How to bring AI into your company without wasting the money

The question we hear most often is: "We'd like to do something with AI. What could we do?"

It is the wrong question, though I understand entirely why it gets asked. The pressure is external — a competitor announced something, it was on every slide at a conference, someone raised it in a meeting. The trouble is that this is how you start a project nobody uses six months later.

The better question is duller: which work in this company is repetitive, tedious, and done by an expensive person?

Start from the pain, not the technology

When hunting for a first project, do not look for "where could AI go". Look for where somebody spends hours copying data from one place to another. Where the same questions come round again and again. Where something takes three days even though the work itself is twenty minutes and the rest is waiting.

Every company has these, and everybody knows where they are. They are usually described as "that's just how we've always done it".

Once you find one, ask three questions:

• How many hours a month does it eat? Under ten, leave it alone.
• Is there data — examples of how it has been done so far?
• Does it matter if it is occasionally wrong, or is that unacceptable?

The third one matters most. If the answer is "it can never be wrong", do not start there. Not because it cannot be done, but because a first project needs to succeed.

The first project should be small and boring

I understand the pull toward something big and visible. Resist it.

A good first project ships in weeks, touches one team, can be switched off with no consequences, and has a measurable effect. It needs no process redesign and no company-wide buy-in.

Candidates that reliably work out: pulling fields out of documents (invoices, delivery notes, contracts), sorting and routing inbound email, drafting first-pass support replies, searching your own documentation, generating product copy from specifications.

All boring. All effective.

What to measure — and measure it first

This is the step almost everyone skips, and then nobody can say whether it was worth it.

Before anything is deployed, measure the status quo. How long one item takes. How many arrive a month. How many are wrong. How many hours someone spends on it.

Without that number you will be having a feelings-based argument in six months. With it, you have an answer in ten minutes.

Measure the uncomfortable one too: how many outputs a human had to correct. That is the only figure that tells you whether the system genuinely helps or has merely moved the work from writing to checking.

A human in the loop is not a weakness

Nearly everyone wants full automation. Nearly nobody needs it.

A model that drafts and a person who confirms with one click saves most of the time and keeps control. A model that acts alone saves slightly more time and introduces a risk nobody is watching.

The sensible path is incremental anyway: let AI suggest and have a human check everything. When months of data show it is not wrong within some category, let that category through automatically. Keep reviewing the rest. Never advance without the numbers to justify it.

The money, and what nobody mentions up front

Model calls are cheaper than people expect. At sensible volumes most business deployments land in the tens to low hundreds of euros a month — negligible against work that used to cost a person hours.

The expensive parts are elsewhere: preparing data, integrating with the systems you already run, and the time of the people who have to explain how any of it currently works. That is where eighty percent of the budget goes. Anyone selling you an AI project who only talks about the model has either never done one or is not telling you everything.

The other thing worth saying out loud: costs scale with use. A twenty-euro pilot can be a two-thousand-euro production system. That is not a problem when you know in advance. It is a problem when you learn it from an invoice.

Where it actually gets stuck

Not on the technology. Almost never.

It gets stuck on data that is a mess, scattered across five systems, owned by nobody. On people who do not want to use it because nobody asked them. On a sponsor who will never see the thing in action. And on a missing decision about what happens when the system is wrong — who catches it and who fixes it.

All of that is solvable. But it has to be solved before anyone writes code, not after.

The whole thing on one screen

Pick one repetitive process costing somebody at least ten hours a month. Measure where it stands today. Deploy AI as a suggestion, not a decision. Track how many suggestions get corrected. When that number is low enough, let a slice run automatically. And if after three months it is not helping, switch it off — that is not a failure, that is information you bought cheaply.