Before you switch on that AI agent, read this

Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from less than 5% a year earlier. Most SME managers will meet them as a renewal email (a new "AI agent" tier, a higher price) or as a general agent, like the dots OpenAI launched in September, plugged into the tools they already use.

Either way, the question that matters rarely gets answered: which of your data the agent works on, and who keeps what it learns. This guide covers how to tell a real agent from a rebranded tool, the questions to ask before signing, and what a Belgian SME owes its own customers once an agent talks to them.

Your AI agent doesn't sell you intelligence. It rents you back your own data.

What is an AI agent in business software? An AI agent in business software is a feature that completes tasks on your data, instead of only suggesting them. Vendors call it "vertical" when it is built for one trade or workflow, such as legal review, customer support or logistics planning. It comes preloaded with that sector's vocabulary, rules and integrations, and it acts on your data to complete tasks. The model underneath is usually a general one from a major provider; the difference lies in the sector knowledge and the access to your systems.

If the model is the same one everybody uses, the vendor's real asset is the domain depth it has accumulated: rules, edge cases, corrections. Part of that depth comes from its customers. When your team feeds the agent your files and fixes its mistakes, it is training the vendor's product as much as using it.

What an SME hands over without noticing

Three things leave the building when an agent goes live:

• Data: documents, messages, customer records the agent reads to do its work.
• Process knowledge: the way you route a case, approve an order or answer a complaint.
• Corrections: every time someone edits the agent's output, that edit tells the vendor what "right" looks like in your trade.

The exposure is concrete in Belgium. According to Statbel's 2025 ICT survey, 34.5% of Belgian companies use AI, and the most common use is automatic analysis of written language (23% of firms). That is documents and messages: exactly where client data sits.

💡 The question to keep in mind: if you cancel in two years, what leaves with you? Your raw data, if the contract allows it. The intelligence built on top of it usually stays with the vendor.

Most "AI agents" on the market are not agents

Gartner has a name for the rebranding wave: agent washing. In its June 2025 analysis, it estimates that only about 130 of the thousands of agentic AI vendors are real. The rest are assistants, chatbots or automation scripts relabelled. The same analysis predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear value or weak risk controls.

For a buyer, the difference shows up in four places.

1. It completes a task, it doesn't just suggest one

An assistant drafts a reply. An agent sends it, files the case and updates the customer record, within limits you set.

2. It works inside your systems, with permissions you control

A real agent connects to your CRM, ERP or mailbox through defined access rights. If the vendor's answer to "how does it reach our data?" is "you upload a file", it is a chatbot with a new name.

3. It shows what it did

Every action should leave a trace you can read: what the agent looked at, changed, and why. OpenAI's dots, for instance, ship with an activity view and approval rules. Without that, nobody can audit a mistake.

4. It handles your exceptions, not only the sector average

Vertical means preloaded with sector rules. Your margin often sits in the cases where you deviate from them. Test those first.

⚠️ Watch out: the demo on the vendor's data. A polished demo on sample files proves the interface works. Insist on one session with your own documents and your hardest real case before any pricing discussion.

Seven questions to ask your vendor before you switch the agent on

Ask them in writing, and keep the answers with the contract.

Who owns the data an AI agent works on?

You do, as the controller of that data. What you may not own is what is derived from it: configuration, rules learned from your corrections, logs. General agents are no exception: OpenAI says it doesn't train on Business content by default, yet what a dot learns about you stays on its platform. Settle it in the contract before signing.

👉 Where the answers point. If most answers fall in the right-hand column and the process involved is the one that makes your margin, the question is no longer which vendor to pick. It is whether that part should run on a tool you own, which we cover in What is custom software?

What a Belgian SME owes its customers once an agent talks to them

Buying the agent does not transfer the responsibility. Two sets of rules apply to the company that deploys it.

GDPR: the vendor is your processor

When an agent handles personal data on your behalf, the vendor acts as a processor under the GDPR, and Article 28 requires a written contract that sets what it may do with that data. The seven questions above are, for the most part, the commercial version of that contract.

AI Act: disclosure applies since 2 August 2026

The Digital Omnibus pushed back the high-risk rules, but not the transparency obligations of Article 50, which include telling people when they are interacting with an AI system. Building that disclosure sits mainly with the provider. Your customers, however, will hold you to it. If the agent answers your support inbox, check that it says so.

Agents used for hiring, credit or access to essential services fall under the high-risk regime, now set to apply from 2 December 2027. That is a delay, not an exemption.

The honest limits of this checklist

An SME rarely has the leverage to rewrite a large vendor's terms. In that case, the seven questions still sort vendors: pick the one whose standard contract answers best, and favour exportability over features. And the checklist cannot predict an acquisition or a price hike, only limit what you lose when one happens. This article is not legal advice; for a contract that touches personal data at scale, have it reviewed.

🔍 Ask your team first: the Conseil supérieur de l'emploi found that one Belgian worker in three already used generative AI at work in 2025. Some of the data you are about to protect in a contract may already be in a free chat tool.

Four mistakes SMEs make when buying an AI agent

1. Accepting the AI tier at renewal without a new review

An upgrade that adds an agent changes what the vendor does with your data, so it deserves the review a new contract would get.

2. Connecting everything on day one

Start with one mailbox or one case type; broad access granted for a pilot rarely gets narrowed afterwards.

3. Judging the agent on its best answers

The cost of an agent shows up in its errors, so measure how often your team corrects it in the first month, not how good the demo looked.

4. Forgetting the exit before the entry

Ask for a test export in the first week; if it takes a ticket and a fortnight now, it will take longer when you actually leave.

💡 What the four share: each one treats the agent as a feature of the software, when it is really a new party with access to your data and your customers.

How we help SMEs evaluate an AI agent at Nightborn

Step 1: map what the agent would touch

In a scoping workshop of four hours at most, we list the systems, data fields and customer touchpoints the agent would access, and mark what must never be shared with an AI. Deliverable: a data map and the success criteria the agent has to meet. The format is on our Product Discovery page.

Step 2: run the seven questions against the shortlist

We put the vendors' written answers side by side, test each agent on your hardest real case, and check the export before anything is signed.

Step 3: integrate with limits, or build the core

If a vendor passes, we connect it to your tools with scoped access and logging, without touching their configuration; see our AI Integration page. If none does and the process is core to your margin, we scope what it would take to run it on a tool you own.

📌 Checklist before you sign: data scope in writing, no training on your data by default, a tested export, EU processing with named subprocessors, AI disclosure to customers, and a cap on usage-based pricing.

AI agents: buy the access, keep the intelligence

What to remember:

• AI agents in business software run on general models; the vendor's value is the domain depth, partly built from customers' data.
• Agent washing is widespread: Gartner counts about 130 real agentic vendors out of thousands.
• Seven written questions separate a vendor that rents you access from one that keeps what you teach it.
• Responsibility stays with you: GDPR processor rules and AI Act disclosure apply to the company that deploys the agent.

The practical step for this quarter: before the next renewal email lands, list the tools in your stack that have added an "AI agent" tier and send each vendor the seven questions. The answers will tell you which contracts to sign, which to narrow, and which process deserves a closer look. If you want a second pair of eyes on those answers, a scoping workshop is where we start.