Almost everything written about AI in business is written for companies with a data team, a platform budget, and someone whose job title contains the word "transformation". If you run a forty-person company, that advice is not wrong so much as irrelevant — it assumes resources you do not have and problems you do not yet have.
So here is the version for a business your size.
The one thing that genuinely changed
Until recently, getting a computer to read a document written by a human was a project. Not a hard problem in principle — a genuinely expensive one in practice. Every invoice layout, every supplier's idea of what a delivery note looks like, every handwritten note in a margin needed its own rule.
That cost has collapsed. A model can now read a document it has never seen, in a format nobody described to it, and pull out the six fields you care about. Not perfectly. Reliably enough to be worth checking rather than worth doing by hand.
That is the change. Everything else people are excited about is either a consequence of it or is not yet real at your scale.
The consequences are worth listing, because they do not look like "AI":
- Invoices, purchase orders and delivery notes stop being typed in by a person.
- Emails that arrive in an unpredictable shape can be routed and summarised.
- Contracts and reports become searchable by meaning, not just by keyword.
- Any process that currently ends in "and then someone reads it and rekeys it" becomes a candidate for removal.
Notice that none of these are new capabilities. They are old capabilities that just became affordable.
What has not changed
Your data is still your data. A model does not fix a customer list with four spellings of the same company. It will confidently work with all four and give you four answers.
Judgement is still judgement. A model will tell you which invoices look unusual. It will not tell you which supplier relationship is worth protecting.
Reliability is still engineering. A demo that works nine times out of ten is a party trick. Getting from there to something you can leave running requires the same unglamorous work it always did: validation, error handling, a way to see what it did and undo it.
Nothing about the pace of adoption is as fast as it looks. The gap between what is possible and what is deployed in ordinary businesses is measured in years, and it always has been. That gap is not a failure. It is where the verification happens.
The honest first step
If you want one concrete thing to do, it is not "explore AI". It is this:
Find the single most expensive place in your business where a person reads something and types it somewhere else. Count how many hours a week it takes. That is your candidate.
If the number is under two hours a week, leave it alone — the integration will cost more than the problem. If it is a day a week or more, it is worth a serious look, and it is worth measuring the error rate honestly before and after rather than assuming the machine is better.
The businesses that will get value from this are not the ones with the boldest strategy. They are the ones who picked one expensive, repetitive, well-defined task and did it properly.
The rest is still a demo.