The enterprise software market is quietly cannibalizing itself. Not through price wars or open-source disruption, but through a shift so fundamental that most analysts are still describing it with outdated frameworks. I spent the last week dissecting McKinsey's latest report on agentic coding tools, and the numbers are stranger than the marketing suggests.
Let's start with the headline stat: 32% of organizations are now choosing to build custom software with agentic AI rather than buy off-the-shelf products. But that's not the interesting part. The interesting part is what happens when you cross-reference that figure with the success rates. MIT's NANDA research shows internal builds succeed only 33% of the time, while purchasing vendor tools succeeds at 67%.
That gap is the story. It's not about whether AI can code. It's about who gets to capture the value when it does.
Context: The Agentic Coding Landscape
Agentic coding tools are LLMs wrapped in a loop: plan, call tools, generate code, run tests, fix errors, repeat. Architecturally, they're compositional innovations—not a new model paradigm, but a new orchestration layer. Deloitte's 2026 Tech Trends report puts production-ready agentic systems at just 11%. Gartner's CIO Survey shows only 17% of organizations have actually deployed agents in production. Yet 75% claim to be adopting the technology.
This gap between adoption intent and production reality is the defining tension of this cycle. The tech industry is notoriously bad at distinguishing between pilots and deployments. I saw the same pattern during the DeFi Summer of 2020: everyone was claiming massive TVL, but when you actually looked at the liquidity distribution, most of it was concentrated in a handful of pools. The same dynamic is playing out here.
The cost structure is the silent bottleneck. McKinsey notes that 20% of organizations are already feeling AI operational cost pressure. An agentic coding workflow can trigger dozens or even hundreds of LLM calls per task. My own simulations of agent-based systems during my research sprint on AI-agent economies showed that a single complex coding task could consume 10-100 times more tokens than a standard chat interaction. That's not a linear scaling problem; it's an exponential one.
Core: The Build-vs-Buy Rebalancing as a Macro Signal
I've been modeling this through the lens of liquidity and capital flows. The adoption data splits into two distinct regimes. High performers—companies with at least 5% of EBIT from AI—are nearly 50% more likely to skip buying software and build internally. Large enterprises are expanding agents at 40% (up from 27% last year). Meanwhile, regular organizations are more likely to fail at internal builds and should stick with vendor tools.
The sector distribution is revealing. Tech leads at 41%, healthcare at 39%, professional services and energy at 38%. These are all knowledge-work-intensive industries with highly customized workflows. They're not adopting agentic tools because they're tech-forward; they're adopting because off-the-shelf SaaS has failed them for years. The AI coding tools are simply the first viable alternative.
This is where I see the clearest parallel to the crypto landscape. When Ordinals injected new narrative and fee revenue into Bitcoin in 2023, it wasn't because people suddenly wanted NFT-like artifacts on the oldest blockchain. It was because Bitcoin's security model needed the fee pressure relief valve. The narrative was secondary; the economic necessity was primary. Same thing here: the "agentic coding" narrative is secondary to the economic necessity of reducing software procurement costs.
The infrastructure angle is where the real value lies. High performers are building internal model fine-tuning, evaluation, and observability stacks. They're not just using APIs; they're building systems around them. The technical barrier has shifted from model capability to systems engineering capability. Code repository semantic understanding, CI/CD integration, secure sandboxing, failure recovery, human review loops—these are the new moats.
Contrarian: The Decoupling Thesis
The conventional reading of the MIT NANDA data is that vendor tools are better and you should buy. That's wrong. The 67% success rate for vendor tools includes traditional software purchases, not just agentic coding tools. The comparison is methodologically muddy. But the direction is clear: building your own agentic systems is harder than buying them.
Here's the counter-intuitive twist: the high performers who choose to build aren't doing it because vendor tools are bad. They're doing it because they want the code to stay inside their perimeter. Privacy and security concerns. Once you send your proprietary codebase to a third-party LLM, that code has left your control. The build-vs-buy decision is increasingly a data sovereignty decision.
And this is where the Gartner prediction of 40% of agentic AI projects being canceled becomes a feature, not a bug. The market will be flooded with projects that fail at the pilot stage because they underestimated the infrastructure costs and complexity. That failure rate creates a massive opportunity for a new layer of services: evaluation platforms, security governance, observability, and cost optimization tools.
Meanwhile, the employee layoff expectation is rising—39% expect layoffs in the next year, up from 32%. That's a governance risk that compounds over time. Employees who fear for their jobs are less likely to share knowledge, more likely to resist knowledge transfer, and the internal build projects that depend on tacit knowledge will fail at even higher rates. Code never lies, but it does omit. The omitted parts are the context, the tribal knowledge, the unwritten requirements that only senior engineers carry in their heads.
Takeaway: Positioning for the Agentic Cycle
The market is pricing agentic coding tools like they're the final evolution of software development. They're not. They're the first iteration of a new category that will bifurcate into commodity tools and premium platforms. The value won't accrue to the tools themselves; it will accrue to the infrastructure that makes them enterprise-ready. Tracing the fault lines before the quake hits means looking at the cost curves, the success rates, and the data sovereignty demands.
Liquidity is just patience disguised as capital. The 32% of enterprises betting on agentic coding tools are the early positioning of a much larger rebalancing. The narrative shifts, but the leverage remains. I've audited enough failed ICO smart contracts to know that the ones that survived weren't the ones with the best marketing. They were the ones with the clearest understanding of the economic incentives. The same principle applies here.
Chaos is the only constant variable. The next 18 months will separate the platforms that can handle the messy reality of production workloads from the demos that look good in a keynote. And for the developers, the enterprises, and the investors who understand that the real bottleneck is not model capability but systems discipline—the opportunity is just beginning.