The right question is not whether AI is worth using. It is whether it is worth building anything when a generic assistant already exists for a few euros per user.
What a generic tool does well
Assistants like ChatGPT, Copilot or Gemini are good at writing, summarising, translating and explaining. For individual work they are hard to beat at the price they cost. If the goal is helping each person write faster, buying licences solves it and nothing else is needed.
Where they stop
They stop the moment the task needs to know something about your company. A generic assistant does not know the price you quoted that client, does not open the CRM, does not know internal rules and leaves no record of what it did. The person remains the link who fetches the information and pastes it in the right place. The manual work does not disappear, it just moves.
The difference in three points
Context. A custom agent answers from the company documents, history and records, and shows the source of every answer. The generic one answers from what it learned on the internet.
Action. A custom agent can create the record, send the email, update the system. The generic one produces text someone has to copy somewhere else.
Control. With a custom agent you define what it decides alone, what needs approval and who can see what. With a generic one, control sits with each user, which is why company information ends up in personal accounts.
When each is worth it
Buy generic licences when the goal is individual productivity and the material involved is not sensitive. It is the right call in most companies, and it is cheap.
Build when the process is repetitive, has volume, needs internal data and is always the same. That is where the return comes from removing the work, not from speeding it up.
In practice the two coexist: generic licences for writing work, custom agents for the processes that make money.
How to decide without spending money experimenting
Pick a process, measure how much time it takes per week and how often it repeats. If the answer is "often, always the same, and it needs information from our systems", there is a case for building. If it is "rarely, and each time is different", there is not.
That is exactly the calculation we run in the diagnosis, and we say when it does not add up.