AI agents for SMEs: 5 concrete use cases

"AI agent" sounds like something for multinationals: huge projects, dedicated teams, six-figure budgets. In my experience the opposite is true: the best use cases are in SMEs, where one person wears ten hats and every hour spent retyping data or chasing emails is an hour taken away from the work that brings in revenue. Here are five concrete scenarios, with what it really takes to make them work.

First, a three-line definition. An AI agent does not just answer: it carries out steps. It reads what comes in (an email, a document, a number), decides according to the rules you have given it and acts inside your systems: email inbox, ERP, calendar, CRM. I explained how it differs from a plain conversational assistant in my post on the difference between a chatbot and an AI agent; here we look at what you can actually do with one.

1. The first filter on incoming email

The typical case is the info@ inbox where everything lands: quote requests, questions about opening hours, supplier invoices, complaints, spam. Someone opens it ten times a day and decides, email by email, what to do with each one.

The agent does the first pass for you: it reads every incoming message, classifies it (sales, support, admin), immediately answers the recurring questions using real data (opening hours, terms, the status of an order), forwards whatever requires judgement to the right person and flags the sensitive cases, an angry customer for example, without touching them. You open the inbox and find the work already sorted, with the easy replies already sent or sitting in draft.

What it takes: access to the inbox, a list of frequent questions with the correct answers, clear rules on what always goes to a human. Low complexity: it is often the first agent I recommend, because the volume is high and the potential for damage is limited.

2. From quote to invoice without retyping

The typical case: a customer asks for an offer by email, someone retypes the details into a Word or Excel template, on acceptance retypes them into the ERP for the order, then again for the invoice. Three rounds of retyping, three chances to make a mistake, hours nobody has.

The agent reads the request, extracts items and quantities, prepares the draft offer by pulling prices from your price lists and puts it in front of you for approval. When the customer accepts, it generates the order and the invoice from the same data, without anyone rewriting anything. On the final stretch of the chain, from invoice to payment reminders, I wrote a dedicated guide on invoice automation in Switzerland.

What it takes: price lists and master data in a readable format (not PDFs scattered across ten folders), an ERP or even just well-structured files, and human approval of the offer before it goes out. Medium complexity: it depends almost entirely on how tidy your price lists are.

3. The assistant for internal documents

The typical case: the answers exist, but they are buried in manuals, contracts and procedures. "What does supplier X's contract say about delivery terms?" means half an hour of searching, or a question to the person who "knows where it's written". If that person is on holiday, the work stops.

The agent indexes your documents and answers questions by citing the exact passage it took the answer from, so you can verify it in ten seconds. It works for supplier contracts, quality procedures, machine manuals, terms of sale.

What it takes: the documents gathered in one place and clear permissions on who can see what (a supplier's contract should not answer to everyone). Low complexity on the technical side: the real work is putting the documents in order.

4. The numbers watchdog that alerts you

The typical case: the numbers are there, in the ERP or on a dashboard, but you look at them when you have time, which means too late. You discover the unpaid invoice at month end, the margin drop at the end of the quarter.

The agent reverses the flow: it monitors collections, due dates and anomalies, and when something goes off track it writes to you. "Customer Y is past 30 days", "this week's collections are below average", "this supplier has raised prices by 12%". You no longer go looking for the data: the data comes to you, when it matters. It is the natural evolution of the work I do on data and management control.

What it takes: accounting data that is accessible and up to date, and thresholds defined together (what counts as an anomaly for your business is your call, not the algorithm's). Medium complexity: the value lies entirely in the quality of the underlying data.

5. The sales follow-up that never forgets

The typical case: quotes sent and never chased. Not by strategy, but for lack of time. An offer with no reply after two weeks is almost always an offer lost in silence.

The agent keeps the list of open quotes, and after the number of days you decide it prepares a reminder personalised on the content of the offer, puts it in draft for you (or sends it on its own, if the rules allow) and updates the CRM with the outcome: replied, postponed, lost. No opportunity fades away just because nobody had the time to pick it back up.

What it takes: a record of the quotes you have sent (even a well-kept spreadsheet will do), the right tone for the reminders and precise rules on when to stop. Low complexity: it is one of the cases with the fastest payback, because it touches revenue directly.

What it takes to make them really work

Here I have to be honest, because this is where projects fail:

• Clean data first of all. An agent running on dirty data works no miracles: it makes the same mistakes as before, only faster. Up-to-date price lists, clean master data, documents gathered in one place.
• Clear rules on what it can do on its own. Answering a question about opening hours, yes; giving a discount, no. You write the perimeter down beforehand, you do not discover it afterwards.
• Human oversight at the start. In the first weeks the agent proposes and a person approves. Its autonomy widens only after it has proven it deserves it.
• One process at a time. The agent that "does everything" does not exist. Five small, reliable agents are worth more than one huge project that stalls halfway.

How much an AI agent costs

Orders of magnitude, no promises: a simple case built on tools you already pay for (Copilot Studio inside Microsoft 365, for example) starts from a few thousand francs of analysis and configuration. A custom agent that touches several systems, with connections to your ERP and complex rules, sits in the range of a well-built automation project: I laid out the detailed numbers in my post on how much it costs to automate a process in an SME.

The criterion for deciding stays the same: hours saved times hourly cost, plus the mistakes you no longer make. If you want to see how I set up these projects, from choosing the process to putting it into production, the reference page is the one on custom AI agents for SMEs.

---

Not sure which case to start with? That is the first thing I clarify in a consultation: I look at your processes, your volumes and the state of your data, and I tell you which agent will really pay off and which is better postponed. My automation and AI service always starts there, from the process, never from the tool.

Frequently asked questions

Can an AI agent make mistakes?

Yes. And unlike a chatbot, which gets an answer wrong, an agent can get an action wrong on real data. That is why you start with strict rules, limited permissions and human oversight on the critical steps, and only widen its autonomy once the agent has proven reliable.

Do you need ChatGPT Enterprise or will less do?

It depends on the process, not the brand. You often start with tools you already have, such as Microsoft 365 with Copilot Studio or the business versions of AI tools, and move to custom builds only when volumes and complexity justify it.

Is my data safe with an AI agent?

It can be, on two conditions: use the business versions of the tools, which do not train models on your data, and give the agent access only to what it needs. In Switzerland the nFADP applies: clear contracts with suppliers and care with personal data. I explored the topic in my post on ChatGPT at work and the nFADP.

Which use case should you start with?

The one with the highest volume and the clearest rules: lots of repetition and little ambiguity. That is where the agent delivers the most and makes the fewest mistakes. A small process that really works beats an ambitious project that stalls halfway.

← Back to the Magazine