AI agency, freelancer or in-house hire: how to choose?

For an SMB's first AI project, the short answer is: a senior freelancer on a flat fee, on a written scope. An agency becomes relevant when the project demands a real team (several developers, design, project management) in parallel. An in-house hire is only justified when AI becomes a core business, with work all year round to keep the role busy. The rest of this article serves to check that answer against your specific case, figures in hand.

In brief

• A senior AI freelancer bills between 600 and 1,000 € excl. tax per day in France (the Malt index puts the AI median around 750 €). On a flat fee, a delivered feature costs thousands of euros, not tens of thousands.
• An agency sells its days between 800 and 1,500 € excl. tax, while the person actually coding your project is often paid two to three times less: the gap funds the sales team, the project manager and the offices.
• A salaried data scientist or AI engineer costs 55 to 80 k€ gross per year depending on experience and region, that is 75 to 110 k€ fully loaded, plus 3 to 6 months of recruitment.
• The criterion that eliminates 80 % of bad providers: insist on seeing something delivered in production, usable in front of you, not slides or POCs.
• A flat fee on a written scope protects you better than time-and-materials: the risk of overrun sits with the provider, not with you.

The 3 options, no spin

I am a freelancer, so judge and party on this subject. I say it up front, and I come back to it at the end of the article. Here, even so, are the three options as I see them after fifteen years on both sides of the fence, in an agency, in-house and independent.

The agency: capacity, at team prices

An agency sells between 800 and 1,500 € excl. tax per day. For an SMB AI project, rarely count on less than 30,000 € of entry budget. That price buys one real thing: capacity. Several profiles in parallel (dev, data, design, project manager), continuity if someone goes on holiday, a structure that survives a departure.

The risks are just as real. The first: who actually codes? The pre-sale is led by a brilliant senior profile, the execution is often handed to a junior supervised from afar. Ask the question head-on: "who will write the code, may I speak to them?". The second: turnover. Agency juniors stay 18 to 24 months; if your project drags on, you will change technical contact midway. The third: the temptation of time-and-materials, billing days rather than a result, which turns a 3-month project into a 12-month subscription.

The freelancer: senior, direct, with one point of fragility

A senior AI freelancer sits between 600 and 1,000 € excl. tax per day. The difference with an agency is not just the price: it is that the person who sells is the one who codes. No sales layer, no hidden junior, direct responsibility for the result. For a bounded scope (a chatbot on your documents, an automation, an MVP), it is the most efficient option at the most legible cost, especially on a flat fee.

The risks, honestly: the bus factor. A single person, so if they fall ill, get overloaded or disappear, the project stops. It can be mitigated (documented code, ownership with you, reversibility set in the contract), it cannot be eliminated. Then availability: good freelancers have a full order book, count a few weeks of lead time. Finally the capacity limit: an independent will never replace a team of five on a project that requires five.

In-house: the right choice when AI becomes the business

A salaried data scientist or AI engineer costs between 55 and 80 k€ gross per year depending on experience (a senior in Paris easily exceeds 65 k€), that is 75 to 110 k€ with payroll charges, workstation and tools. Add 3 to 6 months of recruitment in a market where good profiles choose their employer, and a real risk of a bad hire at 6 months of salary.

The most frequent trap is not the cost, it is under-utilization: an SMB that hires to "do AI" without a year's backlog ends up with an expensive profile maintaining a chatbot and getting bored. They leave after a year, and you start from scratch. Hire when AI is a durable competitive advantage of your business, not for a project. For a project, buy the project.

The provider selection checklist

Whatever the option, here are the seven points I would check if I were in your place. Each is verifiable in under an hour.

• 1. Something delivered in production, demonstrable. Not slides, not a prepared demo: a product or feature live, that you can use during the interview. An AI provider with nothing to show in production has never faced the real problems (costs that drift, hallucinations, scaling up).
• 2. Flat fee rather than time-and-materials for a known scope. A firm price on a defined deliverable puts the risk of overrun with the provider. Time-and-materials (billing by time) is only legitimate for assumed exploratory work, and then with a cap.
• 3. Who operates it afterwards? An AI system lives: models that evolve, API costs, content to update. Ask what happens in month 2: who monitors, who fixes, at what price? A vague answer here signals a vague bill later.
• 4. Ownership of the code and the data. The delivered code belongs to you (written assignment), it is hosted on your Git repository, the cloud accounts are in your name. If the provider hosts everything on their side, you rent, you do not buy.
• 5. Reversibility. Ask the question: "if we stop working together tomorrow, what happens?". The good answer fits in one sentence: you keep the code, the documentation and the access, and another developer can take over. If the answer runs past three sentences, beware.
• 6. Verifiable references. Not logos on a website: a name and an email of a client you can contact. Two references that answer are worth more than twenty logos.
• 7. A written scope before signing. What is included, what is not, the acceptance criteria, the deadlines. A serious provider writes it on their own: it is also their protection.

The classic traps

The endless POC. A proof of concept at 15,000 €, then a second, then a "pilot"... and never a production deployment. The POC is sometimes useful, but it must have a written exit condition: "if X works, we deploy". Otherwise it is a loss leader billed on a loop.

Time billing on a vague scope. The losing combination: no one knows when it is done, and every scoping meeting is billed. If the provider cannot quote a flat fee, it is often because they cannot yet do it, and will learn at your expense.

The agency's proprietary platform. Some agencies deliver on "their" in-house AI platform. The day you want to leave, there is nothing to take over: the code is not yours, nor is the infrastructure. It is an annuity disguised as a service.

AI washing. Some IT service firms and web agencies repainted their brochure as an "AI agency" in 2023 without changing the teams. The point 1 test (something delivered in production, usable in front of you) eliminates them in ten minutes.

My positioning, transparently

I am a freelancer: this article mechanically argues for my own corner, so let me neutralize it by telling you when I am not the right choice. If your project demands a whole team in parallel (a complete overhaul with design, mobile and back office in 4 months), take an agency. If it is a pure data science project (predictive models on your data, research), take a data scientist, that is not my trade. If you need an on-site presence several days a week, I work remotely, it will not be me.

My model, precisely because the "custom AI project" market is murky: I do not sell days, I sell features already proven in production on beforbuild.com, my B2B SaaS, which I customize to your context on a flat fee: from 440 € excl. tax for a feature, 6 600 € excl. tax for a complete MVP (details in how much does a SaaS MVP cost). The scope is written and signed before starting, the code belongs to you, the stack is standard and can be taken over by any developer (Supabase, Cloudflare Workers). Apply the seven-point checklist above to me: it is built for that.

Where to start

1. Write the problem on one page. Not the solution, the problem: "my salespeople spend 4 h a week looking for information in our documents". With an order of magnitude of the current cost. Without that page, any provider will sell you what they know how to do.

2. Ask for two or three flat-fee quotes on that page, from different profiles (a freelancer, an agency). Comparing the answers will teach you more than any article, including this one. Run each candidate through the seven-point checklist.

3. Start small and measurable. One scope, one deliverable in production, one quantified success criterion, in 4 to 8 weeks. You validate the provider AND the use case before committing more. If the first proposed step exceeds 25,000 €, ask why there is no smaller step.

Sources

• Malt · Freelance rate index by profession (daily rate, continuously updated)
• Hellowork · Data scientist salary in France