As a rough guide, in 2026 an AI project for an SMB costs anywhere from €5,000 for a simple automation to more than €200,000 for a platform hosted in-house. The most common ranges: €8,000 to €25,000 for a serious proof of concept, €15,000 to €60,000 for an internal RAG chatbot, €40,000 to €150,000 for a business agent, and €30,000 to €150,000 to embed AI into an existing product. On top of these amounts come recurring costs that typically represent 15 to 30 percent of the initial budget every year.
2026 price guide: ranges by project type
These ranges are estimates drawn from current practice on the French market: every project depends on your context, your data and your information system. They are nonetheless useful to benchmark a quote and to spot proposals that are abnormally low or inflated.
Simple automation: €5,000 to €20,000
Automatic email triage, invoice data extraction, meeting summaries, assisted support replies: these projects rely on existing models accessed through APIs and orchestrated by workflows. Development takes from a few days to a few weeks. It is often the best entry point: fast return on investment, limited risk, and the company learns how to work with AI. For many SMBs, two or three of these quick wins are enough to prove the value of AI internally and to fund the next steps.
Internal chatbot or RAG: €15,000 to €60,000
An assistant that answers employee questions based on your documents (procedures, contracts, technical documentation) relies on a RAG architecture: content indexing, semantic search, controlled generation. The cost varies with the volume and quality of the documents, the access rights to enforce and the level of reliability required. Data preparation often accounts for half of the effort.
Business agent: €40,000 to €150,000
An agent that executes a complete process — qualifying inbound requests, preparing quotes, tracking case files — must connect to your tools (CRM, ERP, email), make decisions within guardrails and know when to hand over to a human. The complexity comes from the integrations, the edge cases and the safeguards. Expect several months of work, with an essential pilot phase.
Embedding AI into an existing product: €30,000 to €150,000 and above
Adding AI features to your software or platform (smart search, recommendations, input assistance) brings production-grade requirements: latency, scalability, controlled unit costs, user experience. The budget depends mostly on the maturity of your code base and the level of industrialisation you are aiming for.
AI hosted in-house: €60,000 to €250,000 and above
Hosting open-source models on your own infrastructure or a private cloud makes sense for highly sensitive data or large volumes. You need to budget for GPU servers, deployment engineering, monitoring and model updates. This option should be assessed on a three-year total cost of ownership, rarely below six figures.
The factors that really move the price
The AI model itself is a small part of the budget. The real cost drivers lie elsewhere:
- The state of your data: scattered, unstructured or poor-quality documents can double the effort.
- The number of integrations: every connection to an existing tool (ERP, CRM, telephony) adds days of work.
- The required level of reliability: going from 90 to 99 percent correct answers costs disproportionately more.
- Regulatory constraints: GDPR, regulated industries or health data call for more demanding architectures.
- Change management: training and supporting the teams, which is often underestimated.
Recurring costs: the budget also plays out after go-live
An AI project is not a one-off purchase, and vendors rarely volunteer the running costs unless you ask for them in writing. As a rough guide, plan for 15 to 30 percent of the initial budget every year:
- API consumption: from a few dozen to several thousand euros per month depending on volumes.
- Hosting and infrastructure: servers, vector databases, monitoring.
- Ongoing maintenance: models and APIs evolve quickly, so regular testing and adaptation are required.
- Quality control: continuous evaluation of answers and correction of drift.
How to scope the project and keep the budget under control
Method matters more than haggling over daily rates:
- Start from a quantified business problem, not from a technology: time lost, error rates, response delays.
- Insist on a short scoping phase (one to three weeks) that delivers a precise perimeter, measurable success criteria and a committed estimate.
- Proceed in stages: POC, pilot on a real perimeter, then rollout, with a budget decision point at each step.
- Put recurring costs in the initial contract, not after the system is in production.
The classic traps that make costs explode
- Starting development without scoping, on the strength of an appealing demo.
- Aiming for a company-wide AI platform from day one instead of a first high-value use case.
- Ignoring data quality and discovering the problem mid-project.
- Overlooking API costs in production, which can exceed the development cost over three years.
- Remaining dependent on a vendor with no documentation and no reversibility.
A well-scoped AI project is a measurable investment, not a gamble. If you are weighing several scenarios or several quotes, an experienced external perspective — one that can challenge both the architectures and the estimates — generally saves far more than it costs, and lets you compare proposals on a like-for-like basis rather than on price alone.