From Managing Your Work to Managing AI Agents

Your work is now to manage AI agents. Are you ready?

The Job Posting Nobody Has Written Yet

Ah, the demo! Someone types one sentence, the AI agent books the flights, drafts the report, fixes the spreadsheet, and the room applauds. Then you try it on Monday, with your actual inbox, and it books the wrong month. 😅

That gap, between the demo and your Monday, is where the next most valuable skill lives. Not building AI agents. Managing AI agents.

I spent years as a pre-sales solutions architect, translating between what software could actually do and what a client needed it to do. Today I run two small businesses with a handful of AI agents doing work I used to do alone, at night, badly. The skill those two lives share is not technical.

It is supervision: deciding what to hand off, how to brief it, how to check it, and what to keep for yourself.

The numbers say the same thing.

In IBM’s 2025 CEO Study, 61% of CEOs say their organization is actively adopting AI agents and preparing to deploy them at scale.
In the same study, only 25% of AI initiatives delivered the return they expected.
The podcast episode Bâtir des équipes IA from TPC (French only) draws the same picture from the ground.

Employers are deploying AI faster than they are learning to manage what they deploy. So, the people who learn to supervise AI agents now have the opportunity to write their own job description…Isn’t that great?

What is an AI agent, and Why isn’t it Just a Chatbot?

You ask a question, a chatbot replies. You ask again, it replies. It basically waits for you to prompt something.

On the other side, an AI agent is given a goal, and autonomously (or semi-autonomously depending on the user) works towards that goal. It breaks the goal into steps, uses tools (your calendar 📆, a spreadsheet, a browser, a code editor). Then it checks its own result, and comes back when it is done or stuck.

The difference is initiative.
➡️ A chatbot is a very fast intern who only speaks when spoken to.
➡️ An agent is an intern you can send out of the room.

Here are two requests, side by side – one to a chatbot, one to an agent:

Chatbot: “What does reconciling accounts mean?”
Agent: “Reconcile this month’s Shopify orders against the Google Sheet, flag anything that doesn’t match, and send me the list.”

The mechanism underneath matters, because it explains why the mechanism might sometimes fail. A large language model predicts text, one word at a time.

Wrap that model in a loop (look at the situation, make a plan, act, check, repeat), give it tools and a stopping rule, and you get an agent.

That loop is why an agent can be brilliant for twenty steps and wrong on the twenty-first. ➡️ It never had understanding. It had a plan and a probability of what the next best step was.

Which is exactly why agents need managers, not people who will believe everything they say.

The Mindset Shift You Need to Make

First, the right mental model is a teammate, not a tool. A tool waits for your hand. A teammate takes a brief, works while you are elsewhere, makes mistakes you have to catch, and gets better when you write down how you want things done.

Companies have started drawing agents on their org charts for that reason; I did too, as you will see below.
The question to ask of an agent is not “what features does it have” but “what job am I giving it, and how will I know it did it well”.

Second, the technical wall has come down. Everything an agent needs from you is written in plain language: the goal, the rules, the steps.

Building a small internal tool, a dashboard or a workflow used to mean hiring a developer or learning to code. Today you describe it in a sentence and watch it appear, a practice the industry has nicknamed “vibe coding”.
The barrier to entry is no longer syntax. It is clarity: knowing exactly what you want, and being able to say it. Which, for most university graduates, is the one thing you were trained to do. 🎓

How it Showed up in my Own Business

Every Saturday at 6 pm, two agents go to work for me.

One is a product manager: it reviews what moved on every product and service across my two ventures and updates a dashboard.
The other is a chief product officer: it reviews the product manager’s update before I ever see it.
The product manager reports to the chief product officer. The chief product officer reports to me.

Yes, I built an org chart for software (yes I am weird like that 😅).

What Actually Happened

The first weeks were a mess, and not for a technical reason. The update came back as a wall of text in a terminal window, and I could not tell at a glance what was moving, what was stagnating and what was stuck waiting on me. So I did what any manager does with a new hire whose reports are unreadable: I changed the deliverable, not the person.

The report became a visual board with three possible statuses per item, and a short list of decisions pending on me at the top. The AI agents stayed the same. I made the decision to change their output.

Second episode. Every time I opened a new chat to write, I was re-teaching the model my own voice: sentence length, the words I refuse to use, where the humor goes. Twenty minutes of onboarding, every single time.

The fix was to write my voice down once, in a reusable instruction file the AI agent reads before it drafts anything. In other words, I created job descriptions for my agents with roles, responsibilities and expectations. Managers write those too right? 👩🏾‍💼

Most of my time managing agents goes where it would go with any new hire: scope, brief, review, feedback. The technical setup took one afternoon. The managerial part never ends.

I think that you understand where I am going with that now: if you want your AI tool of choice to do great work, you need to onboard it like you would any other human employee. ☝🏽

AI tools are really smart and they can predict a lot of things, but guessing what’s right in a new environment they haven’t been trained on? Not so much.🙅🏽‍♀️

The Supervisor Loop: Four Actions You Don’t Need Coding Skills For

After enough Saturdays, the pattern settled into what I now call the Supervisor Loop: Scope, Brief, Verify, Keep. Four actions you already know from managing people, this time, pointed at software.

1️⃣ Scope: One AI Agent, One Job

The beginner’s mistake is “figure it out.” Never say that to an agent. You’ll regret it and you’ll be the only one frustrated at the end of it. 😤
➡️ Give your agent precise instructions: summarize this week’s client emails into a one-page brief, not run my marketing. Then define the escalation rule: the conditions under which it must stop and ask you.

The French health insurer Alan runs its member support this way. As reported in the panel Bâtir des équipes IA: an AI pipeline resolves roughly 40% of member requests on its own and routes the rest to a human team, which let Alan absorb about a third more members without adding head count. Its co-founder and CTO Charles Gorintin put the principle in one line to La Libre in November 2025: we automate tasks, not people.

Humans on exceptions, software on the routine.

Here is the Scope test: Can you describe the job you want the AI to do in one sentence with a verb, an input and an output? If not, the scope is too wide.

2️⃣ Brief: Creating the AI Agent’s Context is the Job

A generic prompt gets you a generic result, and generic is now free.

The brief I use has four parts, and I call them the 4 Rs + Review: give the agent

your Research (what you already know),
the Requirements (what done looks like),
the Rules (what it must never do) and,
the Recipe (the steps)

Then keep a non-negotiable human Review at the end.

Put that brief in a file the agent reads every time, not in a chat you will lose. Think of it as the onboarding binder.
➡️ Writing it is where your expertise, the thing you were afraid was obsolete, becomes the asset: an agent with your methods produces your work; an agent without them produces everyone’s.

Here’s the Brief test: Would a competent stranger, who was handed only this file, be able to produce something you’d vouch for?

3️⃣ Verify: Assume the AI Agent’s Output Looks Right, Check Anyway

Agents fail politely, quietly. When they hallucinate, they neither announce it or warn you. They just do because…

Their raison d’être is not to be right, their purpose is to provide an answer to your request.

That’s a BIG difference, and it should inform how you approach their output: it will look fluent, formatted, confident, and it will sometimes be wrong.

The ethics-and-safety research calls this automation bias: accepting a result because it sounds authoritative. Companies have started hiring people who can exert cognitive distance and critical thinking when handed AI outputs.

Alan hands engineering candidates an AI assistant and code with a hidden bug, to see who catches it.
Shopify watches how candidates clean up flawed AI-generated code rather than how fast they produce it (both reported in Évaluer la maîtrise de l’IA en entretien).

Your version needs no code. There are two habits I encourage you to develop when reviewing an AI output:

Trace 3 claims it included in its output: each time, pick one number, one name or one date and follow it back to the source. 3 random checks should be enough to calibrate your trust.
Require a log: ask the AI agent to end every task with three lines on what it did, what it skipped and why. You read the log to check the AI agent’s work process, not the raw work.

Here is the Verify test. Don’t hesitate to evolve this test as your comfort with AI increases.

4️⃣ Keep: the Offline Test

Career coach Simone White asks the only question that matters here: if the AI agent is offline today, can I still do this work myself?

That’s a big question! In the past few months, research and debates have talked a lot about how too much AI usage could contribute to the shrinking of our brains. I don’t personally believe it will shrink our brain, but I do think that if we get into the habit of never questioning what it serves to us, never digging further in a topic in our own time or never formulating our own opinions on various subjects, we will lose the habit of thinking critically – beyond the surface.

So going back to my previous question :

If the answer is no, you have not delegated a task. You have outsourced a capacity.

And that’s what you want to keep!

Instead, hone the thinking that makes you the manager: the judgment about what is worth doing, the taste to tell world-class from mediocre, the relationships no agent can nurture and hold.

Here is the Keep test: once a month, do one delegated task by hand. Notice where you got rusty, double-down on improving it.

Scope, brief, verify, keep. That is a management job. You have been training for it since your first team meeting. 💪🏾

Three Old Skills With New Direct Reports

Every step of the loop is a skill you were already told mattered, usually in a course, a presentation or a webinar you did not take seriously. Here is the mapping, because it changes what “upskilling for AI” means.

Communication is the Brief. An agent executes exactly what you wrote, including the parts you did not mean. The person who can state a goal, the constraints and what done looks like in one page now outperforms the person who knows more but cannot write it down. Written clarity is no longer a nice-to-have. It is the instruction set.
Critical thinking is the Verify step. Agents produce fluent, confident, well-formatted output. The only defence is a person who asks “how do you know” of a paragraph that sounds right. The hiring exercises at Alan and Shopify test exactly this: not whether you can produce, but whether you can catch.
Project management is Scope and the escalation rule. Breaking a goal into jobs with a clear owner, a boundary and a handoff condition is project management with a new kind of team member. If you have ever run a group project where one person disappeared for two weeks, you have already managed an agent that lost its context.

None of the three is a technical skill, yet all three are what Zapier, Alan and Shopify are now testing for, in those words or close to them. The curriculum did not change. The direct reports did.

Why “I’m afraid of AI” is the wrong sentence

The fear is not silly. The data backs part of it.

In the World Economic Forum’s Future of Jobs Report 2025, employers estimate that 47% of work tasks are done by humans alone today, 22% by technology alone and 30% by the two together; by 2030 they expect a roughly even three-way split.
Anthropic’s March 2026 study Labor market impacts of AI found no rise in unemployment among AI-exposed workers so far. There’s one early signal: the rate at which 22-to-25-year-olds start jobs in highly exposed occupations fell about 14% compared with 2022, a result the authors call barely statistically significant. The early-career rung is where the pressure shows first.

But “afraid of AI” is too blunt a sentence to act on. A 2025 study on job-related anxiety in the age of AI (Rizkina and colleagues) splits it into three distinct fears, and each one has a different fix:

Displacement anxiety (my role disappears): the fix is to move toward the work that is growing, and supervising agents is that work. Nobody is deploying 61% of anything without someone to manage it.
Obsolescence anxiety (my expertise is worthless now): the fix is step 2 of the loop – the Brief step. Your expertise is the brief. Agents without it produce generic output, and generic is free.
Control anxiety (the machine decides, not me): the fix is step 1 and step 3. You set the scope. You set the escalation rule. You verify. Control was never the agent’s to take.

There is a fourth risk the fear conversation skips, and it is the one that worries me most for university-educated professionals: cognitive debt. Sinead Bovell, Simone White and the ethics-and-safety literature all land on the same warning.

Hand the model your first draft, your problem framing, your brainstorming, and within months you no longer trust yourself to decide without it.

The ‘Keep’ step exists for that reason. The skill is not using AI agent. The skill is staying the person who could do without it.

So the right sentence is not “I’m afraid of AI.” It is “I haven’t tried working alongside AI yet so I don’t know how I can measure up to it.” That is fixable in a weekend. 🗓️

A Weekend Experiment That Needs no Technical Background

Pick one recurring task that drains you: the weekly summary, the meeting prep, the follow-up emails.
Open one AI tool and stay with it. Claude, Gemini, ChatGPT; the choice matters less than the consistency.
Write the brief using the 4 Rs: what you know, what done looks like, what it must never do, the steps.
Run it once a day for five days. Each day, trace one claim back to its source.
Keep a one-line log of every mistake you caught.

That log is your first line of evidence that you can manage an AI agent. The remediation you come up with to handle AI mistakes is your differentiation on the market. It is also, in 2026, something that very few people can write honestly on a résumé.

Don’t believe the hype, the LinkedIn posts or the Instagram/TikTok posts. We’re all catching up to a technology that moves faster that we can blink an eye. – YOU ARE NOT BEHIND.

And What About Employers?

Let’s be real: most of the fear is really about them.

A little refreshed on the employer side of the data I cited at the beginning so you don’t have to scroll back up:

In IBM’s 2025 CEO Study, 61% of CEOs say they are actively adopting AI agents, and only 25% of their AI initiatives have delivered the expected return.
The French panel Bâtir des équipes IA tells the same story company by company.

That is not an AI model problem. That is a management problem, the same one this whole article is about, but at company scale.

Klarna made the loudest version of the mistake: in early 2024 it announced its AI assistant was doing the work of 700 agents and froze hiring. Fifteen months later its CEO told Bloomberg the focus on cost had gone too far and the company was recruiting humans again so a customer could always reach a person.
Alan went the other way, keeping humans on exceptions from day one, and kept both the automation and the quality.

Three Things Employers Can do Differently

Three things an employer owes the people it is asking to supervise agents:

Measure outcomes, not licences: the first wave of metrics counts tool connections and seats. The second counts lead times, volume and quality. Only the second one pays anyone’s salary.
Hire and promote for supervision, not prompting: Zapier’s AI fluency rubric places every hire on four levels (Unacceptable, Capable, Adoptive, Transformative). Since March 2026, assesses four components: AI mindset, strategy, building and accountability. Note the last one.
- Zapier’s own phrasing: with AI, you can delegate the work, but not the accountability. That is the manager’s job, and it is now in the hiring criteria.
Keep a human on the exceptions, on purpose: not as a transition measure but as the design. Alan’s CTO again: tasks, not people.

If You’re a Recent Graduate – Read This Part Twice

The entry-level market is where the pressure lands first; Anthropic’s data on 22-to-25-year-olds says so.

Here is what the employers who are hiring now actually expect, going by what they publish and how they interview:

AI in your core work, not at the edges. Zapier’s floor for every new hire since March 2026 is “Capable”: AI embedded in your main tasks, repeatable systems rather than one-off prompts, and a before-and-after you can describe. Using a chatbot to tidy an email does not clear the bar.
A process you can defend out loud. Interviews are moving from “show me the result” to “walk me through how you delegated, what you checked and what you rejected”. Zapier watches candidates work with AI live; Alan plants a bug and hands you an assistant; Shopify watches you clean up flawed AI code. A polished artifact with no visible reasoning behind it is now a weaker signal than a rough one with good judgment.
Accountability for the output. The line companies are converging on: you can delegate the work, not the accountability. If the agent’s mistake ships under your name, it is your mistake. Employers want people who already behave that way.
Proof, not credentials. A degree opens the door less than it did; a log of five agent tasks you ran, the errors you caught and what you changed opens it more. Bring the log to the interview.

The good news you can observe in that list: none of it requires a computer science degree. All of it requires the Supervisor Loop.

If You’re an Employee

If your employer has not done any of the three, this article is your argument. Bring the weekend log and show what you caught in your personal life. Why?
➡️ I think we can all agree that Knowledge Management, Process Management, Information Management and Data Management are far more complex in an organization 📋…So if your weekend log caught AI mistakes at the scale of your home, all the more reason to have more humans in the loop catching similar (and potentially riskier) mistakes at the scale of your organization.

Hindsight is 20/20

I did not come to this because I love technology. I came to it because I was stuck: two ventures, one pair of hands, and a to-do list that kept growing. Creating AI agents did not fix that. Managing them did. The first time an AI agent shared its report with me, I realized the thing I had learned was not a prompt trick.

It was the same thing my best manager taught me in my first job: say what you want, say what done looks like, check the work, and support your team with constructive feedback.

The skill is old. The direct [AI] reports are new.

Which is why I think managing AI agents is the most valuable skill of the next few years for exactly the people who think it is not for them: the humanities graduate, the first-time manager, the professional who has never written a line of code and does not intend to. You have been supervising work for years; yours and that of others. You just never had a teammate or report that works at 5 am and never gets offended by feedback.

So here is the challenge. This week, hand one task to an AI agent. Catch one mistake. Write down what you would tell a human who made it.

➡️ If you want to know where you stand first, my AI adoption test gives you three recommendations based on your current level: the AI Fluency Quiz.

TL;DR – From Managing Your work to Managing AI Agents

An AI agent is a model in a loop with tools and a goal: it goes, you supervise. 61% of CEOs say they are actively adopting them, and only 25% of AI initiatives deliver the expected return (IBM, 2025). The AI management gap is your opportunity.
Manage one with the Supervisor Loop: Scope (one agent, one job, one title, an escalation rule), Brief (the 4 Rs + Review, in a file the agent reads every time), Verify (trace one claim per output, require a log), Keep (if the AI is offline, can you still do it?).
“I’m afraid of AI” splits into displacement, obsolescence and control anxiety; each one is answered by a step of the loop. The risk nobody warns you about is cognitive debt, and ‘Keep’ is the fix. Start this weekend: one task, one tool, five days, one error log.

Sources

World Economic Forum, The Future of Jobs Report 2025, January 2025 — 47% / 22% / 30% task split today; near-even three-way split expected by 2030.
IBM Institute for Business Value, 2025 CEO Study (announced May 6, 2025) — 61% of CEOs actively adopting AI agents; 25% of AI initiatives delivered expected ROI.
Massenkoff & McCrory, Anthropic, Labor market impacts of AI: A new measure and early evidence, March 5, 2026 — ~14% drop in the job-finding rate for 22-to-25-year-olds entering highly exposed occupations vs 2022; no rise in unemployment detected.
La Libre, Chez Alan, IA et emploi font la paire, November 9, 2025 — Charles Gorintin, co-founder and CTO of Alan: “On automatise des tâches, on n’automatise pas des personnes.”
Bloomberg, Klarna Turns From AI to Real Person Customer Service, May 8, 2025 — CEO Sebastian Siemiatkowski on recruiting human agents again; the February 2024 “700 agents” claim is Klarna’s own press release.
Zapier, One year later: Raising the AI fluency bar for every Zapier hire, March 31, 2026 — four levels; four components (AI mindset, strategy, building, accountability); “you can delegate the work, but not the accountability.”
Bâtir des équipes IA (French panel, 2026, in MC’s research notebook) — Alan support pipeline: ~40% of member requests resolved autonomously, ~33% more members absorbed without added head count. Figures as reported by the panel; not independently verified.
Évaluer la maîtrise de l’IA en entretien (French panel, 2026, in MC’s research notebook) — Alan hidden-bug interview exercise; Shopify flawed-code exercise.
Rizkina et al., Job-Related Anxiety in the Age of AI, 2025 (PDF in MC’s research notebook) — displacement, obsolescence and control anxiety. Author names to confirm against the PDF.
Simone White, How to GET and KEEP a job in the Age of A.I. (video) — the offline test.
Sinead Bovell, A.I. Futurist: What Your Life Looks Like In 2028 (video) — cognitive debt.