What AI can and cannot do for your business

I have been writing software long enough to know what it is like before this, and I have spent the last few years building with AI every day. Both experiences matter, because the useful opinion is somewhere between “it changes everything” and “it is a gimmick”, and getting there requires knowing which is which.

Here is the split as I have found it in practice.

What AI is genuinely good at

Drafting. First versions of copy, documentation, email, job descriptions, SQL queries, boilerplate. Not final versions. First versions, which you then edit. This is a real and substantial saving, and it is underclaimed because people expect the output to be finished.

Summarising. Reducing a long document to its usable points, pulling the action items out of a thread, turning a codebase into a summary of what it does. This is close to solved and saves more time than anything else on the list.

Searching your own documents. This is where AI earns its keep in a business context. An assistant that answers questions from your internal documentation, your contracts, your regulations or your product manuals is genuinely useful, because the information exists and nobody can find it.

I built a chatbot that answers questions about Romanian fiscal legislation from official sources. This is a good example of AI doing something that is hard to do any other way: the questions are ordinary, the corpus is large and precise, and the alternative — hiring someone to read legislation and answer questions all day — is expensive and slow.

First-pass code. For well-specified, bounded problems, the first version is good. You should still read it. But the cost of going from nothing to a working draft has dropped dramatically.

What it cannot do

Arithmetic and anything requiring precision. Models approximate. They do not calculate. Ask one to count characters in a string or total a column of numbers and you will get a plausible, confident, wrong answer. Anything where being approximately right is worthless needs a real implementation.

Know your business. An AI does not know your pricing rules, which customers are allowed discounts, why the Q3 process is different from Q2, or what you tried last year and abandoned. It knows what is in the documents you give it and what is generally true. That is a meaningful gap between an assistant that is useful and one that is trusted.

Guarantee anything. The confidence is not derived from correctness. This is the failure mode that actually matters commercially: a wrong answer delivered with total certainty is much more expensive than an admission of uncertainty, because nobody checks what never sounded unsure.

Replace the judgement about what to build. AI is very good at producing a plausible implementation of a thing you have described clearly. It has no view on whether you should build that thing. Nobody’s does.

The part I care most about

The single most common misunderstanding is that AI removes the need for expertise. In my experience it inverts it.

I have used these tools heavily for years. They make me faster at things I already knew how to do. They make me more dangerous at things I know just enough about to be confident. The bottleneck has moved from “can I write this” to “do I understand what is being written” — and the second question is not one these tools help much with.

What has genuinely changed is the threshold. Things that were not worth building because they were 80% boring and 20% interesting are now worth building, because the 80% is cheap. That is a real and significant change for a small team, and it is independent of the model’s quality.

What I would tell a business considering AI

Start with the boring, high-volume, internal problem. Not the customer-facing feature, not the strategic bet. The internal thing that a person does several times a week and nobody enjoys.

Make the system able to say “I don’t know”. This is harder to build than it sounds and it is the difference between an assistant people trust and one they learn to double-check.

Have a domain expert review the output. Not to catch every error — to catch the ones that matter. In the fiscal chatbot, the person who knows Romanian tax law reviewing responses is not optional. Without that step you have a confident system that will eventually say something wrong to someone who relies on it.

Keep the fallback path. Know what the system does when it is wrong, because it will be wrong.

And measure whether it removed work or moved it. If your team is now reviewing AI output as much as they used to do the task, the automation did not land.

In one line

AI is a very good way to remove the boring 80% of a task, and a bad way to remove the domain judgement in the other 20%. Build accordingly.

If you want to talk about where AI actually fits into your specific process, get in touch. I am happy to tell you that it is not worth it, which is an answer a lot of organisations need and few get.