The question always arrives in the same order. First “does it really work”, then “what would it do for me”. The first has no useful answer. The second does.
An AI installed in a company does six things well. It does four others badly. The rest is sales talk.
What it does well
Reading and sorting what comes in. E-mails, supplier invoices, forms, messages. It classifies, extracts what is useful and puts each thing in the right place. It is the task where it is most reliable, because a wrong classification takes one second to fix.
Writing a first draft. A customer reply, a quote, a product sheet, meeting minutes. The first draft is the long work, the proofread is the short work. Reversing that ratio is most of the gain.
Restating the same information in several formats. A technical sheet becomes a sales description, an e-mail, an answer to a precise question. In Switzerland, that also means in three languages.
Watching and alerting. A stock level going down, a payment running late, a negative review published this morning. The machine never gets tired of looking every day.
Searching your own documents. Finding the clause of a contract signed three years ago, or the reason a price was set the way it was. It is the most underrated gain, because nobody counts the time spent searching.
Chaining several steps without being prompted again. An e-mail arrives, it identifies the customer, finds the order, prepares the reply, updates the record. That is where a system differs from an assistant: it finishes.
What it does badly
Deciding when information is missing. It does not say “I do not know”, it produces a plausible answer. On a price, a delivery date or a legal rule, a plausible answer is a wrong answer. Every task with stakes goes back through a human, and that is an installation rule, not a communication precaution.
Judging a human situation. An angry customer, an employee who is struggling, a partner who needs careful handling. It can write, it cannot feel.
Working on what it does not have. An AI connected to nothing invents; an AI connected to your data answers. The quality of an installation is decided almost entirely there, and almost not at all by the choice of model.
Replacing a profession. It removes tasks, not jobs. The difference is not diplomacy: a task has a beginning and an end, a profession is a series of judgment calls between the tasks.
How to know what it would change for you
Take a normal week and write down what repeats. A task that comes back more than three times a week, always follows the same steps and whose errors are easy to catch can be automated today. A task that requires deciding with incomplete information cannot be, and will not be next year.
That list fits on one sheet of paper. It is worth more than any demonstration, because it is yours.