To choose your AI agency in 2026, five criteria decide it: does it execute with you or merely advise you? Does it show quantified proof (ROI, client cases) and a clear price, or a "contact us" instead? Is it certified by a serious model provider, such as an Anthropic partner? Does it stay at your side after delivery, or vanish once the invoice is paid? And does the code it produces truly belong to you? A good agency ticks these five boxes. The others sell slides.
In 2026, everyone is an "AI agency". Yesterday's webmaster, the consulting firm that added three pages to its website, the IT services company that rebranded its offer: the same label covers radically different realities. For an SME leader who finally wants to move to AI, the problem is no longer finding a provider — it's telling apart the one who will deliver a result from the one who will deliver a presentation.
This guide gives the concrete criteria to make that call. Not generalities, but the precise questions to ask, the signals that reassure, the red flags that should make you run. Enough to walk into any meeting — including one with us — knowing exactly what you should demand.
Why has choosing an AI agency become a headache?
Because the barrier to talking has collapsed, while the barrier to executing has stayed high. Anyone today can talk about agents, RAG, multi-agent systems and LLMs fluently enough to impress in a meeting. Far fewer people know how to install an AI system that holds up in production, integrates with a real information system, and creates measurable value six months later.
That's where the real risk sits. Most AI projects never reach their promise — not for lack of technology, but for lack of a deep enough understanding of the business and its information flows. A nice demo proves nothing: it runs on a hand-picked case, under ideal conditions. What matters is what survives contact with reality. Your job, when choosing an agency, is to spot who can do that last part.
In practice, three families hide under the same label. The web agency or IT services company that added "AI" to its offer without changing trade: it knows how to deliver a website or software, far less a system that reasons and improves. The consulting firm that produces AI strategy in slide decks, brilliant on paper but with no hands to execute. And the tool reseller, who plugs in a third-party subscription and bills the integration — until the day the tool changes its rules or its price. None of these families is illegitimate, but none truly does the job of installing, custom-built and over time, an AI that creates value for you. That last category is precisely the one to learn to recognize.
Advice or execution: the question that decides everything
This is the first filter, and by far the most discriminating. Ask yourself one thing: at the end of the engagement, do I walk away with a document, or with something that works?
A good share of the market sells consulting dressed up as AI: an audit that ends in recommendations, a "transformation plan", a workshop — then execution falls back on you and on teams that have neither the time nor the expertise. The deliverable is smart, but nothing has changed in the company. That's the classic consulting model, repainted in AI colours.
At Codito Ergo Sum, the positioning is the exact opposite: we don't explain what you should do, we do it with you. The monthly AI Partnership embeds real habits month after month, alongside the teams; Custom Development ships agents and tools that actually run. Advice is only the first step — never the final product. When you evaluate an agency, look for that shift: does it talk about what it will tell you, or about what it will build?
A simple test to run in the meeting: ask "who, on your side, will write the code and set up the agents?". If the answer stays vague — "our partners", "it depends on the project" — you're dealing with a layer of consulting that will subcontract execution, with the margin and loss of information that implies. If the agency names its people, shows code and describes precisely how it delivers, you're holding someone who actually executes.
What objective criteria should you use to evaluate an AI agency?
Beyond the impression in a meeting, four signals can be checked factually.
- Certification. Being an official partner of a model provider — Anthropic, for example — is not a decorative logo: it's a mark of access, technical seriousness and commitment. Codito Ergo Sum is a certified Anthropic (Claude) partner. Always ask whether the agency has one, and which.
- Quantified proof. A serious agency shows results in euros and hours, not vague "successful projects". On two concrete cases, our audits documented €527,000 of hidden costs per year at VLT Notaires and €109,000 at the Amourdedieu group — more than €636,000 in total — with a return on investment materializing in about two months. That's the level of precision to demand from everyone.
- A named method. Improvising doesn't scale. A real agency has a repeatable method — for us, the Information Trinity, the Triple ROI and the AI Maturity Score — that explains how it decides what to automate first, rather than sprinkling AI at random.
- Code ownership. What the agency builds for you must belong to you. An agent or software whose code and access you don't hold is a disguised dependency. Custom that truly belongs to you is an asset; the same tool locked at the provider's is a leash.
To dig into who runs an agency and how it works, its "about" page speaks volumes: see ours, the Agency page, and above all its case studies — client cases beat any promise.
How much should an AI agency cost — and how do you read a price?
The first sign of seriousness is a published price. An agency that replaces its rates with "contact us" asks you to enter a negotiation before you even know whether you're in the right ballpark. That's a bad start.
Our real ranges, displayed in plain sight: a strategic AI Audit from €2,000 excl. VAT, a monthly partnership at €399 (Essential) or €699 (Premium) excl. VAT with no commitment, and custom development from €3,000 excl. VAT on quote. We laid out the full reasoning, line by line, in our article "How much does AI really cost an SME in 2026?". The gist: the right budget isn't the biggest one, it's the one that tackles first whatever pays for itself fastest. An agency that pushes you straight to the most expensive build, with no prior audit, is selling its catalogue — not your interest.
What red flags should you absolutely avoid?
Five signals should trigger caution, however charming the salesperson across the table.
- Opaque pricing. No starting rate, everything "custom" from the first word: often a sign the price is tuned to what you seem able to pay.
- No proof. Vague testimonials, no numbers, no client to talk to. An agency that truly delivers has results to show.
- The one-shot. An engagement that ends at delivery, with no follow-up. Models change several times a year; an AI system without maintenance goes stale. Prefer a partner that stays — that's the whole point of long-term support versus a one-off consultant.
- Lock-in. A tool whose code, data or access you can't retrieve. You become captive.
- Jargon without ROI. Lots of buzzwords, not a sentence on what it saves you. If the agency doesn't tie its tech to a euro or an hour saved, be wary.
What questions should you ask before signing?
Here's the list to take into the meeting. The best agencies answer without dodging; the others get flustered.
- Show me a quantified client case: what problem, what did you deliver, what measured gain?
- Who actually writes the code — and does the final code belong to me 100%?
- What's your published starting price for a first scope?
- What happens after delivery: do you stay, or do you leave?
- What's your method for deciding what to automate first?
- Are you certified by, or a partner of, a model provider?
- How do you measure ROI, and over what horizon?
- What are the recurring costs once the agent is in production?
- Can I speak to one of your current clients?
- If I stop, am I locked into a fixed term?
None of these questions is a trap; they're simply the ones an agency that truly executes is happy to answer, because the answers play in its favour.
Generalist or specialized AI agency?
Both make sense — the question is how specific your sector is. A generalist agency that has mastered the method of integrating AI serves most SMEs very well, across all sectors: the frictions of a firm, a services group or a distributor resemble each other more than you'd think, and it's the method that makes the difference, not the sector veneer.
Some trades, however, have codes strong enough to justify a specialization. That's why we created a dedicated sector subsidiary, Codito BTP, for construction and public works. The right instinct: choose an agency that has mastered the AI method first, then, if your sector demands it, check that it knows the codes — in that order. Sector expertise without an execution method remains just talk.
In summary
Choosing your AI agency in 2026 comes down to a simple conviction: prefer the one that builds to the one that comments. Check that it executes with you, that it shows quantified proof and a price, that it's certified, that it stays after delivery and that the code belongs to you. Flee opaque pricing, the one-shot and jargon without ROI. And take the list of questions along: it sorts the field in ten minutes.
If you want to apply these criteria to your situation rather than to generalities, that's exactly what the first meeting is for.