You have a sound AI project, a budget signed off, and one question blocking everything else: who is going to do the work? Hiring a data scientist or an AI project lead takes months for a profile that large companies compete for. Going through a consultancy costs more per day and leaves the skill outside your walls. Both answers are defensible. The bad one is to choose before you have framed the real problem, which is almost never a hiring problem.
Key Takeaways
- An in-house AI profile costs between €35k and €60k gross per year according to the Apec, before payroll charges, tooling and onboarding time.
- Lack of expertise is the number one brake reported by French companies, whether they use AI or not, according to Insee (2026).
- In-house makes sense when AI touches sensitive data, a recurring workload or a future in-house product.
- Outsourcing makes sense for a first project, a scope that is still unclear, or a rare and temporary skill.
- In a small company, the most solid setup is almost always hybrid: a partner frames and builds, an employee takes over operations.
What an in-house AI profile really brings, and what it does not fix
An employee who knows AI understands your data, your processes and your constraints. They are there on Monday morning when an automation breaks, they understand why a salesperson ignores the tool that was delivered, and they can train colleagues as they go. That proximity has real value, hard to put a number on but easy to observe: projects that fail usually fail for lack of internal ownership, not for lack of technique.
Hiring does not solve everything. A data scientist knows how to train a model, rarely how to frame a project, arbitrate a budget and bring a business team on board. A senior profile who can do all three is expensive and hard to find. The Apec is blunt about it: the dual expertise in statistics and software development makes this profile rare for companies. Add knowledge of your own business and you understand why a standard job ad attracts thirty applications, of which three are relevant.
Another blind spot: one AI employee on their own does not produce much. They need clean data, an internal sponsor, trained users and a minimum of governance. Without that base, the hire becomes a fixed cost that watches projects go by.
The full cost of a hire, beyond the salary
Salary is the visible part. For a data scientist, the Apec reports that 80% of salaries offered in job ads fall between €35k and €60k, with an average of €46k. Add payroll charges, tooling, a workstation, sometimes a test cloud account, then recruitment and onboarding time. Count several months before the person produces anything useful, plus a manager to supervise them.
The real hidden cost sits elsewhere: commitment. A consultancy stops when the engagement ends. An employee stays, even if the project is postponed, even if management changes its mind, even if adoption never takes off. In a company of twenty or thirty people, that rigidity weighs more than a high day rate.
There is also a replacement cost. AI skills move fast, and a junior hired for a specific language or tool can be out of date within eighteen months. In-house therefore requires a continuous upskilling plan, otherwise the investment dilutes.
The expertise gap, a documented brake
In 2025, 18% of French companies with 10 or more employees reported using at least one AI technology, up from 6% in 2023, according to Insee (2026). The figure hides a wide size gap: 15% for companies with 10 to 49 employees, 31% for those with 50 to 249, 58% above 250. The smaller the structure, the more adoption costs in time and fixed skills.
The study also gives the nature of the blockage. Among companies that do not use AI, 54% mention a lack of expertise. And more than half of the companies that do use AI say they are held back by a lack of expertise. In other words, the problem does not disappear when you start: it changes shape. You can run a tool, but you cannot extend it, audit it or adapt it to a new department.
That finding moves the debate. The question is not "in-house or external" in the abstract, but "where do we place the expertise so that it is available when the need appears". A full-time expert for a need of two days a month is waste. No expert available for six months is a blockage.
When in-house is the right call
Some contexts make hiring hard to avoid.
- Sensitive or regulated data. Health, finance, legal, HR: the tighter the scope, the more reassuring it is to have an identified person accountable for processing and able to answer an audit.
- Recurring volume. If AI has been processing a hundred cases a day for a year, the capability has become a business function. It needs an owner, not a visitor.
- Cross-functional need. An AI profile that equips marketing, support and operations creates more value than an isolated project. That coordination works poorly from the outside.
- A project that becomes a product. An internal conversational agent or a quoting engine used by clients requires continuous changes, testing and a roadmap. That is a job, not an engagement.
- Transmission. Training teams continuously requires presence. A consultant who shows up every two months leaves gaps.
In those cases, hiring must be prepared. An AI profile without clean data or a sponsor will not last six months. Before opening the role, you need to know which use cases will be handled, who supplies the data and how results will be measured. The use cases page gives concrete examples to frame that work upstream.
When outsourcing is more rational
Conversely, four situations clearly favour a partner.
The first project. Nobody in the company knows what a successful AI project looks like. A partner brings a method, benchmarks for delivery time and a view of the traps. It is expensive, but less than a failed hire.
An unclear scope. As long as the use case is not settled, hiring freezes a hypothesis. A short engagement to explore, then a decision, is cheaper.
A rare and temporary skill. Model engineering, agent security, quality evaluation: these skills are better rented than employed full time in a thirty-person company.
A tight deadline. A need that must deliver before the end of the quarter cannot absorb four months of recruitment. An external audit lets you start within weeks and decide afterwards.
Outsourcing has its own limits. A partner leaves with what they learned at your company, and dependency can become expensive at renewal. Hence the importance of clear terms: ownership of code and models, data access, documentation, and skills transfer written into the contract. The audit and advisory page describes that kind of upstream framing.
The hybrid model that fits a small company
Between the two, one setup works better than either extreme: have a partner build, have an employee operate.
| Criterion | In-house hire | Partner or agency | Hybrid model |
|---|---|---|---|
| Entry cost | Annual salary, charges, tooling | Billed per engagement or per day | Framing engagement, then an internalised role |
| Time to start | Several months | A few weeks | A few weeks, gradual ramp-up |
| Business knowledge | Low at first, high later | Rebuilt on every engagement | Built jointly, then absorbed |
| Risk if the project stops | High (fixed cost) | Low (end of engagement) | Medium, limited by a bounded scope |
| Skill after the project | Stays in the company | Leaves with the partner | Stays, carried by a trained employee |
| Best fit | Recurring, structural project | First project, rare expertise | A company building a capability |
The principle is simple. The partner takes the riskiest part: framing, technical choices, first version, documentation. The employee takes the most durable part: operations, improvement, training colleagues, working with business teams. That employee does not need to be a data scientist. An operational profile trained in AI, who knows the business and can steer a tool, is often enough, provided they are supported at the start.
That is why AI training comes before hiring, not after. Training two or three key people costs less than a role and makes the company able to absorb an engagement without depending on it.
Five questions to decide
Before opening a role or signing a quote, ask these questions out loud.
- Is the need recurring or one-off? One-off and unclear: outsource. Recurring and measured: hire.
- Is the data available and clean? If not, no hire will save the project. Start there.
- Who will own this in two years? If there is no answer, the partner will leave with the skill.
- What annual budget can you freeze? A salary is paid twelve months, an engagement is bounded.
- What must you be able to do yourselves at the end? The answer defines what must be documented and transferred, and therefore what belongs in the contract.
If three answers out of five point outside, a project of AI agents framed by a third party remains the fastest route. If they point inside, prepare the ground before the job ad, otherwise the role will stay vacant or be badly used.
FAQ
Do you need a data scientist to start an AI project?
No. Many business projects rely on existing tools, configured agents and process automation. A data scientist becomes useful when you build your own models or must handle large data volumes. For a first agent or a quoting automation, a well-trained operational profile is often enough.
How long does it take to hire an AI profile?
It depends on the level and the rarity. On experienced profiles, the Apec reports strong expectations, often beyond three years of experience for a complete profile, which lengthens searches. Plan several months between the job ad and the start date, then onboarding time before the first results.
Can you outsource an AI project without losing control of the data?
Yes, provided it is written down. Hosting, retention period, access, deletion at the end of the engagement, subcontractors: these points belong in the contract. A serious partner documents them without being asked. If the answer stays vague, that is a signal.
Should you hire before or after choosing use cases?
After. A hire launched without a settled use case produces an employee looking for a mission, or assigned to unrelated tasks. Start with a short list of use cases, estimate their value, then deduce the profile and the way of working.
Can a small company train an existing employee instead of hiring?
It is often the best compromise. An employee who knows the company learns to use and steer AI faster than an outside expert learns the business. It still requires a structured programme and real use cases to practise on, otherwise the training stays theoretical.
Conclusion
There is no universal answer between hiring and outsourcing, but there is a method. Look at how recurring the need is, the state of your data, who will own the topic in two years, and the budget you can freeze. Most companies that succeed with their first project do not pick a camp: they have a third party frame the work, train an employee, and internalise what lasts.
If you are still unsure about the setup, describe your situation on the contact page: a thirty-minute conversation is usually enough to see whether the need calls for a role, an engagement or a training programme. To go further on uses and costs, the article on AI agent use cases and costs in business details what a project really involves. And if you are comparing several service options, the checklist of questions to ask before signing avoids unpleasant surprises.
