AI in SMEs: playtime is over
Updated on 19 November 2026.
Why now
In 2025, 34.5% of Belgian businesses already used at least one AI technology, according to Statbel. In many SMEs this use crept in from the bottom: one employee tries ChatGPT, another takes out a subscription, a third connects a tool to the mailbox. At the end of 2026, three signals show that this trial phase is ending:
- clients are asking questions: in professional services, buyers increasingly demand to know whether and how AI is used, and they are renegotiating prices;
- bills are surprising people: even large companies, with dedicated teams, are overrunning their AI budget;
- "AI agents" are being sold everywhere, including for tasks that a simple automation does better and more cheaply.
These three subjects are linked: what you promise your clients depends on what you use, and what you use determines what you pay. Let us take them in order.
Decision 1: what you tell your clients
What buyers are starting to demand
The 2026 barometer of professional-services purchasing, presented on 18 November by Decision Achats (a French survey; the article gives neither the sample nor the method), shows the trend:
- 58% of the buyers surveyed ask their suppliers for a statement of AI use;
- 31% have already added an AI clause to their contracts, and nearly two in three have done so or are preparing to;
- 36% have already revised their pricing terms;
- nine in ten want to know who actually delivers the service.
In tenders, some procurement departments even ask suppliers to commit to a percentage of productivity gain from AI. A 10 to 15% reduction in workload is mentioned as a basis for negotiation, to be checked against results.
It is a French survey, focused on large buyers. But if you are a consultant, engineering firm, accountant, translator, agency or developer working for large companies or public bodies, the question will reach you too, if it has not already.
What an "AI clause" contains
An AI clause is nothing exotic. It answers five questions your client is asking anyway:
- Which tools? The names of the AI services used, and whether they are business plans or personal accounts.
- Which data? What you put into them: public data, anonymised data, or the client's confidential documents.
- Where does its data go? If you process personal data for your client and the AI provider processes it on your behalf, it is in principle a processor under the GDPR. Article 28 then requires the client's prior written authorisation, specific or general, before engaging it. A free or personal account, which may be used to train models, does not pass this test.
- Who validates? The name of the person who reviewed and approved the deliverable. This is what distinguishes professional work from generated text.
- Who owns the result? European copyright protects a human intellectual creation. For content largely produced by AI, protection is uncertain: the contract must state clearly what the client receives and what you guarantee.
The right answer: take the lead
Rather than filling in every client's questionnaire, write a short statement, one page at most, to attach to your proposals:
- the tools used, all in business versions, with no training on your data or your clients' data;
- the tasks where AI is involved (research, first draft, translation, proofreading) and those where it never is;
- the rule "no confidential client data in an unapproved tool";
- systematic human validation, with the name of the person responsible;
- a contact for any questions.
This page costs you an hour, it reassures and it sets you apart. It assumes your internal use is already framed: if not, start with an AI use policy and reread our advice on AI and confidential data.
And the price?
That is the real question behind the clauses: if AI saves you time, the client wants its share. Two pieces of advice:
- do not cut prices by default: time saved on a first draft is often lost again in checking, and your liability does not shrink. An AI can be wrong with confidence: we explained it here;
- sell a result rather than hours: when the scope and acceptance criteria are clear, a fixed price or a price per deliverable protects your margin better than time-based billing, which AI mechanically reduces. Time and materials remain suitable when the need evolves along the way.
Decision 2: what AI really costs you
Even large groups get it wrong
In September 2026 the Capgemini Research Institute surveyed 500 executives of companies with more than one billion dollars in revenue, in ten countries. Results published on 12 November:
- 59% say their AI spending was significantly higher than forecast over the past year, with an average overrun of 24%;
- half admit that AI spending is approved without sufficient visibility of the total cost;
- the most underestimated costs are not technical: human oversight, reworking generated content, governance;
- more than three executives in five want to move to a multi-model strategy: choosing the model by task, and cheaper models when they are good enough.
These are very large companies, not SMEs. But the causes of the overrun are the same at your scale, only smaller.
The three cost items of an SME
| Item | Example | The trap |
|---|---|---|
| Subscriptions | "Team" plans from ChatGPT, Claude, Copilot, Gemini or Mistral: around $20 to $25 per user per month (prices checked on 10 November 2026, see our comparison) | Individual subscriptions that pile up, paid on different cards, with nobody holding the list. Some are consumer plans that do not protect your data. |
| Usage (API) | An automation or agent that calls a model for every email, document or ticket | A variable bill that climbs with volume or a badly tuned loop. A top-end model used for a simple task. |
| Human time | Reviewing, correcting, checking figures and sources | The most forgotten item, and often the heaviest. If reviewing takes as long as writing, the tool earns nothing. |
Four rules to stay in control
- Take stock: who uses what, paid how, with which data. An hour with the company credit card statement is often enough.
- Consolidate on a single business plan, managed by the company, rather than five individual subscriptions.
- Cap usage: API platforms let you set a monthly budget or alert. No API key without a cap.
- Measure what it brings in: for two or three specific tasks, note the time before and after, review included. If the gain is not visible after three months, stop or change tools.
The choice of model matters too: to sort emails or extract a date, a small model is enough and costs a fraction of a top-end model. An open model installed on your premises can even remove the usage cost: we covered this in our article on local AI.
Decision 3: simple automation or AI agent?
The deciding question
An AI agent is a program that uses a language model to understand a request, decide what to do and act in your tools (read an email, search the management software, prepare a reply). A rule-based automation (also called RPA or an automated flow) always performs the same steps, without understanding anything.
The deciding question: does the task require understanding, or only execution? Three criteria help to answer:
- the format of the input: if it always arrives in the same form (a file, a form, a date), a rule is enough. If it arrives as free text, an attachment or a photo, an agent becomes useful;
- the cost of an error: the higher it is, the more human validation is needed, whatever the tool;
- frequency: a rare task justifies neither a robot nor an agent. A task repeated hundreds of times a month deserves a tool.
Concrete examples
| Task | Right answer | Why |
|---|---|---|
| Chasing unpaid invoices at 30 days | Feature of your accounting software, or a simple automation | Fixed date, fixed text template: no understanding needed. |
| Entering invoices from your Belgian suppliers | Peppol e-invoice | Since 1 January 2026, structured electronic invoicing is mandatory between Belgian businesses: it arrives directly readable by your software. No AI needed to read it. |
| Copying orders from a client portal into your software | Rule-based automation | Same format every time. |
| Sorting incoming emails (quote request, complaint, invoice, advertising) | AI agent, with human checks at first | Free text, varied wording. |
| Preparing a quote from a freely written request | AI agent that prepares, human who validates and sends | The request must be understood; a pricing error is expensive. |
| Monthly sales report | Automation, or simply a well-built spreadsheet | Structured data, repetitive calculation. |
In most SMEs, the best solution combines all three: an automation for what is repetitive, an agent for what requires understanding, and a person to validate whatever commits the business.
The tools, from simplest to most powerful
- The features of your software: reminders, bank reconciliation, alerts. Often already paid for and unused.
- Power Automate: included in Microsoft 365 Business subscriptions for standard connectors (Outlook, SharePoint, Teams, Excel…). So-called "premium" connectors and desktop automation are charged separately.
- n8n: an automation tool that can be installed on your own server. Its licence (Sustainable Use License) allows internal business use free of charge, but it is not an open-source licence in the strict sense.
- Make or Zapier: online services, easy to pick up, priced by volume.
- An AI agent connected to your software: powerful, but it acts with the rights you give it. Before connecting one, read what to check when connecting AI to your business software.
A rule of caution: never let an agent take an irreversible action without human validation, whether a payment, a message to a client or a deletion.
An illustrative case
Invented case, but built from common situations. A twelve-person engineering firm near Liège works for municipalities and two large industrial groups.
In October, one of the groups sends a questionnaire: AI tools used, data concerned, validation of deliverables. Nobody really knows how to answer. The company card statement reveals seven different AI subscriptions, including three personal accounts into which engineers had pasted client reports. An employee has also connected an agent to the "info@" mailbox to sort requests: the API bill has tripled in two months, because the agent uses the most expensive model for every newsletter received.
Within a week, the manager takes the three decisions. A single business subscription for the whole team, with no training on the data. A one-page statement attached to every proposal. Advertising and newsletters filtered by a simple rule before reaching the agent, which switches to a lighter model with a monthly cap. Sorting requests stays with the agent, but a person validates every quote before it is sent. The client's questionnaire is completed in twenty minutes.
In practice: three options
| Option | Cost | Time | Risk |
|---|---|---|---|
| Do nothing | €0 today | 0 | A client questionnaire without a credible answer, subscriptions piling up, client data in personal accounts, agents billing for every email. |
| Do it yourself | €0 (and often savings) | Half a day per decision: AI statement, inventory and consolidation of subscriptions, list of tasks to automate | Low for decisions 1 and 2. Decision 3 requires technical skills as soon as tools are connected together. |
| Get support | A few hours of service | One to two days, spread out | The lowest: statement reviewed, subscriptions consolidated and secured, caps in place, one or two useful automations put into service and documented. |
The first step costs nothing: this week, list every AI tool the company pays for, and those your team uses without paying.
For your IT person
What follows is for the person who manages your tools. For the rest of the team, the three decisions are enough.
- Accounts: a single business plan (ChatGPT Business, Claude Team, Copilot Business, Gemini Enterprise or equivalent), accounts created by the company, single sign-on if possible, training disabled or excluded by contract. Remove personal accounts used for work.
- Contracts: keep each AI provider's data processing agreement (DPA) and its list of sub-processors; this is what a client will ask for under Article 28 of the GDPR.
- API: one key per use, a monthly budget or alert on each account, call logging. Choose the smallest model that does the task properly, and filter upstream whatever does not need AI.
- Power Automate: stay on the standard connectors included in Microsoft 365; a premium connector triggers an additional per-user licence. Daily action quotas exist: check them for high-volume flows.
- Self-hosted n8n: do not expose it directly to the internet, keep it updated, protect the credentials it stores. See our external access audit.
- Agents: one service account per agent, with minimum rights; read-only by default; any action that commits the business goes through human validation.
Test yourself
Sources
Boulnois, E. (2026, November 13). Agent IA ou RPA : ce que les PME doivent savoir avant d'investir [in French]. Journal du Net. https://www.journaldunet.com/intelligence-artificielle/1557877-agent-ia-ou-rpa-ce-que-les-pme-doivent-savoir-avant-d-investir/
Brasseur, L. (2026, November 18). IA et prestations intellectuelles : les acheteurs reprennent la main sur les prix [in French]. Decision Achats. https://www.decision-achats.fr/familles-dachats-1300/prestations-intellectuelles-2293/ia-et-prestations-intellectuelles-les-acheteurs-reprennent-la-main-sur-les-prix-52861
Capgemini. (2026, November 12). La plupart des entreprises prévoient d'adopter une stratégie d'IA multi-modèles, alors que près de six sur dix ont subi des coûts liés à l'IA significativement supérieurs aux prévisions [in French]. https://www.capgemini.com/fr-fr/actualites/communiques-de-presse/la-plupart-des-entreprises-prevoient-dadopter-une-strategie-dia-multi-modeles-alors-que-pres-de-six-sur-dix-ont-subi-des-couts-lies-a-lia-significativement-superieurs-aux-p/
FPS Finance. (2025). Period of tolerance during the first three months of 2026. e-invoice. https://einvoice.belgium.be/en/news/period-tolerance-during-first-three-months-2026
Microsoft. (2026, November 10). Power Automate licensing FAQ. Microsoft Learn. https://learn.microsoft.com/en-us/power-platform/admin/power-automate-licensing/faqs
n8n. (n.d.). License FAQ. n8n Docs. https://docs.n8n.io/n8n-community-license/community-license/license-faq
Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation). (2016). Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2016/679/oj
Statbel. (2025). Artificial intelligence is gaining ground in Belgian enterprises. https://statbel.fgov.be/en/themes/enterprises/ict-and-e-commerce-enterprises