AI agents need an SME job description

The first useful question about an AI agent in a small company is not whether it is autonomous. It is whether anyone has written its job down.

The OECD’s 2026 D4SME survey, published in Empowering SMEs in the age of AI, gives the operational clue. SME adoption of AI tools is rising quickly, most use is still off-the-shelf, and some firms are experimenting with more tailored applications, including AI agents. The same summary warns that strategic, targeted and secure integration inside business operations remains uneven.

That is the whole agent problem in one sentence. The software can now act across tools. The company often has not decided what action means.

Off-the-shelf is the default

For most SMEs, AI adoption begins without a project. A writing assistant appears in the office suite. A meeting recorder enters the video tool. The CRM offers generated replies. The accounting package adds document extraction. Staff try a general assistant because it is already in the browser.

This is not a failure. Off-the-shelf AI is often the cheapest way to learn what a team actually needs. It exposes repetitive drafting, search, translation, classification and summarisation work. It also exposes where the company’s data is messy, where approval is oral, and where a process depends on one person remembering the exception.

The mistake is to treat that discovery layer as a production system.

An agent raises the stakes because it does not merely suggest text. It may open a ticket, move a record, draft a customer reply, update a spreadsheet, trigger a workflow or call another tool. Even when every step is reversible, the operating question changes from “was the answer plausible?” to “was the action allowed?”

A job description beats an autonomy slider

The cleanest control is boring: write the job description before choosing the agent.

Field	What the SME should write
Task	The exact work the agent performs, in one sentence.
Inputs	The files, systems, forms or messages it may read.
Actions	The fields it may change and the messages it may prepare.
Forbidden actions	Anything it must never send, delete, approve or decide.
Reviewer	The person or role that checks exceptions.
Evidence	The log, note or record left after each run.
Stop rule	The error rate, complaint, missing data or confidence threshold that pauses use.

Source: OECD SME AI adoption publications and European Commission AI literacy guidance. Last verified 2026-06-24.

This document does not need to become a governance framework. One page is enough for a narrow workflow. If the task cannot be described on one page, the company is probably not ready to automate it with an agent.

The job description also prevents the most common false promise: “the agent will handle the process.” No agent handles a process in the abstract. It handles a slice of work under constraints. The constraints are the product.

The risky middle is where SMEs live

Large organisations can surround agents with security teams, procurement reviews, model risk committees, internal audit and platform engineering. Very small firms may use only a general assistant and keep every action manual. The risky middle is the 20-to-150-person company that has enough software to connect, but not enough governance to absorb mistakes quietly.

That company often has a CRM, email, spreadsheets, shared drives, an accounting tool, a website form and a few internal databases. The agent pitch is obvious: connect the fragments. The operating reality is less elegant. The customer name is spelled three ways. The order status is not authoritative. The warranty rule changed but the old PDF is still in the folder. The sales team uses one field for two meanings. The service team knows which exceptions matter, but never wrote them down.

An agent can accelerate this mess. It can also make it visible.

The useful first project is therefore not the broadest agent. It is the narrowest one that forces the company to name the work. For example: classify inbound support messages into five queues, prepare but not send a reply, attach the source message, and escalate anything containing a refund, legal threat or product-safety claim.

That is small enough to test. It is also real enough to teach the company where its rules are missing.

AI literacy is now operational

The Commission’s AI literacy Q&A is easy to misread as a training story. For SMEs, it is more practical than that. Article 4 asks providers and deployers to ensure a sufficient level of AI literacy, taking into account the staff, the context and the persons affected by the systems.

In agent terms, literacy means knowing what the system is allowed to do, where it gets information, what can go wrong, when to intervene and how to read the trace it leaves. A generic “AI awareness” session will not teach that for a warehouse inventory assistant, a customer-support triage agent and a finance document extractor at the same level of detail.

The training follows the job description. If the agent drafts customer replies, staff need to know the escalation categories and the banned claims. If it updates stock records, they need to know which source wins when two systems disagree. If it prepares invoices, they need to know where human approval starts.

This is why agent adoption belongs close to operations. The people who understand the exception need to shape the rule.

Europe’s policy push does not remove the local work

The 2026 State of the Digital Decade package pushes support for businesses, especially SMEs, to adopt advanced digital solutions through skills, infrastructure and innovation ecosystems. The Apply AI Strategy similarly frames adoption as a competitiveness issue, with European Digital Innovation Hubs, AI Factories, testing facilities and sandboxes in the support machinery.

That infrastructure matters. It can reduce confusion, open access to expertise and make experimentation less lonely.

But it cannot write the local job description. The public layer can help a company understand available support. It cannot know which customer messages should never be answered automatically, which spreadsheet is trusted, which manager accepts risk, or which exception breaks the workflow.

SME AI adoption will therefore split into two kinds of progress. The visible kind is tool uptake: more assistants, more embedded features, more agent demos. The valuable kind is operating clarity: fewer unnamed rules, fewer orphaned spreadsheets, better escalation, clearer logs and less automation theatre.

Start where an error has a name

The safest first agent is not the one with the most impressive demo. It is the one where an error has a name and a route.

If the agent misclassifies a lead, who notices? If it drafts the wrong reply, who approves before sending? If it cannot find a source document, what does it do instead? If the same customer appears twice, which record wins? If the model changes behaviour, which test catches it?

Those questions sound defensive. They are actually what make autonomy useful. A company that can answer them has already done half the work. A company that cannot answer them is not blocked by AI capability. It is blocked by process ambiguity.

The OECD survey points to a practical 2026 reality: SMEs are adopting AI quickly, but integration remains the hard part. Agents do not change that. They make the integration problem sharper.

Before giving an agent a tool key, give it a job description. The first version can be short, plain and imperfect. It only has to be specific enough that a human can say, after one run, whether the agent did the job or merely looked busy.