AI at work: where to start without wasting money
Most of the money lost on AI goes on licences bought before anyone worked out what they were for.
8 min read · Reviewed 2026-08-21
Start with the task, not the tool
The right question is not which AI tool to buy. It is which task steals hours every week and nobody enjoys doing. Write the worst five on a sheet, with an estimate of hours per week next to each. That sheet is worth more than any vendor demo.
Then pick the first one using the three criteria below, in this order. The winner is almost never the most expensive task — it is the most repeated one.
- It is repetitive and written. Text in, text out: summaries, replies to similar requests, sorting email, first drafts of proposals.
- A mistake is not serious. If getting it wrong costs a client or a fine, that is not the task to start with.
- Someone can check the result in seconds. If verifying takes as long as doing, you saved nothing.
If the first task you picked fails any of the three, pick another one. Cheaper than finding out in three months.
The two-week test
Before buying anything, do this with the cheapest individual subscription you can find, or even the free tier.
- Pick two people who actually do the task. Not the managers.
- Two weeks using the tool on that task only, nothing else.
- At the end, one question: how much time did you save, and what had to be fixed by hand?
- If the answer is 'not much' or 'not sure', the task was badly chosen — or the tool does not fit. Either way you spent tens of euros, not thousands.
Decide about data first, not later
If nobody has decided anything, the team is already pasting things into public chatbots — including things that should not leave the company. A ban does not fix this; it just pushes the problem out of sight. A three-line sheet does.
- What never leaves: identifiable customer data, contracts, health data, credentials. No exceptions.
- What leaves freely: public text, internal drafts with no names, technical documentation already online.
- What leaves only in approved tools: the middle ground — and for that, someone has to keep a short list of what is approved.
Five minutes in a team meeting and one shared sheet cover ninety per cent of this. GDPR applies exactly as it always did; what changed is the number of places data can leak from without anyone noticing.
The three expensive mistakes
- Buying licences for everyone before a small team has proven it works. Two months later half the seats are unused and nobody wants to be the one who cancels.
- Automating a bad process. AI makes the process faster — if the process was wrong, it is now wrong faster and at greater volume.
- Leaving out the people doing the work. Tools chosen in a management meeting and pushed down do not get used, and nobody says why.
If you would rather do this with an outsider who has been through it, that is exactly what the diagnostic is.See the consulting
Frequently asked
Do I need someone technical on the team?
To start, no. The first useful automations run on tools that already exist and need no code. Programming only comes in when you want to connect AI to your own systems.
How much should I budget for the first year?
Less than you think for experimenting — tens of euros a month is enough to prove or kill an idea. Real money should only come in once one task has a measured saving.
Will this replace people on my team?
On the tasks where it makes sense to start, no. It takes the boring work off the people doing it and gives hours back. Companies that begin by planning cuts tend to pick the wrong tasks, because they are looking at costs instead of at the work.
What if the AI gets things wrong?
It does, regularly. That is why the third criterion is that someone can check the result in seconds. Automation without human checking is only justified when the cost of an error is close to zero.
And if you would rather I did it
I help companies work out where AI is actually worth it, build teams of autonomous agents, and the software around that — custom apps and websites.