Guide
Where should an SME start with AI?
Updated on
Start with your tasks, not with a tool. List the admin tasks that come back every week, note how long they take, then pick just one, frequent and easy to check, for a first project. An AI audit draws up that list with your teams and points to the task to start with.
Should you start with the tool or with the task?
With the task. The question “which AI should we get?” often comes too early. It assumes you already know what you want to hand over to it.
A tool chosen first then goes looking for a use. It gets installed, a few people try it, and everyone drifts back to their habits because the tool matches no particular task.
The better question is more down to earth: what takes up time every week, in our company, without requiring any thought? The answer points to the project. The tool comes afterwards, and it is not always the one you had in mind. To tell the families of tools apart, see our guide AI agent, chatbot or automation.
Which tasks suit AI in an SME?
Not every tedious task is a good candidate. The ones that suit a first project best share several of these traits:
- They come back often. A weekly or daily task generally pays back more than a yearly one, even if the yearly one is more painful.
- They follow rules that can be explained. If an experienced colleague can describe how they go about it, those rules can usually be given to an agent.
- The information already exists in digital form, or can be dictated: an email, a document, a record, a few sentences spoken after a meeting.
- The result is easy to review. Meeting notes, a draft reply or a completed record can be reviewed quickly.
- A mistake can be caught. A person sees the result before it reaches a client.
- Several people do the same thing. The gain multiplies, and the rules are already more or less shared.
In an SME, that often means meeting or visit reports, keeping records up to date, preparing replies and follow-ups, finding a piece of information in files and emails, and recurring reporting.
Which tasks are better left aside at first?
Three kinds of task make poor first projects.
Rare tasks. They are dreaded, but they do not come back often enough to justify the investment.
Decisions that commit a person or the company: choosing a candidate, granting a discount, settling a dispute. An agent can prepare the file. The decision stays with a person.
Tasks where the information exists nowhere except in one employee’s head. It has to be brought out first, for example by dictating it to a voice agent, before anyone can put it to use.
How can you draw up a first list yourself?
You can make progress before calling anyone in. Over an ordinary week, ask each team to note its repetitive tasks in a table as simple as this one:
| Column | What goes in it |
|---|---|
| The task | What is done, in one line |
| Who | The people who do it |
| Frequency | Per day, per week or per month |
| Time | Roughly how long it takes each time |
| Where the information comes from | Email, document, call, visit |
| Where it has to end up | The software or file to update |
The list does not need to be precise. It often brings out a few tasks that show up in several teams, and that is already a solid basis for discussion.
Its limits are well known. People underestimate tasks that have become automatic, and overestimate the ones that irritate them. That is where an outside eye becomes useful.
What does an AI audit look at?
An AI audit draws up that list on site, with the people concerned, and goes further. It looks at four things.
The work as it is actually done. Not the written procedure: the real steps, with the tools really used, workarounds included.
Time. Each repetitive task is recorded with how long it takes and how often it happens. That starting figure is what will later tell you whether the investment was worth it.
Tools and data. Can the software in question be connected? Is the information in it complete, current and organised in a usable way?
The order of priority. The tasks an agent can take over are ranked from most to least worthwhile.
You receive the record of time lost, the ranked list of tasks, a roadmap and an estimated return-on-investment plan. Those documents are yours: you can carry them out with us, with someone else, or not at all. The audit starts from €800 excl. VAT.
Which mistakes should you avoid?
Starting with the tool
This is the mistake described above. You can recognise it by one sentence: “We bought licences, now we need to find something to do with them.” A subscription to a general-purpose assistant is not useless, but, as delivered, it knows nothing about your clients or your procedures. It mostly helps each person individually, and rarely changes a process on its own.
Wanting to automate everything
A project that takes on ten tasks at once risks finishing none of them. Every task has its exceptions, its unwritten rules and its people to win over. Better one agent that does one thing properly and that the team genuinely uses. You widen the scope afterwards, building on what works.
Some tasks are not worth automating anyway, because they are too rare or because the human contact is where their value lies.
Forgetting the data
An agent works with the information it is given. If client records exist twice, if documents are filed however each person sees fit, or if half the history sits in personal mailboxes, the agent will inherit the mess.
That does not doom the project. It changes where it starts: sometimes the first step is to put one source in order, or to connect the tools that do not talk to each other.
Confidentiality is settled at that point too, not later. Our commitments are simple. We do not resell your data and we do not use it to train models. We sign a data processing agreement and tell you, before we start, where it is hosted and which services it passes through.
Deciding without the teams
They are the ones who know where time is lost, and they are the ones who will use the agent or leave it aside. A project decided without them risks not being adopted. Involving them from the first list onwards, then training them on their own work, does more for success than the choice of technology.
What does a good first project look like?
It is small and it is useful. One task, in one team, with a person who reviews what the agent produces. The time spent beforehand has been measured, so there is something to compare against afterwards.
That first project does two things. It should give time back to the people behind it. And it teaches the company how to work with an agent: what to hand over, what to check, what to keep. What comes next, whether a broader custom AI agent or process automation, is then decided on that experience rather than on a promise.
That leaves the budget. It is covered in the guide How much does a custom AI agent cost?