I automated my company with AI for six months. Unfiltered feedback.

What nobody sells you in those LinkedIn threads about "I replaced my team with 12 agents and generate 40K MRR while sleeping."

A few months ago, I did what everyone promises to do: I handed over a large part of my operations to AI agents. Sales follow-ups, candidate sourcing, content publishing, CRM updates, SEO monitoring, calendar cleanup. The idea was simple and vaguely megalomaniacal: transform a tech studio CEO into the conductor of an army of polite robots.

Six months later, it's running. Really. But if you're looking for the ecstatic testimonial from the guy who "unlocked 10x productivity," move along. Here's the real assessment — the one with numbers, bugs, and irony.

The inventory, or how you end up with 60 robots without realizing it

To date, the setup counts 42 skills (specialized modules) and about sixty scheduled tasks, of which roughly a third are already disabled. Disabled, not because they didn't work, but because they worked too well, each in its own corner, and we had to merge a dozen into a single engine to stop the proliferation.

First lesson, free: automation doesn't reduce complexity, it relocates it. You no longer manage tasks, you manage automata that manage tasks. And automata, like interns, need you to explain very precisely not to send the email to the wrong recipient.

What works (and it's real)

Let's be fair, because there are real wins.

Content publishing is the most mature pipeline: posts go out on their own, to the right profiles, at the right times, with a self-repair mechanism that replays a failed post in the following days. Zero intervention for weeks.

Filing meeting notes in the CRM had its day of glory: a single pass transformed 28 transcripts into 60 CRM actions and 19 candidate profiles, without a single duplicate. The kind of task nobody ever did on a Friday night willingly.

And my favorite, because it's modest: cleaning up orphaned calendar slots, which faithfully reports, day after day, "50 meetings, 0 orphans." A conscientious little robot doing its job in silence. You get attached.

What breaks (and it's instructive)

Now, the part the dream sellers omit.

The bug that sends the email to yourself. On a follow-up where the prospect never replied, the system would fetch "the last sender in the thread" to determine the recipient… and that last sender was me. Result: carefully crafted follow-ups, sent enthusiastically, to myself. The robot returned a proud "OK, sent." Technically, it hadn't lied.

The day the variables didn't fill. Three emails went out to prospects with, right in the middle of the text, a chunk of raw uninterpreted code instead of the intended content. Internally, we politely called it "brand damage." The prospect mostly had to wonder if they were talking to a company or a sick cron script.

The drafts that multiply. At one point, the anti-duplicate check verified internal memory but not the actual state of the mailbox. Consequence: up to five drafts stacked on the same thread. The AI wasn't lazy. It was too zealous. That's worse.

And then there's the category "real-world infrastructure doesn't cooperate": entire waves of tasks in degraded mode because a third-party API was unreachable, LinkedIn quotas blocking invitations in a loop, and — cherry on top — a recent internal audit that was itself interrupted by rate limiting. In other words: the robot in charge of checking the health of robots got kicked out by the same type of failure it was supposed to monitor. There's a lesson in humility there.

The three truths never written on slides

1. Automating is stacking guardrails. My follow-up module is on its twentieth version. Its history reads like a cemetery: each bug bears the name of a prospect or candidate, and each incident gave birth to a new rule. Check the calendar before proposing a slot. Check that the day actually matches the date. Filter automatic replies. Don't recontact someone twice in twelve hours. The ratio between "safety" code and "business logic" code is crushing. AI does 80% of the work in 20% of the time; the remaining 20% takes 80% of the time, and those are what determine whether you look serious or ridiculous.

2. On sensitive topics, humans stay in control — and that's what saves you. Anything touching legal matters, unpaid invoices, delicate negotiations is never sent automatically: it's escalated for validation. This rule wasn't born from an abstract governance principle. It was born from incidents. We automate what's reversible; we keep our hand on what isn't. It's less sexy than an autonomous agent, but it avoids disasters signed with your name.

3. The real bottleneck wasn't technical. It was market size. Here's the most counterintuitive discovery. By optimizing the people-contacting machine, it started running on empty: "strong saturation, 17 out of 20 accounts already contacted less than 30 days ago." The problem was no longer "how to send more," it was "there's nobody new left to send to." You can build the smartest pipeline in the world; it doesn't manufacture prospects that don't exist. Automation doesn't create a market. It exhausts faster the one you have.

So, do we continue?

Yes. Without hesitation. The machine saves me real time, absorbs tasks nobody liked, and forces me to formalize processes I kept in my head. But it didn't replace me, and it didn't replace anyone. It created a new job: robot keeper. A job where you spend your days writing rules to prevent very fast systems from very quickly doing stupid things.

If you're getting started, remember this: AI isn't an employee, it's a power plant. Extremely powerful, indifferent to consequences, and needing a bunch of circuit breakers. The fantasy is autonomy. The reality is supervision — just at a new scale.

And honestly? That's already huge. It just has nothing to do with what they're selling you.

Sébastien Lonjon is the founder of Abbeal, a tech studio present in Paris, Montréal and Tokyo. This text is an internal feedback with numbers to back it up. No robots were harmed during its writing — a few were just unplugged.