Where AI is not yet good enough for SMEs

This article is about the limit. Anyone who knows that limit sets up their automations better than someone who has to discover it.

Why does AI not calculate reliably?

A language model predicts text. It does not calculate, it writes down what is a plausible number. With small sums that usually goes well, and on a VAT check across three hundred rows it goes wrong once — and you cannot see where.

The answer is not a better model but a different division of labour: the model writes the calculation script, the script calculates. Then the sum is fixed, it is repeatable, and you can read it back. Every calculation in the Studio works that way. Always ask whether you can see the sum; if you cannot, it is an estimate that looks like an outcome.

What happens with an exception nobody has written down?

This is where the biggest pitfall sits. An automation learns from what is in your systems. What is not in them — that this one customer always calls first, that with that supplier you check the price by hand since that business last year — does not exist for the system.

That knowledge sits in the heads of people who have worked there a long time. It is also exactly the knowledge that only comes out when something goes wrong.

That is why the order matters: watch along for a few weeks first, and only then automate. And that is why approval is on by default for everything to do with money and customers. Not because the system is stupid, but because it does not know your history.

Which work should you not automate?

Work	Why not	What instead
A judgement about a person (hiring, appraising)	High risk under the AI Act, and you want to do it yourself	Have AI summarise the documents, you judge
A difficult conversation or a complaint with emotion in it	The tone is the message	AI prepares, you deliver
An exception that happens once a year	Setting it up costs more than it delivers	By hand, and write it down
Something only you understand	Nobody can check whether it is going well	Record how it works first
A one-off job	An automation pays for itself through repetition	Just do it

The last row is the most underestimated. Automating pays off through repetition. Something that happens three times a year is quicker by hand than setting it up.

How do you know AI is doing it well?

By three things, and all three can be checked without knowing anything about technology.

Can the outcome be read back? Can you see what came in, what happened to it and why? If not, you will not know where to look when something goes wrong.

Does the system say so when it does not know something? An automation that fills in an amount anyway when the amount is unreadable is more dangerous than one that stops and asks.

Is the original kept? With every piece of processing, the original document should be findable afterwards.

Does this mean AI does not replace people?

For the work in this article: no. The work that does move over is the retyping, the forwarding, the looking up and the summarising — the work nobody enjoys at the end of the day.

What remains is the craft: the judgement, the conversation, the exception, the responsibility. That is also why we train the same people whose retyping we take away. A team that understands what the automation does is the first to see when it does something odd.