AI is getting better. I thought I needed 100 people. I need ten. Codex keeps getting faster. Better. I have now people that used to do customer experience, now building code, building product. So, to answer your question, I don't know timelines anymore. They are literally a figment of my imagination. Every time I think I know something, AI or Claude says nah.
That's from a conversation I had on the Fargo Talks podcast with Jeff Fargo, and I want to expand on it, because every part of that answer would have sounded insane to me eighteen months ago.
How many people does it take to build a company now?
Fewer than you think. Far fewer.
When I started planning this company, I did what every founder does. I sketched an org chart. I costed it out. Engineering here, product there, support over here, and the number that fell out the bottom was around a hundred people to build what I wanted to build at the quality I wanted to build it.
That number is now ten. Not because I lowered my ambition. Because the tools changed underneath the plan.
I want to be precise about what actually changed, because "AI makes you more productive" is a useless sentence that means nothing to anyone. Here's what changed. A coding agent stopped being an autocomplete and became a thing you delegate to. You hand it a task, it clones the repository, reads the codebase, writes across multiple files, runs the tests, fixes what broke, and hands you back a pull request. That is not a faster typist. That is a teammate who works while you sleep.
What does AI actually do to an engineering team?
It changes who can be on one.
The number I keep coming back to is from OpenAI's own research, published this summer. Their engineers now generate 99 percent of their output through Codex rather than through chat. That alone is remarkable. But the number that actually matters, the one that describes my company, is this: between November 2025 and June 2026, Codex usage on OpenAI's Customer Support team rose 32 times. Their Legal, Finance, and Recruiting teams crossed over to Codex as their primary AI tool around April.
Read that again. The support team. The lawyers. The recruiters. All of them shipping technical work.
That's not a story about engineers getting faster. That's a story about the wall between technical and non-technical falling down. And it's exactly what happened inside my own company. I have people who came in doing customer experience who are now building code and building product. Not because I retrained them. Because the barrier that used to keep them out disappeared, and they walked through the hole in the wall.
The most useful person on a small team in 2026 is not the person with the deepest technical credential. It's the person who understands the customer's problem most precisely and has the judgment to specify a solution. That person used to hand their understanding to an engineer and lose 30 percent of it in translation. Now they hand it to an agent and keep all of it.
Why can't founders predict timelines anymore?
Because the tool is improving faster than the plan can survive.
This is the part I find genuinely disorienting. I don't know timelines anymore. They are literally a figment of my imagination. I say a feature will take six weeks, and then a model ships, and it takes four days. Or I say something is impossible with current tools, and two weeks later it is not only possible, it's easy. Every time I think I know something, AI or Claude says nah.
And I want to be honest that this is uncomfortable. Founders are supposed to project confidence about timelines. Investors ask. Teams ask. Customers ask. The honest answer right now is that anyone giving you a confident twelve-month engineering roadmap in this environment is either not paying attention or telling you what you want to hear. The ground is moving too fast for that kind of certainty to be anything but theater.
So I've stopped pretending. I plan in short horizons. I commit to direction rather than dates. And I've made peace with the fact that half of what I think I know about what's hard and what's easy will be wrong by the time we get there.
What should founders and operators actually do about this?
Stop planning around the tools you have today.
The headcount plan you wrote last year assumed a set of constraints that no longer exists. The role definitions you're hiring against assume a wall between technical and non-technical that's coming down. The roadmap you committed to assumes a rate of change that has already been exceeded.
The move is not to make a better plan. The move is to build a team that adapts faster than the plan breaks. Ten people who use these tools daily, who have the reflex to try the thing that was impossible last month, who don't need permission to reach for an agent, will beat a hundred people executing a plan written in a world that no longer exists.
I thought I needed 100 people. I need ten. That's not a cost-cutting story. It's a story about what a small group of people can now reach, and how badly the old math underestimated them.
The full conversation is on the Fargo Talks podcast with Jeff Fargo.
*Judd Hoffman is CEO and Co-Founder of Ethica AI, building AI-powered tools for real estate transaction workflows.*