Which AI for which task? The SME guide

Daniel Platteau

Updated on 10 November 2026. Prices taken from the vendors' official pages on that date, in US dollars excluding VAT unless stated otherwise. Plans change often: we will keep this guide up to date.

"Which AI should I get?" is the wrong question

It is the question we hear most, and also the most common one on forums: one subscription or several? Free or paid? ChatGPT, Claude or Gemini? It sounds like "which car should I buy?" asked without saying whether you carry children, pianos or customers. The answer depends on what you want to do, what you already use, and what you are going to hand over to the machine.

You are not alone in asking. According to Statbel, more than one Belgian company in three (34.5%) used at least one AI technology in 2025. According to Eurostat figures analysed by Acerta, language applications (analysing, producing or translating text) remain by far the most used.

This guide is a starting point. It helps you choose, then points to our detailed articles for each topic.

What people call "an AI": three different things

Before comparing, you need to know what you are comparing. The word "AI" covers three realities.

• The model: the "engine" that understands and produces text, images or sound. Models have technical names that change every three months. No need to remember them.
• The assistant: the application where you type your request (ChatGPT, Copilot, Gemini, Claude, Mistral's Vibe, Perplexity…). This is what you buy. It chooses the model and adds web search, file analysis, memory and connections to your tools.
• Built-in AI: the same technology, housed in software you already use. Copilot in Word and Outlook, Gemini in Gmail, transcription in Teams, the assistant in your accounting software or CRM.

For an SME owner, the right question is therefore rarely "which model?". It is rather "which assistant, or which built-in AI, for which task, and with what guarantees for my data?".

Free, personal subscription or business plan: what really changes

The major assistants all have three levels.

1. Free: useful for discovering. Usage limits, less powerful models, and sometimes advertising: OpenAI funds its Free and Go plans with ads, while Plus, Pro, Business and Enterprise remain ad-free.
2. Personal subscription (about 20 dollars a month): more capacity, better models, file analysis, in-depth research.
3. Business plan (about 20 to 25 dollars per user per month): admin console, central billing and, above all, a data processing agreement with a commitment not to train the models on your content.

The point that surprises most people: paying for a personal subscription does not protect your data by default.

• At OpenAI, conversations on the Free, Plus and Pro plans may be used to train the models as long as you have not switched off "Improve the model for everyone". Business, Enterprise and the API are not used for training by default.
• At Mistral, conversations with Vibe (formerly Le Chat) are used for training by default unless you opt out. On the Team plan, the administrator can switch this off for the whole organisation. Enterprise customers are excluded by default.
• At Google, with a personal account, your Gemini history ("Keep Activity" setting) is used to improve the models, with the help of human reviewers. Reviewed conversations are kept for up to three years, even if you delete your history. With a business Google Workspace account, there is neither human review nor training.
• At Anthropic (Claude), the Free, Pro, Max and Ultra plans offer a training setting that you can switch off. The Team and Enterprise plans are not used for training by default.
• At Microsoft, Copilot Chat used with a work account benefits from "enterprise data protection": your questions and the answers are not used to train the models.

The legal detail (GDPR, what you can and cannot share) is in our article AI and GDPR: what can you really entrust to ChatGPT or Copilot?.

The main assistants at a glance

The prices below are those displayed by the vendors in November 2026, in dollars excluding VAT. In Belgium, invoices are issued in euros and the final amount varies with VAT and the exchange rate. As an indication, Google AI Pro is billed €21.99 a month in France.

Two remarks. First, these tools are often on a par for everyday tasks: for an email or a summary, the difference lies more in the quality of the request than in the choice of tool. Second, an American or European vendor does not tell the whole story: a European vendor that trains its models on your conversations by default does not protect your data better than an American vendor under a business contract.

Which AI for which task: the table

This is the heart of the guide. For each task, three levels: the beginner who wants to try without paying, the freelancer ready to pay for a subscription, and the team that needs guarantees.

Let us go through the tasks one by one, with the traps to avoid.

Writing: the first use, and the most profitable

Delicate emails, reminders, quotes, job offers, LinkedIn posts: this is where AI saves time from day one. All the assistants do well. If your company uses Microsoft 365, start with Copilot Chat, already included in eligible business subscriptions. Same logic with Gemini for Google Workspace.

A tip worth more than a subscription: give context. "Write a reminder" produces a bland text. "Write a polite but firm reminder for a €2,400 invoice unpaid for 45 days, loyal customer for 6 years, Belgian tone without stilted phrases" produces a usable one. Our list of 10 administrative tasks to hand over to AI gives other examples.

Summarising and analysing documents

A 40-page contract, a specification, a competitor's annual report: AI summarises, compares two versions, flags unusual clauses. Paid assistants accept much longer documents than free versions.

Two reflexes. First, always check important passages in the original document: an AI can invent a clause that does not exist, with the greatest confidence. Second, a customer contract almost always contains personal data: no free version for that. On AI errors, see our article Can an AI lie?.

Finding up-to-date information

An assistant without Internet access answers from memory, and its memory dates from when it was built. For a regulation, a price or a news item, use a tool that searches the web and cites its sources: Perplexity, or the search mode of ChatGPT, Gemini and Claude. Then click on the sources. "Deep research" features, which read dozens of pages before answering, save hours on a market study, provided you reread the result like the work of a gifted but hurried intern.

Translating: an asset for a Belgian company

French, Dutch, English, German: translation is a task where AI excels. DeepL, a German vendor, remains a reference, and its Pro version deletes texts immediately after translation. General-purpose assistants also translate very well and can adapt the tone, for example for an offer aimed at a Flemish customer. For a contract or a legal document, human proofreading remains essential.

Meetings and transcription

Automatic transcription, summary, list of agreed actions: if you use Teams or Google Meet, these features exist in some plans. With Copilot Business, Copilot summarises Teams meetings and answers questions based on the transcript, which the organisation must allow. Be careful: recording a meeting means informing the participants, and an automatic summary is reread before it is circulated. An AI struggles to tell who said what when three people talk at once, which, let's admit it, also happens to humans.

Excel, figures and tables

Assistants can write a formula, explain a pivot table or analyse a sales file. ChatGPT and Claude do calculations by running code, which is more reliable than a calculation "in their head". Copilot works directly in Excel with the Copilot Business licence. Two rules: anonymise files before sending them to a free tool, and always check a total that will feed a decision.

Images and visuals

To illustrate a LinkedIn post or a poster, ChatGPT and Google (AI Pro) generate good-quality images. Three precautions: no image of a real person, no logo of another brand, and a mention when useful. The European AI regulation (AI Act) imposes transparency obligations on generated content: see our article AI Act: a year's grace, not a holiday.

Programming and automating

This is the field that changed most in 2026: "coding agents" write, test and fix programs on their own. Claude Pro includes Claude Code, ChatGPT includes Codex, Mistral Pro includes the Vibe coding agent. For a freelancer who wants to automate a repetitive task (renaming files, extracting data from PDF invoices, linking two pieces of software), it is spectacular. Before connecting AI to your accounting or CRM, first read Connecting AI to your accounts: gains and risks.

Truly confidential documents

Medical records, HR data, the professional secrecy of a lawyer or an accountant: even a business contract may not satisfy your professional ethics. The solution is then an AI installed on your premises, on an office PC, with no connection to an outside service. "Open" models, which you download and run yourself, have made great progress. The analyst firm Gartner even estimates that 90% of the steps in tasks handed to AI agents in companies could run on these models by 2030, for a fraction of the cost of frontier models. An analyst forecast, not an observation, but the trend is clear.

Hardware, budget and limits: see An AI installed at your premises, without the cloud and our comparison of hardware for local AI.

Three profiles, three set-ups

The beginner: 0 euros

You have never used AI, or hardly. Start with the tool already there: Copilot Chat if you have business Microsoft 365, Gemini if you are on Google Workspace. Otherwise, a free version of ChatGPT, Gemini, Claude or Mistral. Open two side by side for a week, ask them the same questions about your real working day, and keep the one that suits you best. Switch off training in the settings, put no customer names in, and you take no risk. For the step-by-step method (choosing a pilot, measuring, training), see AI in SMEs: where to start?.

The freelancer: 15 to 25 dollars a month

You use AI every day. A single subscription is almost always enough: ChatGPT Plus for versatility, Claude Pro for long-form writing and analysis, Mistral Pro for a cheaper European vendor, Google AI Pro if you live in Gmail (it also includes 5 TB of storage). Paying for three subscriptions "to compare" is a common expense and rarely useful. Forums are full of it.

The SME with 5 to 30 people: 20 to 35 dollars per person per month

Here, the question is no longer just the tool, it is the organisation: who uses what, with which data, and who pays. Three logical paths:

• You are on Microsoft 365: Copilot Chat for everyone (included), Copilot Business for the staff who spend their day in Outlook, Word and Excel.
• You are on Google Workspace: Gemini is already included in the business plans. Start by enabling it and training the team.
• You want an assistant independent of your email suite: ChatGPT Business, Claude Team or Mistral Team, with one account per person and an admin console.

In all three cases, you need an acceptable use policy: your employees are probably already using AI, with their personal accounts. Our article Shadow AI: govern AI with a simple policy offers a template, and the checklist before deployment covers the obligations (staff, GDPR, AI Act).

The demanding professional

Developer, analyst, consultant who delegates tasks of several hours to agents: the 100 or 200 dollar plans (ChatGPT Pro, Claude Max or Ultra, Google AI Ultra) are justified when you hit the limits of the standard subscription every day. Beyond that, API access, billed by usage, lets you build your own tools. And this is where open models, hosted on site or in a European cloud, become attractive for repetitive tasks.

The options compared

Measure the result. A survey reported in the French press in October 2026 indicated that only a minority of user companies reported concrete economic gains. The difference is rarely made by the tool chosen, almost always by the uses: which tasks, who does them, how they are checked.

True or false: five common beliefs

"The paid version protects my data." False, not by default. Personal subscriptions to ChatGPT, Mistral and Claude leave training switched on until you turn it off. What protects you is a business plan with a contract, or the right setting.

"A European AI is automatically GDPR-compliant." Not automatically. Mistral is a French vendor, which simplifies jurisdiction questions, but its personal plans also use your conversations for training by default. The vendor's origin matters, and so do the settings and the contract.

"I need the most powerful model." Rarely. For an email, a summary or a translation, entry-level models are more than enough. Gartner even advises companies to keep the most expensive models for complex tasks.

"An AI PC makes ChatGPT faster." False. ChatGPT, Copilot, Gemini and Claude run in the cloud. See AI PC: should you replace your laptop?.

"AI no longer makes mistakes." False. It makes fewer, but with as much confidence as ever. A figure, a date, a legal reference: check them at the source.

Example: an accounting firm with six people

Illustrative case. An accounting firm with six people works on Microsoft 365. Two staff members already use free ChatGPT on their phones to rephrase emails, and one of them has already pasted an extract of a balance sheet "to check a calculation".

We would start by enabling Copilot Chat for everyone (no extra cost, enterprise data protection) and writing a one-page policy: what goes into Copilot, what never does. The two people who prepare files in Excel test Copilot Business for a month. For translations into Dutch, DeepL Pro. For the most sensitive files, nothing leaves the office: if the need is confirmed, a local AI on a dedicated workstation. Budget: around forty dollars a month for the two Copilot licences, plus DeepL, and half a day of training.

What has changed in recent months

The market moves fast. A few landmarks from autumn 2026:

• ChatGPT now offers three Pro plans (100, 200 and 500 dollars) above Plus, and its Business plan costs 20 dollars per user with annual billing.
• Mistral's Le Chat is now called Vibe, an assistant that combines conversation, work tasks and programming, with the same Free, Pro and Team plans.
• Claude has renamed its plans: Max at 100 dollars, Ultra at 200 dollars.
• Google offers an entry-level AI Plus plan, and AI Pro now includes 5 TB of storage.
• Microsoft charges 21 dollars per user for Copilot Business, with a promotion at 18 dollars until 31 December 2026 (annual commitment).
• In Belgium, Combell, delaware and imec announced a "European Sovereign AI Cloud": computing power in European data centres, to keep control of costs and data. One to watch for companies that want AI hosted in Europe.
• AI Act: since July 2026, Article 4 requires companies that use AI to take measures to develop their staff's AI literacy. Training the team is therefore no longer just a good idea.

Test yourself

At ExsIT

ExsIT has no commercial ties with the vendors mentioned. The first contact is free: we listen to your needs, then propose an action plan. Choosing tools according to your real uses, privacy settings, acceptable use policy, team training, and local AI when your data must not leave the office.

For your IT person

• Training and contracts: OpenAI (Business, Enterprise, API), Anthropic (Team, Enterprise, API), Google Workspace and Microsoft 365 (Copilot Chat with enterprise data protection, Copilot Business) do not train their models on business customers' content by default. Mistral: Vibe Team with organisation-level opt-out, Enterprise excluded by default. Check the DPA, data location (EU Data Boundary at Microsoft) and retention period.
• Personal accounts: block or govern the use of consumer assistants with company data (policy, Purview DLP for Copilot, DNS filtering if needed). Switch off "Improve the model for everyone" (ChatGPT), "Keep Activity" (Gemini) and the training setting (Claude, Mistral) on any tolerated personal accounts.
• User feedback: at several vendors, a thumbs up or down can make the conversation usable for training, even after opting out. Mention it in the policy.
• Connections to software: MCP connectors and agents inherit the user's rights. Start read-only, keep logs.
• Open models and local AI: Ollama, LM Studio or llama.cpp on a machine with plenty of graphics or unified memory; route repetitive tasks to an open model and complex tasks to a frontier model under contract. Details in our technical guide to local AI.
• API: billed per token, 30-day retention by default at several providers, zero retention possible on request and under conditions.

This article was prepared with the help of an AI, then checked and proofread by us. One of the tools compared (Claude) was used to write it: we applied the same criteria to all.

Sources

Acerta. (2026). L'IA, nouvel allié des entreprises belges, avec 1 sur 3 qui s'y engage activement. Forum for the Future. https://blog.forumforthefuture.be/fr/article/lia-nouvel-allie-des-entreprises-belges-avec-1-sur-3-qui-sy-engage-activement-.../29797

Anthropic. (n.d.-a). Is my data used for model training? Anthropic Privacy Center. https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training

Anthropic. (n.d.-b). Plans & pricing. Claude. https://claude.com/pricing

Cryptobriefing. (2026). Gartner predicts open-weight models will run 90% of enterprise agent workflow steps by 2030. https://cryptobriefing.com/gartner-open-weight-models-enterprise-2030/

DeepL. (n.d.). Traductions IA plus vraies que nature. https://www.deepl.com/fr/translator-ppc

European Union. (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence. Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2024/1689/oj

Google. (n.d.-a). Gemini Apps Privacy Hub. Gemini Apps Help. https://support.google.com/gemini/answer/13594961

Google. (n.d.-b). Generative AI in Google Workspace Privacy Hub. Google Workspace Help. https://support.google.com/a/answer/15706919

Google. (n.d.-c). Google AI Pro & Ultra. Gemini. https://gemini.google/subscriptions/

IT-Connect. (2026, 3 July). Google AI Pro : le stockage passe de 2 To à 5 To pour le même prix. https://www.it-connect.fr/google-ai-pro-le-stockage-passe-de-2-to-a-5-to-pour-le-meme-prix/

ITdaily. (2026). European cloud with a strong Belgian flavour: Combell, delaware and imec join forces. https://itdaily.com/blogs/cloud/combell-delaware-imec-european-cloud/

Journal du Palais. (2026). IA : l'adoption s'accélère, les gains économiques se font attendre. https://journal-du-palais.fr/au-sommaire/entreprises/ia-l-adoption-s-accelere-les-gains-economiques-se-font-attendre

Microsoft. (n.d.-a). Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat. Microsoft Learn. https://learn.microsoft.com/en-us/microsoft-365/copilot/enterprise-data-protection

Microsoft. (n.d.-b). Microsoft 365 Personnel. Microsoft Store Belgium. https://www.microsoft.com/fr-be/microsoft-365/p/publisher/cfq7ttc0k7c3

Microsoft. (n.d.-c). Microsoft Copilot for small business. https://www.microsoft.com/en-us/microsoft-365-copilot/business

Mistral AI. (n.d.-a). Do you use my user data to train your Artificial Intelligence models? Mistral Help Center. https://help.mistral.ai/en/articles/347617-do-you-use-my-user-data-to-train-your-artificial-intelligence-models

Mistral AI. (n.d.-b). Pricing. https://mistral.ai/pricing/

Mistral AI. (n.d.-c). Vibe gets to work. https://mistral.ai/news/vibe-agent/

OpenAI. (n.d.-a). About ChatGPT Pro tiers. OpenAI Help Center. https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers

OpenAI. (n.d.-b). ChatGPT Business: General FAQ. OpenAI Help Center. https://help.openai.com/en/articles/8542115-chatgpt-business-general-faq

OpenAI. (n.d.-c). How your data is used to improve model performance. OpenAI Help Center. https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance

OpenAI. (n.d.-d). Introducing ChatGPT Go, now available worldwide. https://openai.com/index/introducing-chatgpt-go/

OpenAI. (n.d.-e). Pricing. ChatGPT. https://learn.chatgpt.com/docs/pricing

Perplexity. (2026). AI costs: what teams pay to use AI tools. https://www.perplexity.ai/hub/blog/ai-costs

Statbel. (2025). Artificial intelligence is gaining ground in Belgian enterprises. https://statbel.fgov.be/en/themes/enterprises/ict-and-e-commerce-enterprises