Track · AI in Product4 min read
AI-generated code breaks the build-versus-buy math for CIOs
AI coding tools are upending the build-versus-buy call, but METR found AI made devs 19% slower and Veracode found 45% of AI code has flaws. CIOs should re-run the math on TCO, not demos.

Context
METR's 2025 randomized controlled study found experienced open-source developers using AI tools took about 19% longer on tasks despite expecting speedups.
Veracode's 2025 report found 45% of AI-generated code samples introduced security vulnerabilities.
Enterprises commonly run hundreds of SaaS applications with overlapping capabilities and rising subscription costs.
The article proposes five renewed build-versus-buy criteria centered on total cost of ownership.
The same AI economics apply to SaaS vendors, raising renewal-time questions about whether efficiencies reach subscription prices.
A 2025 randomized controlled study by METR found that experienced open-source developers using AI tools took roughly 19% longer to complete tasks on familiar codebases — despite expecting the tools to make them faster. That number should anchor every build-versus-buy conversation CIOs are about to have, because AI is genuinely rewriting the economics of enterprise software, just not in the straightforward way the demo suggests.
The old logic favored buying. SaaS delivered mature capabilities, ongoing enhancements and support, while custom builds demanded heavy investment in developers, infrastructure, testing, security and maintenance. Now, with AI-powered development tools, applications that once took weeks or months can sometimes be prototyped in hours or days. Even employees with limited programming experience can describe what they want and get a working solution — the practice known as vibe coding.
But as the argument goes, AI is reducing the cost of creating software. It has not eliminated the cost of owning software.
Why the SaaS math is shifting?
Many enterprises now run hundreds of applications — some overlapping, some barely adopted, all with subscription costs that keep climbing. Business units buy narrow tools that create integration and governance headaches. AI-assisted development makes replacing selected applications, consolidating overlapping functionality or building targeted capabilities around existing platforms economically plausible for the first time in years.
The METR study, caveat noted, focused on one development environment and doesn't prove AI slows development generally. It does prove something CIOs keep forgetting: generating code faster doesn't equal delivering reliable software faster. Measure actual productivity before assuming savings.
What breaks when you replace SaaS with vibe coding?
The failure modes are concrete:
- Security. A 2025 Veracode report on AI-generated code found 45% of samples introduced security vulnerabilities. Enterprise apps still need authentication, authorization, encryption, vulnerability management and data protection — accountability that stays with the enterprise regardless of how fast the app was built.
- Maintenance. SaaS bundles patches, enhancements, infrastructure and support. Build it yourself and those become internal obligations for the application's entire life.
- Shadow IT. When an employee's vibe-coded app becomes essential to a department, who maintains it when they leave? The question sharpens when the app touches financial transactions or customer data.
- Integration. A "simple" SaaS app may connect to ERP, CRM and identity systems; replacing it means rebuilding integrations, migrating data and retesting workflows at scale — thousands of users behave differently from a few.
The relevant measure is total cost of ownership, not initial development cost. Vibe coding may make software easier to build; it doesn't make software easier to own.
How should CIOs decide now?
Don't indiscriminately replace major platforms. Mature ERP, HR, payroll and financial applications hold years of accumulated business rules and compliance capabilities — recreating them invites risk even with faster development. The near-term opportunity sits in narrow, non-critical SaaS apps that duplicate platform capabilities or could be built cheaper long-term.
Five questions should replace the five-year-old criteria:
- Is the capability strategically differentiating?
- How complex is the application — simple workflow or mission-critical transaction system?
- What is the true total cost of ownership, including subscription fees, integration, security, maintenance and internal operating resources?
- Do we have the architecture standards, secure development practices and operational support to sustain what we build?
- Could we get the same outcome by extending platforms we already pay for but underuse?
That last question matters most: the best alternative to buying another SaaS app may not be building one, but using what you already own. CIOs should also separate AI-assisted professional development (governed by engineering discipline) from business-user vibe coding (governed by defined boundaries) — same tools, different oversight.
Will SaaS prices fall?
One more lever: vendors have the same AI capabilities. If AI cuts their development and support costs, CIOs should ask at renewal whether those efficiencies will show up in subscription prices — or whether vendors will keep raising them while layering on AI features and new licensing models.
The likely endgame is hybrid: keep core platforms, build targeted workflows and AI agents at the edges. The question is no longer whether you can build cheaper than you can buy — it's which path delivers the best business outcome over the application's whole life.
via cio.com (Source)
More from Daniel Okafor
Show full bio
Staff writer covering technology and product news in product management at Roadmap File.
60 articles