AI en automatisering
AI agents for SMEs: hype or growth accelerator?
· 3 min read · Oddbird
From lead follow-up to content production: where AI agents already save real time, where they mainly move work around, and how to pick a first use that does not fall over.
What you take away
- An agent is only useful for work that recurs often, has clear rules, and where a mistake is manageable.
- The gain rarely sits in the model and almost always in the connection to the systems you already use.
- Start with one process with a measurable time saving, not with a platform.
A lot is currently being sold under the heading of AI agent, and some of it is a chat window with an expensive name. At the same time we see uses at clients that save hours weekly and could not be done two years ago. The distinction can be made, and it has little to do with technology.
What separates an agent from a chatbot
A chatbot answers a question. An agent completes a task with several steps, may decide the order itself, and can touch things outside the conversation: your inbox, your calendar, your CRM. That last part is exactly where both the value and the risk sit.
| Chatbot | Agent | |
|---|---|---|
| Does | Give an answer | Finish a task |
| Steps | One | Several, in varying order |
| Touches | Only the conversation | Systems: mail, calendar, CRM, documents |
| A mistake costs | A wrong answer | A wrong action, sometimes at a client |
| Needs to succeed | A good knowledge source | Good connections and clear limits |
Where it already works
The uses that prove themselves share three things: the work recurs often, the rules can be written down, and a mistake is manageable and recoverable.
- Lead follow-up. Enriching an incoming enquiry with what is publicly findable, categorising it, and drafting a reply a human still approves.
- Preparing quotes. Producing a draft from earlier quotes and the enquiry that is eighty percent right.
- Preparing content. Gathering research, making an outline, writing channel variants. The writing itself stays human work if it has to carry your tone.
- Reporting. Pulling figures from several systems and turning them into a plain-language summary.
- Internal questions. Staff who have to search documentation nobody can find.
Where it mainly moves work around
To be honest about the other side. Some uses look impressive in a demo and mainly produce checking work in practice.
- Fully automated client communication with no human in between. The times it goes wrong cost more than the times it goes right return.
- Decisions with consequences: prices, discounts, commitments. Those want a human signature.
- Work that occurs once a month. You never earn back the time you put into setting it up.
- Anything where your systems cannot be connected. Then the agent becomes a step someone has to copy in and out of by hand.
The question is not whether it can be done. The question is whether it happens often enough to be worth setting up.
How to pick a first use
Do not choose on what sounds impressive but on a simple calculation. Score candidates on four points.
| Criterion | Ask yourself | Strong signal |
|---|---|---|
| Frequency | How often does this happen? | Daily or several times a week |
| Time per instance | How many minutes does it take now? | More than ten minutes |
| Regularity | Can you write down the steps? | Yes, without listing exceptions |
| Cost of a mistake | What if it goes wrong? | Someone sees it and corrects it before it goes out |
What setting it up actually takes
The biggest surprise for most companies is that the model is the easy part. The work sits in connecting to what you already have.
- Describing the process as it really goes, including the exceptions everyone knows by heart.
- Arranging access to the systems, with the smallest rights that suffice.
- Testing with real cases from your own archive, not with invented examples.
- Setting limits: what it may do itself, where it must stop and ask.
- Recording what runs and why, so your successor understands it.
And privacy
As soon as an agent touches client data, it is no longer just an IT project but also a processing question. Record which data ends up where, choose EU processing where possible, and add it to your processing register. That sounds heavier than it is, but putting it right afterwards is heavier still.
The summary: it is not hype and it is not a miracle cure. It is a tool that pays off on repeated work with clear rules. Start small, pick something whose time saving you can measure, and keep a human at the controls wherever a mistake reaches a client.