what happens when the model is wrong. Ask for the monthly running cost. Ask to see an extract of documentation delivered to a client. Ask for the name of the person who will build it.
The answers matter less than the speed at which they come. A provider who has shipped answers in concrete sentences, with figures on their side and limits named. A provider selling a pitch returns to the model, the power of AI and the number of possible use cases. To prepare the comparison, the questions to ask before signing complete this test, and choosing between an agency, a freelancer and an IT services firm helps you frame the shape of the engagement.
FAQ
How many signals are enough to rule out a provider?
One, if it touches scope or contract: no acceptance criteria, no ownership clause, no running cost. Those three points determine how the rest of the project goes. Signals of posture, such as a tool name offered as a technical answer, are open to discussion: they point to a lack of experience, not necessarily bad intent.
Can an independent provider be serious on this kind of project?
Yes, and often more so than a firm that subcontracts. What counts is continuity: the person who designs talks to the person who builds, and the same person answers when something breaks. A serious independent documents their work and plans their exit, precisely because they know they may disappear.
Should you insist on a paid test phase before committing?
It is often the healthiest formula for a first project. A short, paid pilot on a single use case with a real dataset lets you measure quality before funding the rest. A provider who refuses any paid scoping exercise is selling a commitment larger than what they can demonstrate.
What if the provider you already chose shows these signals?
Handle the two most expensive points first: acceptance criteria and the exit clause. They can be added to a running contract as an amendment. Then ask for a technical session with the person who actually builds. If that session is refused, you have your answer.
Conclusion
An unreliable AI provider does not give themselves away through price, size or the model they use. They give themselves away through what they avoid: your data, your acceptance criteria, the running cost, the person who will build. Those four subjects can be checked in a single meeting, before any financial commitment.
If you want to move from sorting providers to a scoped project, an audit of use cases helps you pick the first workstream and set the success criteria before you talk to anyone. To see what AI agents look like inside a small company, the use cases give concrete examples, and our AI agent offers describe the typical scope of a project. The simplest next step is a conversation: contact us with two or three processes that cost you time, and we will tell you which ones hold up.
