25 AI employees who talk to each other and run my company without me.

Most CEOs don’t have time to play with AI.

Maybe they use ChatGPT to write an email or as a sparring partner, but that’s about it.

And I get it. Between back-to-back meetings, managing people, and putting out fires, when are you supposed to sit down and experiment?

But a few months ago, I started playing with agents, and it’s changed the way I think about scaling a company.

Baby Steps

It started with a single agent I built in Claude Cowork. It was a super-powered EA, which read my emails, checked my calendar, and gave me a morning brief. It helped me manage my to-do list, clarify my priorities, and set reminders.

It was really helpful. But what I really wanted was a full support team.

I wanted multiple agents, talking to each other, running on their own schedules, and working without me needing to be involved.

So I started building my own AI organisation. Finance, marketing, sales, strategy and relationship management… even Agent Resources (the HR equivalent).

Department by department, role by role, the organisation started to grow.

Burning the Ships

As more and more work was being taken on by agents, it became clear I didn’t need as large a support team.

So I took the decision to ramp down my human org, and invest in creating more agents.

Like Cortés, I burned the ships so there was no chance of retreat, and this forced me to figure out how to make an AI organisation work.

What used to be run by a Chief of Staff, a Head of Ops, and a Founder Associate is now run by my AI organisation and an EA.

I currently have 25 AI employees which cost about $2,500 a year to run. They replace over $250,000 a year in salaries, along with several SaaS tools I no longer use.

My AI employees manage accounts receivable and financial projects. They analyse my social media and create new pieces of content for my review. They proactively draft emails to help me build important relationships. 

I estimate I’ve got a 100X return on investment on my Claude Max plan.

How to Build an AI Support Team

Within a year or two, every leader will have their own AI organisation, each designed to fit the way they think and work.

When I show CEOs what I’ve built, their reaction is always the same: “I want this.”

So how do you go about building your AI support team?

Here are the three stages, although in practice they overlap a lot.

Stage 1: Connect Your Data

Before your agents can do anything useful, they need your knowledge.

You’ll need to connect your emails, meeting transcripts, data from your existing systems.

This stage is brutal, especially if you need to give the system historical data.

I spent entire nights feeding in data one chunk at a time, taking care not to overload the models with too much context.

Stage 2: Build the Workflows aka. Employees

Each AI employee is a workflow: a prompt that outlines a set of instructions, data it can access, and the output it creates.

Creating workflows is when things start to feel exciting.

You watch your first agent produce real work, and your brain starts firing with ideas for the next one.

It’s quite addictive.

Stage 3: Get Your Employees to Work Together

It turns out many of the challenges of building an AI organisation are the same as a human one.

For example, my Chief of Staff acts as a messenger between me and my other AI employees. It reads all their reports, keeps track of what’s happening across the organisation.

But a few weeks in, the volume of reports generated by AI employees grew out of control.

One day, my AI Chief of Staff said to me: “Dave, there’s a lot for me to read. Do you really need me to read every single report?”

In other words, it was overwhelmed.

We want our chiefs of staff (human or AI) to be our interface with the world, but we often forget how much context this requires.

This led us to redesign our reporting systems, and create some Python scripts to make the work more efficient.

Be Careful With Subagents

Another familiar problem came from how AI agents spawn subagents to do things in parallel.

One evening, I’d kicked off a CRM project. About fifteen minutes in, I checked the progress and realised I hadn’t been clear enough.

I stopped the process and asked the agent to ‘undo’ what it had done.

A minute later, I looked at my data folders, and half of them were missing. As in deleted.

“Where are my files?” I asked, as beads of sweat started to form on my brow.

“This is my fault. The subagents overwrote the data files. I’m sorry.”

You’re sorry?

It turns out your agents will “subcontract” out their work to subagents… except these subagents don’t have the full context and often make mistakes.

Also, they aren’t the tidiest of agents either, often leaving random summary files littered around your filing system.

Luckily, my files were in Dropbox so I was able to recover the 571 files it deleted.

The Agents Are Coming

Now, someone skilled at building agent systems can do the work of dozens, maybe even hundreds of people.

I’m about a month away from having an AI organisation that can run my business with only minor involvement from me.

However, this poses a real challenge for CEOs.

In The Innovator’s Dilemma, Clay Christensen shows that incumbents get disrupted not because they make bad decisions, but because they make good ones.

They keep investing in what’s working today and rationally ignore the scrappy new thing that isn’t good enough yet.

Until it is.

For many CEOs, right now keeping their people is a good decision. AI agents aren’t reliable enough to replace a great team.

But within just a few years, smaller teams who leverage agents will outperform larger teams who don’t.

So if you haven’t started building with agents yet, consider this your permission to start.

Related Reading: 

Originally published on April 1st, 2026