HowtoChooseanAIAgencyin2026:VerifiableCriteria,NotPromises
A practical guide to evaluating any AI agency, ours included: what to check before signing, what to ask in the RFP, which signals should make you walk away, and when you simply don't need an agency. Every criterion in this guide can be verified without taking anyone's word for it.
How do you choose an AI agency?
To choose an AI agency, evaluate six verifiable criteria: (1) systems in production with metrics, not demos; (2) a team with names and checkable credentials; (3) public prices or clear ranges by project type; (4) partnerships and a stack the agency can justify; (5) EU AI Act and GDPR compliance by design; and (6) dogfooding, meaning the agency uses AI in its own operations. If a vendor fails three or more, keep looking.
Who Signs This Guide
An AI agency writes it: an interested party. Hence, everything verifiable.
Over 15 years delivering digital projects and, since 2023, applied AI.
More than 30 projects with language models since 2023, with 15+ agents in production.
Ranges published on every service page. Almost nobody in the industry does this.
With names and verifiable profiles, across Spain and Argentina.
The 6 Verifiable Criteria for Evaluating an AI Agency
None of them depends on believing a sales pitch. All of them can be checked in an hour.
Systems in production, not demos
Ask for 2-3 systems running today, with a metric (hours saved, resolution rate, revenue) and a referenceable client. A demo takes an afternoon; a system holding up real users for months is another league. If the entire portfolio is pilots and proofs of concept, there is no portfolio.
A team with names and credentials
Who will actually work on your project, with name, role and a checkable profile (LinkedIn, GitHub). Watch out for the classic consultancy pattern: the star team sells, the junior team executes. Ask about each profile's real dedication and insist on meeting the people running the day-to-day before signing.
Public prices or clear ranges
Anyone who has quoted dozens of projects can give ranges by project type without blinking. "It depends on many factors" as the only answer signals inexperience or made-to-measure pricing. Publishing prices is still extremely rare in this industry; at minimum, demand indicative ranges in writing.
Justified partnerships and stack
Partner badges (cloud, platforms) are easy to verify in official directories and filter out newcomers. But beware the opposite effect: an agency married to a single vendor will always prescribe its own hammer. The best model changes every few months; the stack should be justified by your case, not by their reseller agreement.
EU AI Act and GDPR compliance by design
The EU AI Act is in force and its obligations arrive in phases; the AI literacy duty of Article 4 already applies. Ask how they classify the risk of your use case, where your data is processed and which contracts (DPAs) they sign. If the answer is vague, the legal risk stays with you.
Dogfooding: they use AI in their own house
If an agency sells AI but its own operations (support, project management, marketing) are 100% manual, something doesn't add up. Ask them to show you how they use AI internally, with the same rigor you'd demand as a client. Whoever operates their own systems knows the real costs and failure modes, not the brochure ones.
The 12 Questions for Your RFP
Copy them verbatim into your request for proposals. Evasive answers are also an answer.
- Which AI systems do you have in production right now, and what metric backs each one?
- Who will work on my project? Names, roles and percentage of dedication.
- What price range do you handle for a project like mine, and which factors move it?
- How do you evaluate the quality of the model's answers before and after going to production (evals)?
- Where is my data processed, who has access to it, and under which contracts (DPA, EU region)?
- How do you classify my use case under the EU AI Act, and which obligations fall on me?
- Which part of the solution is your own development, and which part is a third-party product with a margin?
- Which business metric will the project move, and what happens if it hasn't moved after 3 months?
- When it's done, do I own the code, the prompts and the data? What does leaving cost?
- What monthly operating cost will the system have (tokens, hosting, maintenance)?
- Do you use AI in your own operations? Show me.
- Which project have you turned down recently, and why?
Red Flags When Choosing an AI Agency
None is conclusive on its own. Two or more together, yes.
Case studies without a single metric
Ask for concrete numbers: resolution rate, hours saved, cost per query, attributable revenue. If all you get is adjectives ("transformative", "innovative"), there is no case study: there is a press release.
A "proprietary AI platform" that is a wrapper
Ask what's underneath. A thin layer over the ChatGPT API charged as a proprietary product means overpricing plus lock-in. A wrapper can be legitimate, but only if it's transparent and priced as what it is.
Accuracy promises without evals
A "99% accuracy" without an evaluation set, methodology or test data is marketing, not engineering. Ask how they measured it, on which cases, and what happens with the remaining 1%. Anyone who evaluates seriously will gladly show you the process.
Everything is possible and everything is urgent
An experienced agency scopes, discards, and will tell you no to something in the first meeting. The one that says yes to everything will learn on your money. Ask what they would NOT do in your case: the answer is worth more than the proposal.
A fixed quote before seeing your data
In AI, the state of your data is the biggest cost factor. Whoever gives you a fixed price without looking at data and systems is either improvising or padding the margin to cover themselves. The serious approach: a range first, and a discovery kickoff that closes it.
Do You Need an AI Agency?
The honest answer is: not always. This is the line we use ourselves to say no.
Who it's for
- You have a process with measurable volume and cost where AI can save hours or generate revenue.
- Your data exists and is accessible, even if messy: there is raw material to work with.
- You have no internal AI team, and building one would take longer than the project itself.
- Your case touches compliance (EU AI Act, GDPR) and you don't want to learn it through scares.
- You need the system integrated with your CRM, your ERP or your real operations, not a toy on the side.
Who it's not for
- A ChatGPT, Claude or Copilot subscription for the team (€20-60/user/month) solves your case: start there.
- You already have a data or development team: you may only need training or a few weeks of consulting, not a turnkey build.
- You want AI because "you have to have AI": without a business case, not even the best agency in the world will save the project.
- Your budget is under €3,000: existing SaaS tools will give you more than any custom development.
- Nobody in your organization will own the system after delivery: without an internal owner, every project dies alone.
How to Apply This Guide to Us
Kiwop is an AI agency, so this guide is signed by an interested party. The right answer to that conflict isn't to fake neutrality: it's to hand you the means of verification. Our prices are published (starting with the AI audit and what each type of project costs), we operate Nexo, our own platform with AI in production, and we've used AI agents in our daily operations since 2023. Send us the RFP above, question by question: that's exactly what it was written for.
Frequently Asked Questions When Choosing an AI Agency
What companies ask us while they compare.
How much does an AI agency charge?
With indicative ranges from our experience and published prices: an AI audit costs €3,000-8,000, automating a process €3,000-15,000, an AI chatbot €8,000-30,000, an agent in production from €25,000 and an enterprise RAG from €45,000. Ongoing support starts at €1,500/month. You'll find the full breakdown, with the factors that move each range, on our page about how much an AI project costs.
How do I check that an agency's case studies are real?
Three routes: ask to talk to the client behind the case (a 5-minute call is enough), ask for the specific metric and how it was measured, and ask to see the system running live with real data, not a recorded video. An agency with real cases arranges this within a week. Evasion ("confidentiality", "the client isn't available") on all three fronts at once is your answer.
Is a big agency better than a small studio?
Size matters less than the share of work actually in production and who really touches your project. A large consultancy can sell you the senior team and staff you with juniors; a three-person studio can give you the founder but fall short on coverage and continuity. The right question isn't how many they are, but who will work on yours and what those specific people have shipped to production.
What is a "wrapper" and why should I care?
A wrapper is a thin layer over a third-party model (ChatGPT, Claude) presented as a proprietary product. It isn't bad per se: sometimes a well-built layer solves your case at a good price. The problem is the opaque wrapper: charged as a "proprietary AI platform", with lock-in and a markup. The test is asking what's underneath and what happens if you want to leave. If the answer makes them uncomfortable, you have it.
Does the agency need to be an OpenAI, Microsoft or Google partner?
It helps, but it doesn't replace systems in production. Partnerships can be verified in official directories and filter out opportunists, but they reward reselling volume, not engineering quality. More valuable than a badge: an agency that is model-agnostic and can argue when to use GPT, Claude, Gemini or an open model. The best model changes every few months; your vendor shouldn't be married to any of them.
What role does the EU AI Act play in choosing an agency?
It's a fast, cheap filter: ask how they would classify the risk of your use case and which obligations fall on you. A serious agency answers without checking notes, because it already does this on every project; the Act's obligations are phasing in and the AI literacy duty already applies. If they tell you "that's your lawyer's job", you'll end up paying twice for the technical file the regulation requires.
When do I NOT need an AI agency?
When a €20-60/user/month subscription to ChatGPT, Claude or Copilot solves your case; when you already have a technical team and only need training or a few weeks of consulting; when there's no measurable business case behind it, just the feeling that "we should do something with AI"; and when there will be no internal owner of the system after delivery. In all those scenarios, a custom project is money thrown away, whoever sells it to you.
How do I start with the least possible risk?
With a bounded, cheap project that has a metric: an AI audit (€3,000-8,000, 2-3 weeks) if you don't know where to start, or an automation of a single process (€3,000-15,000) if the case is already clear. In 4-8 weeks you know whether the vendor delivers, with an investment that doesn't hurt. Distrust anyone who proposes the 12-month platform as a first date.
Apply the Guide to Us: Ask for the Cases, the Team and the Prices
These criteria work against any agency, ours included. The prices are already published; the cases and the team we'll show you in a 30-minute call.
Put us to the test Initial technical
consultation.
AI, security and performance. Diagnosis with phased proposal.
Your first meeting is with a Solutions Architect, not a salesperson.
Request diagnosis