Your Team Has ChatGPT. Nobody's Workload Went Down.

On paper the maths was obvious. Forty people, a chatbot licence each, twenty euros a month. If every one of them saved twenty minutes a day it would pay for itself before lunch on the first Tuesday.

A year later the licences are still being paid and nobody can point at a job that stopped being done. People genuinely use the thing. They like it. Ask them and they will say it helps. Ask them which task left their week and the answer is usually a pause.

This is not a failure of the tools and it is not a failure of the people using them. It is a structural property of the chat window, and once you see it you cannot unsee it.

You are still the transport layer

Open a chat, paste something in, read the answer, copy it back out. That loop has a person in the middle of it at every step. The AI did the thinking part faster, which feels like progress, but the carrying is still yours: finding the file, opening it, deciding what to paste, judging the output, putting it back where it belongs.

For a one-off question that is a good trade. For something you do ninety times a month it is close to worthless, because the carrying was always most of the work. The thinking was the short part.

This is why the licence maths never lands. It assumes AI replaces the expensive minutes. In practice it replaces the cheap ones and leaves the tedious ones exactly where they were.

Three shapes of AI work, and almost everyone stops at the first

Answer. You ask, it responds. A better search engine that writes in sentences. Genuinely useful, almost never time-saving at scale, because you are still initiating every single interaction.

Assist. It sits inside a tool you already use and offers something before you ask. A suggested reply, a draft, a summary in the sidebar. Better, because the carrying is partly gone, but you are still the one deciding when it runs and what happens to the output.

Act. It has a job, it has access to the place the job lives, and it does the job whether or not you are looking. The email gets filed. The report gets built. The exception gets flagged for a human and everything else gets handled.

The gap between the second and the third is where every real hour of saving lives. It is also the gap that no amount of prompt training will cross, because it is not a prompting problem. It is an access problem.

What "access" actually means

It is less exotic than it sounds. Access means the automation can reach one of four things:

• ✓a file or folder, so it can read what arrives and write what leaves
• ✓a mailbox, so it can see what came in and act without being asked
• ✓a database or spreadsheet, so it can look things up rather than be told
• ✓an API, so it can talk to the system that already holds the answer

Nearly every job worth automating needs one or two of these. The reason your chatbot licence has not changed anything is that a chat window has none of them. It cannot see your mailbox. It has no idea what is in the shared drive. Every fact it uses had to be carried in by a person, which is precisely the labour you were hoping to remove.

The uncomfortable part: most teams cannot name the job

When a room is asked to list the repetitive work in their own week, the first list is almost always wrong. People name the thing they find annoying rather than the thing that consumes the most time, and those are rarely the same. Annoying is memorable. Time-consuming is invisible, because it is distributed in twenty-minute slices across five days.

The jobs that turn out to be worth automating tend to share four properties. They are repetitive, they are rule-shaped rather than judgement-shaped, they happen at volume, and getting them slightly wrong is recoverable. Score your candidates against those four honestly and the ranking usually rearranges itself.

That scoring exercise is worth more than any tool recommendation, and it is why we spend a full part of the workshop on it rather than on software.

What a team should do instead of buying more licences

Stop asking what AI can do. It can do almost anything badly and a narrow set of things extremely well, so the question has no useful answer.

Ask instead: which job in this team could be described completely in writing, happens at least weekly, and would be survivable if it were occasionally wrong in a way somebody noticed? That question has a small number of answers and every one of them is a candidate.

Then check the honest constraint before committing: can the thing that would do this job actually reach the data? If the answer is no, the project is an integration project wearing an AI costume, and it should be costed as one.

The part nobody sells you

Some of your candidates should stay manual. A task done four times a year by one person who is good at it is not an automation opportunity, whatever the deck says. Automating it costs more than it saves and adds a thing that can break quietly.

A team that leaves a session knowing which three of their five candidates are not worth automating has saved more money than a team that automates all five. That is an unpopular thing to say when you sell automation, and it is the most reliably true thing in this entire subject.

If you want to run this properly with your own team and your own files, that is exactly what the half-day workshop does. If you already know what you want built, start with a paid pilot instead and skip the workshop entirely.