Half the companies we talk to want to “adopt AI.” Ask them for what, and you get silence. That is the problem right there. AI is not a goal, it is a tool, and like any tool it is good for a few things and useless for the rest. This article is about telling those apart before you spend your budget on a chatbot nobody uses.

We are mostly talking about large language models here, which is what most people mean by AI these days. They generate text, read text, search through text. They are good at that. What they do not do is guarantee you the truth or reliably do math. That single fact explains everything that follows.

Where it actually saves time and money

The common thread in every good deployment: it is fine if it gets things wrong now and then, because a human checks the result anyway or the mistake is cheap. That is where AI works beautifully.

Support triage and pre-processing. An incoming ticket gets read, tagged, prioritized, and routed to the right person. A human still answers, but nobody has to wade through a hundred emails by hand every morning. You feel that one immediately.
Drafts and first passes. A suggested reply, a summary of a long thread, call notes turned into a clean writeup. Nobody sends it blind, but starting from a reasonable draft beats starting from a blank page.
Search over your own documents. This is the most underrated use case, in our opinion. Hook AI up to your internal wiki, contracts, or docs, and you stop spending half an hour hunting for the one paragraph you need. You ask, and you get an answer with a link to the source.
Repetitive, annoying ops. Reshaping data from one format to another, pulling structure out of unstructured text, tagging a thousand items. The kind of work an intern would do and resent.

Notice that in none of these cases does AI make the final call on its own. It prepares, sorts, suggests. The decision stays with a person, and that is exactly how it should be.

Where it is a bad idea

Anything that has to be guaranteed correct with no human in the loop. Calculating an invoice, drafting a legal clause, a dosage, anything with numbers or binding consequences. The model will write you confident nonsense without so much as blushing. If a mistake hurts, it cannot be the last word.
Replacing judgement. “Let the AI decide who to hire, who gets the loan, who gets let go.” No. This is not about technology, it is about responsibility. You cannot delegate that to a model that will not even remember its own decision a week later.
The vanity chatbot. The little bubble in the corner that answers “please contact our support” to everything. It helps no one, annoys the customer, and embarrasses you. If you cannot name the specific question it should answer, do not build it.
Long tasks with many iterations. If you will be working on it again every week, the model knows nothing from last time and you end up in a loop fixing the same mistakes. Proper automation, or just a person, pays off more here.

How to start small and measure it

Do not kick off with a big project. Start with one ugly, repetitive task that measurably annoys someone in the company.

Pick one process with a clear metric. Not “improve support,” but “time from ticket received to first reply.”
Measure where it is now. Without a baseline number you will never know whether AI helped or you just bought yourself a feeling of progress.
Build the smallest possible version and run it on real data alongside the existing process, not instead of it.
Compare after two or three weeks. Time saved, error count, how often a human had to step in. If it is not saving anything, switch it off. That is not a loss, that is budget saved.

And a few things we insist on. Host it so your data stays yours, ideally self-hosted or at least with a vendor that will not train on it. Own the accounts and the keys, so you are not hostage to a single platform. And work out token costs up front, not when the invoice lands, because “a few cents per query” adds up fast across thousands of queries a day.

The one-line version

AI makes sense where it prepares work for a person and a mistake is cheap. It does not make sense where it is supposed to carry the responsibility or guarantee correctness on its own. Start small, measure it, and if it saves neither time nor money, feel free to throw it out. Adopting AI because everyone else is doing it is the most expensive way to end up with no advantage at all.