Agency vs In-House AI: Build or Buy Your AI Project in an SME?
For an SME, the right answer is rarely "all in-house" or "all agency". Most of the time, you bring in a provider to scope and launch the first project, then bring in-house whatever comes back every week. Here are the criteria to decide, with figures you can check.
This article is part of our complete guide to AI in business for SMEs. It assumes you have already identified one or two use cases; if not, start with an AI audit.
Build or buy: the real question is not the one you think
The decision is not about AI in general, but about each project taken separately. An SME can perfectly well hand its invoicing automation to a provider and train its sales team to write their own prompts. These are two different decisions with different logic.
Three situations come up most often:
- you want to use existing tools (a conversational assistant, an office copilot): this is mostly a training and governance topic, not a development one;
- you want to connect tools with no-code or automation: a trained in-house lead can handle a good part of it;
- you want custom development (an agent, an integration with your business software, processing of your documents): this is where the agency versus in-house question really applies.
Insee figures show where most companies stand. In Insee Première no. 2120, 18% of companies with 10 or more employees use AI in 2025: 15% of those with 10 to 49 employees, 31% of those with 50 to 249, and 58% from 250 upwards. Among users, 80% buy ready-made software or systems. Only the group with the most varied uses develops heavily in-house (nearly two thirds of its companies). In other words, building it yourself is the exception, even among advanced companies.
What an in-house team really costs
An in-house team costs far more than the salary of the person you hire. Gross salary is only the first line. To compare fairly with an agency quote, list at least:
- the fully loaded salary of the profile or profiles (not just gross pay);
- recruitment cost, which can be significant for scarce profiles;
- infrastructure: licences, hosting, monitoring tools;
- ongoing training, since tools change several times a year;
- management time, often forgotten, which falls on the owner or a technical manager;
- replacement cost if the person leaves, which happens quickly in a tight market.
We do not give a salary range here: they vary widely by region and profile, and figures circulating online are rarely sourced. Get yours from your recruiter or accountant, then add the other lines. Our article on the cost of AI for an SME in 2026 details the spending items for tools and integration.
A simple test: if, once all these items are added up, the annual amount exceeds what you would spend on external services over three years, hiring right away is not justified.
What an agency costs, and what it does not give you
An agency or freelancer brings speed and experience, not ownership of the knowledge. A good provider has already seen the mistakes along the way at other clients. It delivers a pilot faster than someone discovering the subject. However, when the engagement ends, the project knowledge stays with them if you have not planned for it.
Points to check before signing:
- Ownership of code and workflows: who owns what, and in what format is it delivered?
- Reversibility: can you take over the project alone, or with another provider, without starting over?
- Documentation: is it included in the price, with a skills transfer to your team?
- Data hosting: where do your documents go, and through which subcontractors? This is a GDPR topic as much as a quality one.
- Maintenance: who fixes things when a tool changes its interface or its pricing?
An agency that will not answer these five questions clearly is selling you a dependency, not a project.
Public funding: a point often left out of the calculation
Public funding mainly reduces the cost of the diagnostic, not of the development. According to the "Osez l'IA" plan strengthened on 17 June 2026 (entreprises.gouv.fr), the Diagnostic Data IA is an 8 person-day service priced at 10,000 euros excluding VAT, 40% covered by France 2030. It is reserved for SMEs and mid-caps with 10 to 2,000 employees registered in the trade register. Its purpose is to identify and prioritise use cases.
For an eligible company, the remaining cost is therefore 6,000 euros excluding VAT out of 10,000 (calculation: 60% of 10,000). The Accélérateurs IA target SMEs and mid-caps with more than 50 employees, more than 8 million euros in revenue and at least three years of existence, with 46% covered over 18 months.
Two limits to keep in mind: very small businesses are not explicitly targeted by these schemes, and conditions can change. Check them before building a budget. For training actions, other funding exists (see our article on the AI budget for an SME in 2026).
The decision table
The more one-off and complex your need, the more an external provider makes sense; the more recurring and close to your core business, the more in-house becomes relevant. This table summarises the most common cases.
| Situation | Best-fit option | Why |
|---|---|---|
| First AI project, no internal skills | Provider, with skills transfer | You avoid early mistakes |
| Off-the-shelf tools (assistant, copilot) | In-house, after training | Little development, a lot of usage |
| Simple automations between your tools | In-house lead or one-off freelancer | A trained lead is often enough |
| Custom development, one-time need | Provider | No reason to hire for one project |
| Several projects a year, lasting backlog | In-house, with external support at peaks | Business knowledge becomes critical |
| Highly sensitive data | In-house or provider under strict contract | Control of data flows comes first |
This table is a starting point, not a verdict. Your AI maturity changes the answer: the lower it is, the more useful external support is at the start. Our article on assessing AI maturity gives a grid to place yourself.
The hybrid model: provider first, in-house later
The pattern that comes up most often is: scope with a third party, build with them, then take over. It runs in three stages.
- Scoping (a few weeks). An audit identifies two or three use cases and estimates the expected gain. You keep the deliverable, whichever provider handles what follows.
- Supported pilot. The provider builds alongside a designated internal pair. That pair documents and asks questions; they will carry the project afterwards.
- Handover. Once the pilot is stable, the company takes over operations and calls on outside help only for complex changes.
This approach requires naming someone internally, often an AI lead, and freeing up their time. It is the most frequent mistake: a project handed to a provider with no counterpart on the company side rarely reaches production. We cover other pitfalls in AI deployment mistakes in SMEs.
Train rather than hire: the option people forget
Upskilling your existing team often costs less than a hire, and it is a prerequisite in every case. Even with an agency, your people must be able to state a need, test a result and spot a wrong answer. Citizen developers, non-IT colleagues who automate their own tasks, are an underused resource in SMEs.
A short course built around your real use cases delivers more than a generic one. That is what we build in our AI training for business, which may be eligible for funding depending on your situation (to be checked case by case with your OPCO).
How to tell your choice was the right one
Set three indicators before you start, not after. Without them, you cannot tell whether the agency kept its promises or the in-house team is progressing.
- A gain indicator: hours saved per week, or shorter processing time.
- A usage indicator: how many people actually use the solution after three months.
- An autonomy indicator: what share of changes your team can make without help.
For calculation methods, see our article on calculating AI ROI in an SME and the definition of AI ROI. A 90-day roadmap, like the one in our article on AI strategy for an SME in 90 days, helps set clear milestones.
Frequently asked questions
Is it better to start with an agency or with a hire?
In the vast majority of SMEs, starting with a provider is less risky, because you do not yet know how many projects you will have. Hiring makes sense when projects repeat and there is lasting work for the person.
Can a freelancer replace an agency?
Yes for a bounded, well-described project. The risk is dependence on one person: require documentation and access to all accounts from day one.
How long does an outsourced AI pilot take?
It depends on scope. We prefer not to quote a generic duration: a serious scoping phase lets you set a realistic schedule for your case, with a checkpoint every two weeks.
Do public schemes cover development?
The schemes we checked mainly target diagnostics and support (Diagnostic Data IA, Accélérateurs IA), with company-size criteria. For development itself, check regional aid in your area.
How do I avoid depending on a provider?
Put in writing the ownership of code and workflows, delivery of documentation, administrator access to your accounts and a skills transfer. Also plan an exit clause.
What if my data is sensitive?
First map which data goes into the tool and where it is processed. Depending on the result, you can stay on tools with controlled hosting, or limit the provider to anonymised data.
How GrowthPerf helps
GrowthPerf is a Qualiopi-certified training organisation specialised in AI, no-code and automation for SMEs and associations. We work both ways: we train your teams so they gain autonomy, and we scope or build with you the projects that need outside help, planning from day one for your internal pair to take over.
If you are torn between an agency and an in-house team, the most useful step is to put your case in writing: use case, expected volume, data involved, budget. Prepare a one-page brief, and we will review it with you and tell you whether the external, internal or hybrid option holds up. See also our guide to AI in business for SMEs to place this choice within your overall strategy.