Before Talking About AI Agents, Make an Inventory

Your company already uses AI.

The problem is that probably nobody has the map.

It is in a browser where someone summarises emails with ChatGPT. In Microsoft 365 or Google Workspace. In the CRM. In marketing tools. In ecommerce plugins. In design, translation, transcription, document analysis, customer service or proposal preparation.

Often it did not arrive as an "AI project". It arrived as a new feature inside a tool already being paid for, or as a quick test by someone trying to save time.

That is not the problem. AI can save real time.

But the conversation is now moving toward agents: systems that do not just answer, but prepare tasks, consult data, fill fields, propose actions, open incidents, draft replies, update records or execute parts of a workflow.

Then the question changes. It is no longer enough to ask: which AI tool should we buy?

The earlier question matters more: where do we use AI today, what data does it touch, what can it do, who reviews it and what should it never decide without approval?

AI Has Already Entered Through Many Small Doors

The European Commission's 2026 State of the Digital Decade package places AI use at almost 20% of EU enterprises and says adoption rose by 48% in 2025. It also recognises the barriers SMEs still face: skills, access to data, infrastructure and resources.

That combination describes reality well. Adoption is moving faster than internal capacity to govern it.

One person tests a tool. Another activates a new feature. Marketing automates text. Administration copies information into an assistant. Support summarises conversations. Sales asks for help drafting a proposal. Management uses AI to organise information before a decision.

The problem is not that this happens. The problem is that nobody knows what is happening as a whole.

The company thinks it is "testing AI". In reality, it is already creating a layer of AI-assisted work without a common map.

Regulation Points Toward Management

This is not legal advice. Nor does it mean every AI use in an SME is high risk. The EU AI Act works with risk levels, and many everyday uses will be minimal or limited risk.

But if you read regulation as a management signal, the direction is clear: literacy, transparency, traceability, documentation, human oversight and responsibility.

Article 4 of the European regulation refers to AI literacy for people operating or using AI systems. The Commission summarises high-risk system obligations that include activity logs, documentation, clear information, human oversight, robustness, cybersecurity and accuracy. Article 50 introduces transparency obligations for certain systems.

The practical conclusion for an SME is not "panic". It is this: you cannot train, supervise, explain, record or correct what you do not know exists.

An Inventory Is Not Bureaucracy

An AI inventory should answer at least these questions:

• which tools are used and for which concrete task;
• which data they read, copy, summarise, transform or generate;
• whether they can affect a customer, employee, supplier, patient or economic decision;
• whether they inform, prepare, recommend, update, send or decide;
• who reviews, corrects, approves or stops the output;
• where input, output, review and decision are recorded.

If AI only summarises public documentation, the control can be light. If it prepares customer replies, tone, sources, permissions and approval matter. If it classifies candidates, patients, risks or complaints, it enters a more delicate area. If it updates CRM, ERP or invoicing data, it touches business systems.

Prepare, Yes. Commit the Company, Not Without Approval.

This should be a first rule for many SMEs testing agents: prepare, yes; commit the company, not without approval.

An agent can gather context, summarise, detect missing information, classify, draft, suggest the next step or prepare an exception queue. But committing the company is different.

Committing the company means sending a sensitive reply, accepting conditions, changing prices, updating critical data, rejecting a request, approving a payment, modifying an invoice, closing an incident or making a decision that affects a person.

There must be a clear boundary. Autonomy is earned with evidence; it is not declared in a demo.

NIST's AI Risk Management Framework insists on responsibility, transparency and defined human-AI roles. The AEPD, from a data-protection perspective, reminds organisations that auditing AI processing means understanding context, purpose, data, responsibilities, traceability and error mechanisms.

In SME language: before automating, write down who watches, who decides and who can stop.

Three Examples Where the Inventory Changes the Decision

Customer inbox. An agent can read a query, identify the customer, find history, classify urgency, detect missing data and prepare a draft reply. The first safe version probably does not answer alone. It prepares a reviewable response.

Budgets and sales administration. An agent can gather opportunity data, check CRM gaps, prepare a base proposal, warn of risks and suggest the next step. But price, discount, delivery commitment and terms are commercial decisions.

Invoices, orders and discrepancies. An agent can compare order, delivery note, invoice, customer data and internal rules. That can reduce administrative time and errors, but amounts, taxes, fiscal data and payments need clear limits.

The Inventory Also Finds Opportunities

There is a defensive side: reducing risk, preparing compliance, protecting data and knowing who is responsible.

But the offensive side may be more interesting. A good inventory shows where value is hiding.

If three people use AI to summarise customer queries, perhaps there is a support flow worth designing properly. If sales uses AI for proposals, perhaps the business needs templates, approved sources and a less fragile proposal workflow. If administration copies document data into generic assistants, perhaps controlled extraction and human review is a good first case.

OutSystems, in its 2026 AI development report, points to concerns around agent sprawl, complexity, technical debt and security. The solution is not to stop adoption. The solution is to make it visible.

The First Serious Agent Should Come From the Map

After the inventory, the business can choose a first workflow. Not the most spectacular one. The clearest one.

A good candidate usually hurts every week, has accessible data, produces a clear output, can be measured and allows a simple human boundary before any external, economic or sensitive impact.

Then it makes sense to design a first agent: what it can see, what it can prepare, what it can propose, what it must record and which person approves.

That approach is less showy than promising a company full of agents. It is much more useful, because it turns AI into better-prepared work instead of another layer of disorder.

Conclusion: Inventory First, Agents Second

The question is not whether an SME should use AI. Many already do.

The question is whether they use it clearly enough to turn it into better work.

The inventory is not the end of the road. It is the starting point for deciding where to invest, what to organise and which first agent can be tested without losing control.

What AI is used. What data it touches. What it can prepare. What it cannot decide. Who reviews. What gets recorded.

Before talking about AI agents, make an inventory.

Sources Used

• European Commission: AI Act regulatory framework
• EUR-Lex: Regulation (EU) 2024/1689
• European Commission: 2026 State of the Digital Decade package
• NIST: AI Risk Management Framework 1.0
• AEPD: AI processing audit requirements
• Deloitte: The agentic reality check
• OutSystems: 2026 State of AI Development