What's left for the CEO when AI does the rest
Cash, collections, checks, meetings, email, studies: what I handed to AI, what I kept, the limits, and where you should start.
Thibault Pons de Vincent
Founder of Oliverlist and VivaSync
In 2026, I made AI my right hand as a CEO. Today I have more than 100 AI agents in production: 24 in Oliverlist's product, 84 on my own desk as CEO. I have built those 84 agents on my own with Claude Code since July, along with 13 applications. They run 200 to 230 times per working day, 99.4% of those runs succeed, and the count has passed 22,000 runs since late July.
Which raises a fair question: if agents do the work, what is left for the CEO? If your company is owner-led, as many are, everything tends to climb to the top: approvals, follow-ups, payments, the occasional crisis. That is exactly the load AI can take off your desk, as long as you are clear about what stays with you. Here is my answer after four months: what I no longer do, what I keep, what does not work yet, and where I would advise you to start.
What I no longer do
| Task | What handles it | What I still do |
|---|---|---|
| Cash | A tracker connected to 3 banks: P&L, forecast, 60-day cash, every week | I read and arbitrate |
| Overdue invoices | Reminders drafted from 7 days late | I approve before sending |
| Checks | 4 diagnostics, about 190 checks every night | I read the anomalies |
| Meetings | Nora, an AI agent that runs the meeting and publishes the minutes | I decide |
| 4 inboxes sorted and summarized about 30 seconds after arrival, replies drafted | I approve the replies | |
| Studies | Pricing, client attrition, a new offer: a few hours instead of several weeks | I choose |
A few reference points. A 21-minute meeting run by Nora costs about $2.45, and 31 meeting actions were turned into tasks in a single pass. The app that sorts my email took two days to build. I have written up the finance tracker in How I rebuilt ten years of accounts in a single day, and the nightly checks in 190 checks every night.
The rest follows the same logic. An appointment-setting agent runs every 15 minutes: built in one day, it went live the next. An automated recruitment firm was built in six days. On the marketing side, 98 website pages have been translated into 3 languages, and articles and pages go out every week. Along the way, the finance spreadsheet, Notion and n8n were all replaced.
What I keep
Three things cannot be handed over.
Deciding. AI prepares, I decide. A reminder goes out once I have approved it. A pricing study gives me options and their consequences in figures; the choice stays mine. Everything arrives prepared, and nothing leaves without approval when the stakes call for it.
Carrying the vision. An agent optimizes what it was asked to do. It does not know what the company should become in three years, or what to stop. Setting the course, choosing the offers, saying no, writing the rules the agents apply: that remains the CEO's job.
Key relationships. A major client, a team member with doubts, a business partner, a banker: those conversations are not delegated. AI can prepare the file, recall the history, draft a note. The relationship itself is held in person. Wherever business rests on personal trust, this part of the job carries even more weight.
The time agents give back only has value if it goes there. If it fills up with operations again, nothing has changed.
The limits, honestly
AI prepares a lot; it does not decide. Several things deserve to be said plainly.
- The big leap did not come from AI. At Oliverlist, automation tripled productivity per person between 2019 and 2023, before generative AI. Since 2020, the chain has been automated from importing contacts to issuing invoices. AI entered the product in May 2023, for what automation could not do: read a reply, classify it, write.
- My CEO layer is young. It is only a few months old. Its effects are measured in places, not everywhere.
- The appointment-setting agent runs live, but its net effect is still being measured.
- The voice agent for client calls is on hold.
- AI makes mistakes. In Oliverlist's product, prospect reply sorting is 99% accurate on urgent replies, but 77% on positive ones. An audit of 7,612 replies found about 4 real errors per day. You only use AI well if you know this, and measure it.
- My evidence is internal. Oliverlist and my own job as CEO. I do not yet have an external client case to show you.
Where to start
- Start from your week, not from a tool. For one week, write down what you do and what you check. Anything that comes back every week and follows rules you could write down is a candidate.
- Connect your data first. Banks, CRM, email, calendars: an agent is only as good as the data it reads. That is the foundation of the method.
- Pick one task and one real workflow. Not a demo: a task your team uses from the first week. That is the principle of the sprint, a first tool in production in 14 days. The selection criteria are covered in How to choose your company's first AI project.
- Keep approval at first. AI prepares, you approve. Loosen your grip when the measurements justify it.
- Measure before you scale. Before and after, on the same indicators. When an effect is not measured yet, say so.
You need neither 84 agents nor some 131,000 lines of code to start. You need one agent that runs every day on a real workflow. Accounts and subscriptions stay in your name, and everything stays with you, documented.
For scale: the 84 agents on my desk as CEO run on a small server and a database, about $59 a month on top of the AI subscription. In Oliverlist's product, AI costs about 1% of revenue. The real cost of a project lies elsewhere: the time it takes to write the right rules.
Your turn
To find out what can leave your desk, book 30 minutes: you walk me through your day and your business, and I tell you what AI can take on. If you would rather start on your own, the online test places your company on one of four levels (manual, tooled, automated, synchronized) and gives you the 3 tasks to hand to AI first. The result appears straight away, and nothing is sent or stored.