How to choose an AI agency and what budget to plan?

AI agencies, consulting firms, software vendors, integrators, freelancers: for the same need, a productivity gain on a process, an answer engine on internal documentation or a new experience for customers and visitors, a company now receives proposals whose amounts range from one to ten.

How do you compare providers, and from what budget is an AI project serious?

In reality, the choice depends less on the technology announced than on the provider's ability to deliver a solution that works in the company's real conditions, to get it adopted by the teams and to make it evolve. An AI project does not stop at delivery.

AI agency, consulting firm or software vendor: what do you need?

An AI agency designs, builds and puts into production a solution adapted to the company's business, data and tools. A consulting firm recommends and scopes, without delivering. A software vendor sells the same product to all its customers.

The first question, the one also asked by the France Num sheet on choosing an AI solution (in French), is not "which tool?" but "what exactly do you want to achieve?". The answer points to the right kind of provider:

• a vendor is enough when the need is standard: transcribing meetings, translating, writing e-mails, summarising documents. A subscription, not a project
• a consulting firm is useful when the question is strategic: an inventory of uses, priorities, organisation, before any investment
• an agency becomes necessary when the value depends on the company's data and tools: an answer engine on internal documentation, an agent connected to the CRM and the calendar, a voice installation in a visitor venue

The boundaries are porous, many agencies also scope the uses. The useful question: who carries the production launch, and who provides support six months later?

What are the criteria that make the difference?

Several criteria make it possible to compare AI agencies on what matters after signing: delivered projects, a method that goes from prototype to industrialisation, adoption by the teams, long-term follow-up, data sovereignty, a method to limit errors, the choice of models and reversibility.

Delivered projects

A serious agency presents a selection of projects: the initial need, how the solution works, what it changed for the teams. Client logos without explanation say nothing. Two simple checks: read the case studies and ask for a demonstration.

A method that goes from prototype to industrialisation

The prototype is a step, not an end. It is there to check, on real data, that the solution brings something. Then comes industrialisation: connection to the tools in place, access rights, security, monitoring, scaling, maintenance. An agency must be able to describe what happens after the prototype, with which timelines and which budget. One that only talks about the prototype has often not planned what comes next.

Adoption by the teams

A solution that is not used is worth nothing, whatever its technical quality. France Num invites companies to ask who will use the tool, with what level of skill, with what enthusiasm or what fears, in order to prepare change management. The agency must take on this part of the project: user training, written conditions of use, an interface suited to the teams' habits, and measurement of real usage in the weeks following go-live.

Follow-up and continuous evolution

An AI project delivered without follow-up degrades. Real usage brings up unforeseen questions, reference documents change, language models are replaced every six months by better ones. Follow-up is therefore needed: analysis of real exchanges, optimisation of answers, updates of the knowledge base and the models, and a regular measurement of the return on investment, as France Num recommends. This follow-up is written into the contract, it is not guessed.

Hosting and data sovereignty

Where do the models run, who has access to the data, and is it used to train third-party models? For customer data, case files or internal documents, sovereign AI means hosting in France or in Europe and, when confidentiality requires it, the company's own servers or a fully local installation. The GDPR and the European AI Act rest on the same principles: human control, robustness, privacy protection, transparency. A company that entrusts personal data to an AI remains responsible for it.

Tests and a method to limit errors and hallucinations

No AI is free of errors. The difference lies in the method to limit and detect them. RAG (Retrieval-Augmented Generation) anchors answers in the company's documents, a set of real questions measures the rate of correct answers before go-live, and a behaviour outside the scope is defined: the solution must be able to say that it does not know. A human review remains planned for answers that commit the company.

Reversibility

Who owns the code, the prompts, the knowledge base and the usage data? Can you change hosting provider, model or agency without rebuilding everything? France Num warns against dependency and loss of know-how: exit conditions are negotiated before signing, not after.

The choice of language models

An agency must explain which language model it proposes and why: a proprietary model accessed through an API, or an open-weight model, whose weights are published and which can be hosted in France or in Europe, or even on the company's own servers. Open models, including those of the French company Mistral, reduce dependency on a supplier and on its pricing policy. The right reflex is to choose a model sized for the need rather than the largest one by default, and, in every case, an architecture that allows the chosen model to be replaced by another, because models change every six months.

What do the public reference points say?

Several free, general-audience documents help frame the choice of an AI provider. The France Num practical AI sheets, written with the CNIL in November 2024 (in French), list the questions to ask before choosing a solution, the precautions of use and the obligations linked to the GDPR. Bpifrance's IA Booster France 2030 programme, launched in June 2023, offers SMEs and mid-caps a four-phase path, including a phase of help in choosing the AI solution, with partial funding of the services under conditions.

The ANSSI security recommendations for a generative AI system, published in April 2024 and also available in English, detail 35 measures, from risk analysis to system segmentation and access logging. Finally, the practical sheet for the responsible purchasing of AI solutions from DINUM and the French State purchasing directorate, released in November 2025, provides questions to ask providers, vendors and hosting companies, including on environmental impacts. These reference points are written for public buyers or SME managers, but their questions apply to any company, and an agency that knows them answers without difficulty.

What budget should you plan for an AI project?

As a guide, for a first project in a small or medium-sized business, a first use case on real data, the proof of concept (POC), costs between €3,000 and €5,000 excluding VAT and takes a few weeks. A first version teams can use, the minimum viable product (MVP), costs between €5,000 and €10,000 excluding VAT, not including hosting and model usage. Going into production then takes one to three months. These are the orders of magnitude practised by Gensai: a few thousand euros to find out, a few tens of thousands to deploy.

As with a website, where a five-page showcase and an e-commerce platform are not priced the same way, the budget of an AI project can go much higher depending on the specifics, the scale of the project and the client's expectations.

There is no public price list for an AI project. One reference point nevertheless gives a scale. Bpifrance's Diag Data IA, eight days of an approved expert to identify and prioritise use cases, is billed €10,000 excluding VAT, 40% of which is funded (les-aides.fr sheet, August 2026, in French). An AI project is therefore priced first of all in days of expertise.

What makes the price vary, above all, is the time spent by the teams and their hourly rate: scoping, development, integration, testing, training. That time depends on a few parameters:

• the number of tools to connect: CRM, calendar, telephony, document base, each connection is a development, and sometimes an API that does not exist yet has to be created
• the state of the data: scattered, duplicated or unclean documents cost more than an already structured base
• the channels: an assistant on WhatsApp, on a website or in the form of an avatar means three different interfaces. WhatsApp is the simplest to set up, a website requires specific development, with web design and code, and an avatar is more complex still
• voice or text: a text assistant and a voice agent do not cost the same. Voice adds speech recognition and synthesis, real time, the handling of interruptions and accents, so more components to assemble and test
• the model: a large language model (LLM) used as is, with RAG, covers most needs. Adding a layer of further training, fine-tuning, to specialise the model on the company's data requires a prepared dataset, computing power and a new round of tests: the cost changes scale, and it is rarely necessary for a first use case
• security and protection: segmentation, access rights, logging, audits. Applying the ANSSI recommendations to a sensitive system can be very expensive. Hosting the solution on your own servers removes token billing, but you then have to buy or rent GPUs, operate them and monitor them

On top of development come recurring costs: hosting, model usage, billed in tokens according to the amount of text processed, maintenance and evolution, and team training. France Num reminds us that the cost is not limited to paying for the solution. An installation that runs locally, like the voice experience installed in an aquarium, removes the usage cost, at the price of dedicated hardware.

A quote without stages or scope cannot be compared. The right reflex is to ask each provider for the price of the first use case, then the price of what follows, starting with a first project chosen for its value and feasibility.

Which questions should you ask at the first meeting?

Seven questions are enough to sort providers, and a good agency answers them with examples:

1. Can you run a demonstration on my documents?
2. Which solutions are running in production at your clients today, and since when?
3. What happens after the prototype: which stages, which timelines, which budget?
4. How do you support adoption by the teams?
5. What do you plan after delivery: follow-up, optimisation, model updates?
6. Where will my data be hosted, will it be used to train a model, and what does the solution do when it does not know?
7. What happens if I want to change model, hosting provider or agency?

Vague answers are an answer. An agency that puts solutions into production regularly has figures and examples for every question.

The right agency starts small, drives adoption and keeps evolving

Technical sophistication does not set AI agencies apart. What sets them apart is accepting a limited first use case, tested on real data, then the ability to get it adopted by the teams and to make it evolve with usage.

It is also the best way to keep the budget under control: each stage is decided with the results of the previous one.

An AI agency to compare? Gensai, AI agency in Paris, applies this method: a first use case in a few weeks, then deployment of the solution, team training and long-term follow-up.