AI agents are your new employees

The way founders talk about AI is changing fast. A year ago everyone called it a tool. Now more of them are calling it an employee, a team member, or “my agents.” The vocabulary swap looks small. The org-design shift behind it is not.

That mental model is doing more work than most founders realize.

The 10-employees thesis

Alex Lieberman, who co-founded Morning Brew and now runs the venture Tenex, has put a number on this shift: every business in five years will operate with roughly ten employees, and nine of them will be AI.

He’s not the only one pointing here. Sam Altman has been writing about the coming one-person billion-dollar company for two years. Andrej Karpathy has been arguing that software development itself is collapsing into prompt-and-review workflows. The pattern keeps showing up: small humans-to-output ratios, big agent bench.

You can disagree with the exact ratio. The number-of-humans-needed argument will land differently in a deep-tech R&D shop than in a service business. What’s harder to argue with is the mental model itself.

Why “employee” beats “tool”

Most founders still treat AI as a tool category. ChatGPT for writing. Claude Code for engineering. A few Zapier flows for ops. Each one a sharp object you pick up when a specific task lands on your desk.

That framing has a ceiling. You’re still doing every job. The tool just makes the job faster.

Treat the same systems as employees and the math changes. You stop asking “can this tool do X” and start asking “what role does this AI fill?” You start writing JDs for them. You think about scope, accountability, and what the handoff looks like when their output becomes someone else’s input. You build a management layer instead of a tool stack.

I wrote about this from the marketing side in the rise of the marketing engineer. The shift from specialist hires to a generalist with agents is partly a cost story and mostly an org-design story. You build the org around the work, with agents filling roles that used to require headcount.

What this looks like inside a business

Take a typical 10-person SaaS company. The traditional org chart has founders, two engineers, one designer, one marketer, one customer success person, one ops person, two SDRs, and a handful of contractors.

The framing pushes you to ask, role by role:

• Which of these jobs is mostly pattern matching at high volume? Hire an agent.
• Which is judgment, relationships, taste, or accountability? Hire a human.
• Which is somewhere in between? Hire a human with an agent reporting to them.

You end up with fewer humans and a lot more output. Not because the humans are working harder. Because the work that used to require a team of five now requires one strong operator running five agent employees. The same shift I covered in your terminal is the new marketing department, stretched across every function in the company.

The mistake most founders are making

The trap is treating agents like contractors instead of employees.

Contractors get a SOW, they go away, they come back with a deliverable, you review it, you pay them. Low management overhead by design. You wouldn’t expect a contractor to learn your voice, internalize your decision-making, or grow into the role.

Most AI implementations are built on the contractor model. Hire an agent for one task, hand it a prompt, accept the output. Done.

Employees work differently. They’re embedded in the business. They carry context. They get feedback. They have a manager who’s responsible for their output. They get better over time because someone is investing in them.

The founders winning right now are the ones who built that second relationship with their agents. They have a brief for each one. They review the output weekly. They tune the prompt the way you’d coach a junior. The agent that was middling at month one is sharp by month four because the manager kept refining the brief.

If you’re treating agents like contractors, you’re getting contractor results. Mediocre, generic, output you don’t trust enough to ship.

The companies that figured this out earliest are already showing the spread. Klarna told the market in 2024 that its AI customer-service assistant was handling chats equivalent to 700 full-time agents inside its first month, with higher customer-satisfaction scores than the human team it replaced. Shopify CEO Tobi Lütke leaked a now-public memo earlier this year telling every team to demonstrate AI can’t do a job before requesting more headcount. Different industries, same gravity: the org chart now starts with what an agent can do, not what a human used to do.

The new headcount math

There’s a more honest version of this thesis I haven’t seen many people articulate. It goes like this.

Old math: you hire ten humans, you get ten humans of output. Linear.

New math: you hire three humans plus seven agents. The three humans are senior operators with taste. The seven agents are pattern-matched, narrow-scope, always-on. The output looks more like fifteen humans, not ten. The cost looks like four.

But this only works if the three humans are good at managing agents. That’s a new skill set, and managing agents is different from being a strong individual contributor.

Most founders haven’t hired for it yet because the JD doesn’t exist. The closest analog is a senior engineering manager who used to run a team of juniors. You need that energy across every function.

What to actually do this week

If this framing is new to you, three concrete moves:

1. Walk through your org chart. For every role, ask: pattern matching, or judgment? Be honest. Most jobs are about 70% pattern matching once you look at the actual work.
2. Pick one role that’s mostly pattern matching and write a JD for an agent version of it. A real job description, not a prompt. What’s the scope, what’s the deliverable, what does success look like, who does the agent report to?
3. Run that JD as your prompt for a week. Treat the output the way you’d treat a new hire’s first week. Coach. Adjust. Don’t fire it after one bad day.

The companies that figure this out in 2026 will look unrecognizable two years later. Smaller headcount, more output, fewer middle management layers because the management is happening at the agent layer instead.

If you want help figuring out which functions to staff with agents at your specific company, the AI Opportunity Sprint is the four-week version of that conversation. It’s how I work with founders mapping where AI actually fits before they start hiring.

Anthropic’s Dario Amodei laid out the macro version of all of this in his “Machines of Loving Grace” essay last fall. The decade ahead is going to compress what would normally be fifty years of org-design change into about five. Whatever the exact ratio turns out to be, the direction is settled. The only real question is whether you’re early to it or late.

Jared Castronova is the founder of JAC Growth Marketing, where he builds AI-powered GTM systems for B2B companies.