There is a lot of pressure to adopt AI and not much guidance on where it pays. After building a number of these systems, the pattern is reasonably clear.
The good fit: high volume, verifiable output
AI earns its keep where the work is repetitive, the volume is high, and the result can be checked automatically.
Document extraction is the strongest example. Pulling reference numbers, dates, line items and totals out of supplier invoices is dull, error-prone work, and it is verifiable, because the line items must sum to the total. Anything that fails the check goes to a person. Anything that passes almost certainly did not need one.
First-line support is another. A well-grounded assistant that answers the same forty questions from your own documents removes a large share of repetitive enquiries, provided it hands over cleanly when unsure.
Drafting at volume works too, whether that is four hundred product descriptions or listing variants for different channels, as long as a person approves before anything is published.
The poor fit: low volume, high consequence, unverifiable
AI is a bad investment where the volume is low and the cost of a subtle error is high. Anything with legal or financial consequence, anything where being confidently wrong is worse than being slow.
The failure mode matters here. AI does not usually fail obviously. It produces something fluent, well formatted and wrong in exactly one detail, such as a transposed figure or an invented clause number. If the reviewer has to check every field against the source anyway, you have added a step rather than removed one.
The test worth applying
Before automating a task, ask: how would I know if the output were wrong?
If the answer is a rule a computer can apply, such as it must balance, it must match a record, or it must be one of five valid values, then automation will pay. If the answer is that an experienced person would have to read it carefully, you are buying a first draft, not an automation. Sometimes a first draft is worth it. Often it is not.
Start narrow
The projects that succeed start with one document type, one process, one measurable number. Measure the current cost in hours and errors, automate that one thing, and check the number again after a month.
Projects that start with a company-wide AI strategy tend to produce a strategy.