A CFO Lets AI Build His Software

Earlier this year I wrote about building with AI agents. Since then the numbers have stopped being a party trick.

This website, its CMS, the image pipeline behind it. All of it built with AI agents in the evenings and on weekends by someone whose day job is CFO.

I want to be clear about what that actually means. The claim that anyone can build software now is true, but also incomplete in a way that matters.

What actually got cheaper

For twenty-five years I’ve watched companies make the same calculation: build or buy. Building custom software was almost always the wrong move for a company our size. Developers cost a lot, projects run late, and the maintenance never ends. So you rent. You pay per seat, every month, forever. As CFO I’ve signed more of those invoices than I care to remember.

AI agents changed the cost side of that equation. The code itself is now close to free. Something a small team would have quoted as six months of work shows up as a working prototype before lunch. The math that made “buy” the default answer for decades just stopped adding up.

What did not get cheaper

Judgement. Knowing what is worth building in the first place. Knowing when the output looks correct but the reasoning is off. Having the taste to leave things out.

The agents write the code. I still own every decision. Maintenance didn’t disappear either. It just changed shape. I don’t maintain syntax anymore. I maintain intent. That’s a different skill, and in some ways a harder one. It’s also the skill you actually pick up from years of running things.

The investor in me can’t unsee it

This is where my two jobs collide. If a family company can build its own internal tools over a weekend, what happens to the vendors it used to rent those tools from? Now multiply that across every small and medium company.

I’m not saying SaaS is dead. That’s as lazy as the hype on the other side. Payroll, compliance, anything where the real product is liability and trust — we’ll keep paying for those. You can’t prompt trust into existence.

Critical systems, the ones where you simply cannot afford downtime, won’t be replaced tomorrow either. The bar for reliability and operational safety is still higher than what most agent setups deliver right now.

But things are moving so fast that I’m not willing to bet those systems are safe even for the next two years. The quality of the output is already better than most humans in a lot of areas. Who knows.

When I look at a SaaS valuation now I ask one new question: what happens to this revenue the day a customer’s fourteen-year-old can rebuild the product in a weekend? Some companies have a real answer. A lot of them only have a slide deck.

The part nobody posts about

The tools change every week. Half of what I learned about one platform in March was already outdated by summer. If you chase the tools you’ll always feel behind, because you are.

So I try to focus on the only thing that actually compounds: knowing what is worth making. It’s slow, it’s not glamorous, and it’s the same skill that builds companies and waits out markets.

The agents can have the syntax.
I’ll keep the judgement.

Interesting times to be alive.
The long game is still the only one I know how to play.