Where AI helps, and where it does not
The question you are asking is probably the wrong one. It is not “how do we use AI”, it is “which of the work we already do has the right shape”. Shape matters more than sector: two very different companies have the same problem if the work looks like reading and sorting text.
Where it helps
Text somebody reads and sorts.
Requests arriving by email, reports, orders written in plain words. Someone opens them, works out what they are about and sends them to the right person. That is repetitive work on language, and it is where AI pays off most.
Dirty data somebody tidies by hand.
The same customer spelled three ways, addresses with no format, duplicates nobody has time to merge. It is not difficult work: it is boring work that nobody ever finishes.
First drafts somebody then corrects.
A reply, a summary of a meeting, a product description. The machine writes the first version, a person puts it right. It works because the human check is already in the process.
Searching through things with no index.
Contracts, old quotes, internal documentation. When the answer exists but is buried in three hundred files, the problem is finding it — and that AI does well.
The common thread: an approximate result is acceptable, and somebody looks at it anyway.
Where it does not
When the rule already exists.
“Over five thousand it needs a second signature” is not something to be learned: it is a line of code, exact and free. You reach for a model when the rule cannot be written down.
When the answer must always be the same.
Calculations, VAT, deadlines. You do not want a system that almost always gets the sum right.
When the problem is two systems not talking.
This one comes up more than all the others, and it deserves a section of its own — just below.
When the data is not there.
AI on wrong data does not produce wrong answers occasionally: it produces them fast and with confidence, which is worse.
The most common case: “we want AI in the CRM”
We get asked this often, and there is nearly always another problem underneath.
A CRM is worth what is inside it. If customers are duplicated, if half the deals never get updated because updating takes six clicks, if the ERP does not hand it the orders and someone retypes them on Monday — that CRM does not have an artificial intelligence problem. It has a data and integration problem. AI on top of a half-empty CRM finds nothing, because there is nothing to find.
The right order is boring and it works: first make the data arrive on its own, then make it clean, then — if it is still needed at that point — put something on top that reads it.
Once the CRM is in order, AI finally has material: working out which requests look like ones already closed, writing the first reply to a customer, telling a salesperson what happened on an account without making them read two years of notes.
Integration first, intelligence second. The other way round you pay twice.
What nobody tells you about the cost
A model is not a purchase, it is a subscription with consequences.
It costs on every use, and the bill grows when things go well. It changes under you: the provider updates, and something that worked starts answering differently. It has to be watched: if nobody ever checks the answers, errors pile up quietly. And above all somebody has to stay responsible for what the system says to a customer. That somebody is always a person.
This is not an argument against AI. It is the reason to put it where the return is clear, and not everywhere.
How to decide, in four questions
Before putting AI into a process, ask yourself:
- Is it repetitive work on text or on untidy data?
- Is an approximate answer good enough?
- Is there already somebody who would look at the result?
- Does the data you need exist, and can it be trusted?
If any answer is no, it is probably not a job for AI — and that is nearly always good news, because the other solutions cost less and break less.
Not sure which case is yours?
Neither are we, until we have looked. It starts with a process analysis: from there you can see whether your problem has the right shape for AI, or whether it is something else in disguise.
How the analysis works