The gut-check questions every leader needs to ask before building with AI

A new pitch is spreading through enterprise tech: skip SAP, Workday, HubSpot and the rest, and use AI tools like Claude Code to build your own software. The appeal is obvious. If a company already has engineers, why not make the app itself and keep the subscription money in-house?

The problem is that “can build” and “should build” are very different questions. The writer remembers a custom analytics dashboard project from his CIO days at a beverage company. The team had the talent and the backing, but after four months of setbacks, it became clear the project was taking time away from work that could have delivered more value elsewhere. They shut it down and bought software instead.

That same logic sits behind the first test he gives CIOs now: is this your core competency, or just a shiny distraction? He points to a global technology company with hundreds of thousands of employees that chose not to build its own billing and HR tools, even though it certainly had the engineering muscle to try. The point is blunt. High-tech companies still have to decide whether custom software is actually part of the business, or just a side quest.

The money question is tougher than the sales pitch suggests. Cutting a subscription can save millions a year in some enterprises, but development is only the beginning. He says companies building their own LLM-based software typically spend five to 10 times more than they would on his company’s workflow automation platform, once development, maintenance, security, and upkeep are counted. He also cites a large financial services company where rewriting a core system with an LLM could take 18 to 24 months just to save 0.5% of annual operating budget.

And then there’s the part people love to wave away until it breaks: security, governance, and leadership turnover. He says roughly half of organizations have seen AI agents exceed their permissions, and cites PocketOS, where a coding agent wiped a production database in nine seconds. He also asks CIOs to think about what happens when the person who designed the system leaves, because average CIO tenure is about 4.5 years and custom systems can take two or more years to build. AI makes building easier. It does not make judgment optional.