Tokens Are Cheap. Your Time Isn't.

The cost of writing software is collapsing. The cost of owning it is not.

Much of the current argument about AI and software begins with a simple claim: if a founder can build an internal tool over a weekend, why keep paying for SaaS? The premise is plausible. The conclusion is not.

For a founder-CEO, build cost has never been the constraint. Opportunity cost is. Tokens are cheap; the weekend spent building the tool is not. Nor are the years that follow: the security patch, the deprecated API, the integration that quietly breaks, the permissions model that worked at 20 employees and fails at 200, and the person who eventually has to maintain code nobody on the team really designed.

Generally speaking, my instinct runs toward building. Across two CEO roles and a CDO role I brought systems in house wherever they differentiated us, because enterprise software cost millions to install, millions to maintain, and still wouldn't do what we needed. That instinct has a limit, and it isn't set by what you can build. It's set by what you're willing to stop doing to build it.

Recently I built a piece of software encoding a hiring methodology I've used for years. Before showing it to founder-CEOs — most of whom write better code than I do — I was reluctant. Wrongly. What came back wasn't that they couldn't build it. It was that they could, and never would. One had started building a hiring app himself and never found the time to finish it. He didn't decide against building it. Nothing stopped him. It lost to pipeline, then to a hire, then to the raise — every week, on the merits, correctly.

That is one founder. Run the arithmetic across the economy. Millions of companies each running their own CRM write the same feature millions of times, then write every subsequent patch, integration fix and regulatory change millions of times as well. That is the same work performed in parallel, permanently.

A central economic advantage of a software vendor is amortization. A vendor builds a capability once, secures it once, maintains it once, and spreads those costs across thousands of customers.

That economics is also why enterprise software is shaped the way it is. On-premise software could be customized — expensively, and at the cost of every upgrade that followed, but the source was yours to bend. SaaS took that off the table. One codebase serves every customer at once, so bending it to any single business was never an option. What replaced customization was settings. Turn this module on. Turn that one off. Rename the field. Ignore the 80% that doesn't apply. Configuration is what you do when variants are impossible.

That constraint is what AI relaxes. Not who produces the software, but how many versions the producer can afford to maintain.

I can see a world where vendors leverage AI to support highly customized solutions for their customers. Picture a CRM vendor that still owns the common architecture: the data model, security, integrations, compliance, reliability, infrastructure and upgrade path. The product each customer experiences is no longer identical. Workflows reflect how that company actually sells. The interface changes by role. Reports are generated around the metrics management cares about. Automations fit the company's approval process. Terminology matches the business rather than forcing the business to adopt the software's language.

The vendor still owns the software. The customer gets something much closer to bespoke. That is a different future from every company building its own tools. It is central production with personalized delivery. For decades enterprise software forced a choice between fit and efficiency. Bespoke software fit the company and was expensive to own. SaaS was cheap to own because the vendor standardized the product and spread the cost. AI separates those two things: the vendor preserves the economics of central ownership while allowing far more variation at the edge.

The first era of SaaS made software cheap by standardizing it. The next era may make it cheap without standardizing it.

AI kills configuration before it kills SaaS.