AI agents for SMEs: what actually works today (and what doesn't yet)

Short answer: agents that retrieve information, summarise, and carry out a fixed, repeated task already work reliably today — provided a human stays in the loop for anything with consequences. Agents that decide and act fully on their own don't yet work reliably enough for most SME use cases. The difference is rarely the language model itself — it's how much autonomy you hand it, and whether there's a brake.

That distinction is exactly why "do AI agents actually work?" doesn't deserve a yes-or-no answer. An agent that summarises your inbox works great. An agent that decides on its own which customer gets a discount is a risk waiting for an incident. This article lays out what we see working in practice, what doesn't yet, and how an SME can get started without falling into either trap.

What makes an agent different from a chatbot

A chatbot answers a question within one conversation and is done. An agent is given a task, not a question: it decides for itself which steps are needed, uses tools (search, reading a document, calling an API, running a calculation) and works towards a result across multiple steps. That ability to choose its own path is exactly what makes an agent useful — and exactly where the risk starts the moment nobody is watching.

What already works today

Four categories of tasks are mature enough in practice to rely on, provided they're set up correctly.

1. Searching documents and answering with a source. An agent that searches contracts, manuals or internal knowledge and answers with a reference to the original document is the most mature application there is — provided the citation is actually verified rather than just sounding convincing. We explain that difference in RAG explained: how to build an AI that doesn't lie.

2. Fixed, repeated workflows. Compiling a report, moving data between systems, running a daily check and emailing a summary — these are tasks with a predictable start and end, where an agent barely has to "decide" anything. The risk is low here because the path is already fixed; the agent executes, it doesn't improvise.

3. Monitoring that proposes instead of executes. An agent that reads signals, forms a diagnosis and prepares a proposal for a human to approve works well — precisely because it isn't allowed to mutate anything itself. We cover that separation between reasoning and acting in detail in why your AI administrator shouldn't be allowed to do anything, with two real examples of what goes wrong once that separation is missing.

4. Early-stage, bounded research work. Gathering public information, compiling a shortlist of candidates, forming an initial assessment based on visible characteristics — work that used to take a morning can be prepared by an agent in minutes. The output is a draft a human reviews, not a final decision.

What doesn't (fully) work yet

Three patterns consistently go wrong, and they all share the same root cause: too much autonomy, not enough brake.

Fully autonomous decision-making. An agent that performs a consequential action without a checkpoint — restarting a system, emailing a customer, making an expense — lacks the human judgement that catches edge cases. Not because the model reasons poorly, but because a plausible diagnosis is not the same as a correct one, and the difference only becomes visible once something has already happened.

Memory that stops at the session. An agent that starts from scratch every morning repeats mistakes the previous session already solved, and costs double: once to build the knowledge, once to reconstruct it. Without a separate memory layer, a "team of agents" remains, in practice, a group of disconnected conversations. More on this in why your AI agents start over every session.

Costs nobody can explain. Agent work drags context, history and retrieved documents along with every step — and that input grows the longer a task runs. Without per-call logging, you only find out at the end of the month that it was expensive, never why. We explain how to make that visible in what does an AI agent actually cost? (Dutch).

The question isn't "can the agent do this", but "what happens in the one percent of cases where it gets it wrong — and who notices".

How an SME gets started

The sequence that works in practice is almost always the same. Start with one well-defined task that's currently done manually and whose outcome is verifiable — not "automate our customer contact", but "summarise incoming invoices into a fixed format". Put a human in the approval loop for anything beyond preparing information. Measure the cost per task from day one, not only once the bill disappoints. And only expand to the next task once the first has run for a while without surprises.

That's less exciting than the promises on a product page, and exactly for that reason more reliable. An agent that does three things well and refuses the rest is more useful than one that attempts everything and occasionally fails in a way nobody catches in time.

Conclusion

AI agents aren't an all-or-nothing technology. For structured, repeated and verifiable work, they're already a real alternative to manual effort today. For autonomous decision-making with consequences, they're not there yet — and the businesses that trust them there anyway usually find out through an incident rather than an evaluation. Start small, keep a human in the loop, and measure what it costs. The rest follows.