To automate repetitive tasks with AI, start with a task you do every week and whose good result you know by heart. Keep a human wherever a mistake would leave the building. At Pillet Grenié Bureau, we started with our meeting notes. An agent reads the transcript, writes the notes and updates our tasks on its own. Here is what worked, what AI gets wrong when it works alone, and where we deliberately keep control.
Why we started with our meeting notes
After every meeting, the same routine. Reread your notes. Write them up. Work out who promised what. Create the tasks by hand. And sometimes forget one. Then a client would be waiting for an answer we had promised them.
We're not the only ones who want to hand off this kind of work. A Stanford study of 1,500 US workers found they mostly want AI to take over repetitive, low-value tasks: data entry, filing, routine processing. Freeing up time for more valuable work drove 69.38% of the responses in favour of automation.
Meeting notes tick every box. They come back every week. We know what good notes look like. And a forgotten task has a real cost. Most importantly, we didn't bolt a tool onto the way we already worked. We redesigned the process itself. According to McKinsey, the companies furthest ahead with AI are about three times more likely to have fundamentally redesigned their workflows. Redesigning the process matters more than the tool budget.
From transcript to an up-to-date task list: what our agent does after each meeting
We follow the standard chain of a well-automated meeting: recording, transcription, structured summary, human review, sharing, then creating or updating tasks.
In practice, our agent does three things.
- It classifies the meeting. Client or internal? Discovery, scoping, ongoing project? If a participant is unknown and no project matches, it treats them as a new contact. When in doubt, it picks the most cautious stage.
- It pulls out what matters. A summary of a few sentences, the decisions with their context, the action items with their owner, and the topics covered.
- It writes and files. The notes follow our usual template. Each action item becomes a task in our project management tool, attached to the right client.
One detail matters more than it seems: every task keeps a link back to the meeting it came from. When you wonder why a task exists, you get back to the conversation in one click. And everything stays editable: summary, decisions, action items.
The time saved is real. One recent source estimates these tools can cut the time spent writing up after a meeting by 90%, while warning that it depends heavily on audio quality. True. And that's not the only trap.
What AI gets wrong when left alone, and what it costs you
Misheard names
Transcription swaps rare proper nouns for common words. A client's name or a product name turns into a word that sounds like it. The result: notes nobody can find and tasks filed in the wrong place. So we give the transcription a list of the words it's likely to hear. A second pass then fixes the known errors that still slip through. Neither is enough on its own.
Wrong times
We work across two time zones, one of them offset by half an hour. In our experience, a language model is unreliable at calculating times, and even less reliable when daylight saving kicks in. One wrong time in a recap and nobody trusts it again. So we do the calculation in code. The model receives a line that's already done, with a clear instruction:
Meeting: 14:00–15:00 Paris · 17:30–18:30 Colombo
Instruction: copy this line exactly as is, do not recalculate any time.
The model has nothing left to calculate. It copies, and the time can no longer be wrong.
False contradictions
We had asked the agent to flag gaps between what was said in a meeting and our project notes. One day it raised what looked like a great alert: the note said the project was finished, the meeting showed a blocker. Except the note had been updated the day after the meeting. Not a contradiction, just chronology. Without dates, the agent mistakes every normal bit of progress for a conflict. Within a week, the team would have stopped reading those alerts. Now it knows both dates. A note newer than the meeting means the project moved forward: nothing to flag. A note older than the meeting means it needs updating, and that alert is useful.
Tasks assigned to the wrong person
A task must belong to one of the two of us, to both, or to the client. The model follows this rule most of the time, but not always. Sometimes it invents an owner or leaves one out. A task without an owner never gets done. So we check every assignment after the AI has run: any value outside the list is rejected. The instruction given to the model isn't enough. The guarantee comes from the check that follows.
How we automate a repetitive task for a client without losing control
When a client asks us to automate a repetitive task, we start from these lessons. The setting is often the same: a team already using AI on their own. According to Salesforce, among employees who use generative AI at work, 55% do so without approval or oversight from their management. One well-scoped agent beats a dozen invisible uses.
The decision. One task, with a result the team knows how to judge. Not a do-everything assistant. That's the principle behind our custom AI agents.
The safeguard. We put human approval on what goes out, not on what the agent reads. Reading a document, looking something up or summarising a conversation happens without waiting. Sending an email, publishing, or turning a contact into a sales proposal takes a click. Clicking "Approve" takes a few seconds. A bad proposal sent automatically can cost you a client. During reviews, we also check for a quieter risk: what the person approves must be exactly what gets executed, not a version that changes after they've said yes. Finally, dates, times and lists of allowed values are checked by code, never left to the model.
The result. The repetitive work disappears. The notes are ready before the next meeting. No task gets forgotten. And the team stays in control of everything that commits the business. That's the difference between an impressive demo and a tool you still trust six months later.


