Where AI actually fits in a company (and where it is just flashy)

Lists titled "50 ways to use AI" are useless, because they do not separate what works from what looks good on a conference slide. Let me try differently: we will walk the departments and, for each, say what is a settled thing today and what is not.

Paperwork and documents

This is the safest bet and it excites nobody. Invoices, delivery notes, contracts, customer orders arriving as PDFs. A model reads the document and pulls out the fields somebody has been retyping by hand.

It works even on documents that look different every time — precisely where template-based approaches (OCR with fixed field positions) failed for years. You do not have to define where on the page the total sits. The model finds it.

One condition: verify what it pulled. The line items must add up; the registration number must exist in your database. That is done in code, not by the model.

Customer support

The second safest thing, with an important distinction.

What works: the model reads an incoming message, classifies it (complaint / enquiry / invoice / noise), routes it to the right person, and drafts a reply. An agent reads it, edits it, sends it. That removes most of the time spent writing the same thing over and over.

What works less well: a bot answering customers unsupervised. We can build one and sometimes it makes sense — but only when it has access to your real data (RAG over your documentation, a live order-status lookup) and a clear boundary at which it hands over to a person. A bot inventing your returns policy is a legal problem, not a saving.

Sales and CRM

Call notes into the CRM. Summarising a hundred-message email thread for the colleague taking over an account. Extracting next steps from a meeting transcript.

These are small things nobody enjoys and everybody postpones. That is exactly why they pay off — not by saving hours, but by finally getting done.

What not to believe: a deal-probability score computed by a language model. Better statistical methods exist, and more importantly you need enough historical data. Most companies do not have it.

Marketing and content

Product copy generated from specifications works well, especially when you have thousands of items that would otherwise get no copy at all. Translation into further languages too — bearing in mind it does not replace a native proofreader, it just makes one cheaper.

Visuals are a real thing now. Product photography, variants for A/B tests, a consistent AI model across a campaign. We do this and it works — on condition that it is labelled as AI content. Platforms increasingly require it, and it is basic honesty toward the customer.

What not to believe: "AI will write your whole blog". It will. It will be average, it will sound like everything else, and Google will have no reason to prefer it. Writing needs somebody who genuinely knows the subject.

Software engineering

This is where the change is most visible. Writing tests, documentation, refactoring, explaining unfamiliar code, a first draft of a function. On routine work the speed-up is substantial.

On non-standard work — architecture, performance problems, concurrency bugs — the benefit is far smaller and occasionally negative. The model writes something that looks right and you spend an hour discovering that it is not.

Operations and inventory

Demand forecasting, stock optimisation, route planning. It works, but note: this is usually not a job for a language model. These are classical statistical and optimisation problems, solved by a different kind of tool — often a simpler and cheaper one.

If someone offers you an LLM for demand forecasting, ask why.

How to read all this

Notice what the working cases share: plenty of examples exist, a human checks the output, and nothing breaks when it occasionally fails.

And what the problematic ones share: they decide instead of a person, they touch money or the law, and nobody reviews them.

That is the whole filter. You do not need a list of fifty ideas — you need one that survives those three questions.