Bringing AI into a small business: where to start, costs, risks
Where AI pays off in a small business, how to start with a single process, what it costs to run, and the risks to manage: data, errors and the EU AI Act.
Updated Published 10 min readLire en français
Start with a single repetitive process, measured before and after, where AI reads, sorts or drafts something that a person then approves. That is where the gain is easiest to measure and the risk stays under control. Before you launch, work out what it will cost to run (subscriptions, usage-based calls, review time) and decide which data the AI is allowed to read.
How far have French businesses gone?
According to Insee, France's national statistics office, 18% of French companies with 10 or more employees used at least one AI technology in 2025, up from 10% in 2024 (Insee Première no. 2120, in French, with an English summary). The gap by company size is still wide:
Share of French companies using AI in 2025, by size
Among companies that already use AI, more than half use it to analyse text (56%), and 43% to generate text or speech. Their brakes are telling: 53% say they limit their use for lack of expertise, 43% because of data protection and 42% because the legal situation is unclear (same source). Across the European Union, Eurostat counts 20% of companies using AI in 2025 (Eurostat).
For a small business, two lessons stand out: text (emails, documents, meeting notes) is the most common ground, and the obstacles cited most often are best tackled with method first.
Four ways to use AI in a business
| Approach | Examples | Billing | Watch out for |
|---|---|---|---|
| Chat assistant | ChatGPT, Claude, Gemini, Mistral | Subscription per user | What staff paste into it |
| AI inside your software | Microsoft 365 Copilot, Gemini in Google Workspace | Per user, add-on or included | Access rights on shared documents |
| Automation with an AI step | n8n or Make, connected to your tools | Platform plus API calls | Silent errors, monitoring |
| Custom AI feature | Chatbot on your documentation, document analysis | Development plus API calls | Security, costs, maintenance |
A chat assistant is the easiest entry point, and the riskiest if left unmanaged. On ChatGPT's consumer plans (Free, Go, Plus, Pro), OpenAI states that content is used to train its models, with an opt-out (ChatGPT plans); business workspaces are excluded from training by default (OpenAI, ChatGPT Business). The CNIL, France's data protection authority, advises never sharing confidential information in a consumer service (CNIL, in French). So give your team business accounts and a one-page usage policy.
AI inside your software (word processor, email, CRM) is billed per user, as an add-on or included in the plan depending on the vendor: see the pages for Microsoft 365 Copilot and Google Workspace. Before switching it on, check who can open which shared folders: Microsoft, for example, says Copilot can draw on any organisational data the user has at least view permission for (Microsoft Learn). A folder shared too widely becomes readable through everyone's assistant.
Automation with an AI step connects your tools and gives AI one precise job: read an attachment, classify a request, prepare a reply. This is often where time saved is easiest to measure, because the task is clearly defined. My comparison of n8n and Make (in French) helps you choose the tool.
A custom AI feature lives inside your own application, through a provider's API. It needs the same care as any product in production: data, costs, quality, GDPR and the EU AI Act. The guide to putting an AI feature into production covers them in detail.
Where AI really pays off
A good first use case ticks these four boxes:
- it comes back every week, with enough volume for the gain to show;
- it deals with text or documents: emails, PDFs, forms, meeting notes;
- a mistake can be caught, because a person approves before anything goes out;
- it can be measured: time spent, response time, number of times data is re-keyed.
Which use cases should you launch first?
Three cases call for caution:
- decisions about people, such as screening job applications or assessing a customer's creditworthiness. The EU AI Act classes these as high-risk, with obligations from 2 December 2027 (Annex III), and GDPR restricts fully automated decisions;
- exact calculations (prices, quantities, due dates): a rule written into your tool is more reliable than a language model;
- sending without review when a mistake is costly, to a customer or a partner.
Often, connecting two tools is enough, with no AI at all. AI earns its place when it beats a simple rule: reading documents whose layout varies, understanding a request written in free text, summarising a long thread.
Where to start: a five-step method
- Pick a processFrequent, with someone who does it today.
- Measure the current stateTime spent, volumes, exceptions, sensitive data.
- Test on real casesExamples from your business, with the expected answer.
- Go live with reviewA person approves, an alert fires on errors.
- Measure and decideCompare with the baseline, then extend or stop.
Involve the people who do the task from day one: they know the exceptions, and they will be the ones using the automation. Set the success criterion before building, for example time spent per week or response time. Then compare several options, with their cost and risk: improve what exists, connect the tools, or add an AI step. Finally, document how it works and train the users, so the team does not depend on a single person.
What it costs to run
Model and subscription prices change often, so this guide describes the cost items and points to the official pricing pages instead of copying figures.
- Per-user subscriptions for assistants and AI add-ons in your software: ChatGPT, Claude, Mistral, Microsoft 365 Copilot, Google Workspace.
- API calls, billed per token, a chunk of a word of about 4 characters in English (OpenAI). Input and output are priced differently, and prices vary by model: see OpenAI, Anthropic, Mistral and Google. For work that can wait, OpenAI and Anthropic both advertise a 50% discount on their batch APIs (OpenAI, Anthropic).
- The automation platform: n8n bills per workflow execution, whatever the number of steps, and its Community edition costs nothing for internal use if you host it yourself (you still pay for the server and its upkeep); Make bills in credits, used up by most actions (n8n pricing).
- People's time: reviewing answers, following up errors, updating things when a tool changes. This item is often left out of estimates.
To decide, write the sum down: monthly cost = subscriptions + number of tasks × average cost per task + platform + supervision time. Compare it with the time saved, measured in step 2. Set-up is a one-off on top: at Techvisor, the AI & automation offer starts at €2,500 excl. VAT for a process with few tools and few exceptions, and the quote is fixed after observing the process.
The risks, and how to manage them
- Data. Decide in writing what the AI may read. With an API, control of the system sits almost entirely with the provider, the CNIL notes: check the contract, the retention period and where processing happens.
- Plausible mistakes. A model can be wrong with total confidence. The CNIL says outputs should never be reused as they are: keep a human review wherever an error is costly.
- Dependence on one provider. Models and prices change. Keep your instructions, test cases and data in-house, so you can switch models without starting over.
- Security. An AI that reads outside emails can find hidden instructions in them: this is prompt injection, first on OWASP's list. Limit what it can do without approval.
- The EU AI Act. Since 2 February 2025, companies that use AI must take measures to build their staff's AI literacy (Article 4). Since 2 August 2026, a chatbot must say it is an AI, unless that is obvious (Article 50).
A one-page usage policy is enough to start. It lists the approved tools, with business accounts, and the data that must stay out of them (customer data, health, HR, contracts) unless a contract covers it. It makes a review mandatory before anything is sent outside, and names the person to alert when in doubt.
My take as a CTO
At Jo's Burger, where I am a partner and the CTO, I built the app and the back office, integrated payment and delivery, then set up process automation. At Welco, I integrated AI to optimise logistics. At Axiome Avocats, a law firm, I compared several scenarios for the firm's tools, with their costs and risks, and assessed AI use cases before proposing a phased roadmap. With Techvisor, I automate with n8n and OpenAI. My advice: start from a specific task rather than from a tool. In many cases, connecting two pieces of software already does most of the work, and AI only steps in to read a document or classify a request. I have also been a tech trainer at Campus Numérique since December 2023, and every automation I deliver includes training for the people who use it.
FAQ
How much does it cost to bring AI into a small business?
There are two amounts. Set-up first: at Techvisor, from €2,500 excl. VAT for a process with few tools. Then running costs: subscriptions, API calls, the automation platform and review time. Estimate them before launch, from your real volumes.
Do we have to use ChatGPT?
No. OpenAI, Anthropic, Google and Mistral all offer models for everyday text tasks. The choice depends on the task, constraints on your data such as processing in Europe, and the price at your volume.
Can staff use AI with a personal account?
It is not advisable. On ChatGPT's consumer plans, content is used to train the models unless the user opts out, and the CNIL advises against sharing confidential information there. Provide business accounts and a usage policy.
Do we need a developer to automate with AI?
Not always. Make is very visual and usable without a developer; n8n is more technical but goes further. As soon as the automation touches sensitive data or plugs into your own application, get help.
What does the EU AI Act mean for a small business using AI?
Three things matter today: train your staff (Article 4), disclose a chatbot as an AI (Article 50), and do not use AI to screen job applicants or assess someone's creditworthiness without preparing for the high-risk obligations that apply from 2 December 2027.
Next step
- To move forward on your own: my comparison of n8n and Make (in French) to choose your automation tool, and the AI sheets for small businesses (in French) published by France Num with the CPME and the CNIL.
- To automate a first process: with the AI & automation offer, I observe the process with your team, connect your tools, add AI where it beats a simple rule and train the users. From €2,500 excl. VAT, quoted after observation.
- To train your team: AI training or coaching comes with a tailored programme and a separate quote.
Tell me about the task that comes back every week: book a 15-minute call.
Sources
- Insee, Les technologies de l’information et de la communication dans les entreprises en 2025, Insee Première no. 2120 (in French), and its English summary
- Eurostat, Use of artificial intelligence in enterprises
- CNIL, Q&A on using a generative AI system (in French)
- CNIL, France Num and CPME, Using generative AI in very small and small businesses (in French)
- European Commission, AI Act Service Desk: Article 4, Article 50, Annex III
- OpenAI, ChatGPT plans, Managing data, sharing, and privacy in ChatGPT Business, API pricing, Batch API, What are tokens
- Anthropic, Claude plans, API pricing, Batch processing
- Mistral AI, Pricing; Google, Gemini API pricing, Google Workspace pricing
- Microsoft, Microsoft 365 Copilot Business, Data, Privacy, and Security for Microsoft Copilot
- n8n, Pricing
- OWASP, Top 10 for LLM Applications 2025
AI & automation
Automate a task that costs you hours
One specific task (re-typing, sorting, looking up information), your tools connected to each other, AI only where it adds something, and your team trained.