The Build-vs-Buy Reckoning: Why 33% Success Rates Are Reshaping Enterprise Software

The silence between the code and the chaos is getting louder. Over the past quarter, I have watched a strange divergence form in the enterprise software market: adoption curves pointing straight up, while production readiness charts barely move. The narrative is the only immutable ledger, and right now, that ledger tells a story of profound imbalance. McKinsey reports that 32% of enterprises are now choosing to build custom software using agentic coding tools instead of buying off-the-shelf products. Among high performers—firms generating at least 5% of EBIT from AI—nearly half are skipping traditional software purchases entirely. Yet Deloitte's 2026 Tech Trends report finds that only 11% of agentic systems are production-ready. Gartner's CIO Survey shows just 17% of organizations have actually deployed agents. The gap between intent and execution is not a gap; it is a canyon. I have spent the last eighteen years mapping this terrain, from the ICO wild west to the DeFi summer to the long bear market solitude. In the wild west, stories are the only compass. The story here is that agentic coding tools—LLM-driven systems that plan, call tools, generate code, execute tests, and self-correct—represent combinatorial innovation, not paradigm shift. They work well on small, well-defined tasks. They stumble on complex, multi-file, legacy codebases. The architecture is sound; the reliability is not. MIT NANDA's research delivers the sharpest data point: internal build success rates hover around 33%, while purchasing vendor tools succeeds at roughly 67%. This is the hidden truth beneath the adoption headlines. Based on my audit experience across dozens of enterprise pilots, the failure mode is rarely the model itself. It is the surrounding infrastructure—semantic understanding of code repositories, CI/CD integration, security sandboxing, failure recovery, human review loops. The technical barrier has shifted from model capability to systems engineering. Gartner predicts that 40% of agentic AI projects will be cancelled by the end of 2027, citing cost overruns, unclear business value, and insufficient risk control. The cost dimension deserves attention. A single agentic coding task can trigger dozens or even hundreds of LLM calls. McKinsey notes that 20% of organizations already feel AI operational cost pressure. One of their senior partners frames it directly: the most successful organizations treat operational cost as a design constraint. This is not a warning; it is an epitaph for projects that ignored unit economics. The contrarian angle cuts against the prevailing narrative of AI-driven productivity. The real winners in this cycle will not be the companies building the flashiest coding assistants. They will be the infrastructure players—observability platforms, evaluation systems, security governance tools, private deployment solutions—that help enterprises fail less. The 33% internal build success rate is not a condemnation of self-building; it is a market signal. High performers are not building from scratch. They are assembling combinations of model APIs, development frameworks, and cloud-native infrastructure. They are purchasing the ability to build, not the finished product. This reframes the competitive landscape. Cloud giants like Microsoft, AWS, and Google benefit from increased compute consumption. Model API providers like OpenAI and Anthropic capture value from token usage. But traditional SaaS vendors face structural compression. When 32% of enterprises decide to build instead of buy, the functional moats of packaged software erode. The new moats are data and workflow integration, not feature checklists. The industry distribution tells a deeper story. Technology leads at 41%, followed by healthcare at 39%, professional services and energy at 38%. These are knowledge-work-intensive sectors with highly customized workflows and strict compliance requirements. Off-the-shelf SaaS cannot satisfy their needs. Agentic tools let them build compliant internal systems at lower cost. This signals the resurgence of shadow IT, but with an AI-native twist. Security concerns compound the complexity. Sending proprietary codebases to third-party LLMs creates data exposure risks. Agents that autonomously modify code can become vectors for prompt injection or dependency poisoning. The 39% of employees expecting layoffs—up from 32%—adds organizational resistance to the equation. Failed internal builds do not just waste development time; they erode morale and trigger a negative feedback loop that makes subsequent attempts even harder. Truth hides in the bear market's quiet shadows. The current market is not a bear market for AI coding tools; it is a bear market for naive execution. The enterprises that will thrive are those with discipline—treating operational cost as a design constraint, choosing build versus buy selectively, and investing in the systems engineering that makes agentic reliability possible. I hunt for the story that the data cannot speak. The data says 32% are building. The silence between the numbers says most of them are not ready for what building actually requires. The next narrative cycle will not be about model intelligence. It will be about deployment maturity. The question is not whether agentic coding tools will reshape enterprise software. They already are. The question is which enterprises will survive their own ambition.