The starkly uneven reality of enterprise AI adoption

Paraphrasing William Gibson, the future of AI is here, but it’s nowhere close to evenly distributed yet.

Last week in London, I had two conversations about enterprise AI that obliterated any semblance of a neat and tidy story of AI adoption. In the first meeting, the head of engineering at a large hedge fund told me about engineering teams with fleets of agents in full production, and (in his personal case) all code is written by LLMs. (Junior hires, interestingly, aren’t allowed to use LLMs for code assistance.) In another meeting, a data engineer at a large retail bank described the exact opposite: No agents and sparse use of LLMs. Maybe other parts of the bank are moving faster on AI, but his division clearly isn’t.

This isn’t about one company “getting” AI and the other not. Rather, it’s a reminder that even within the same company there are wildly divergent adoption curves for new technologies. AI is widening the gap between teams that can absorb it operationally and teams that can’t. That’s what the best recent data suggests, too. McKinsey found that 88% of respondents say their organizations are using AI in at least one business function, but only about one-third say their companies have begun scaling AI programs. As for agents, 23% report scaling an agentic AI system somewhere in the enterprise, while 39% are still just experimenting. And in any given function, no more than 10% say they’re scaling agents.

Broad usage, in other words, is not the same thing as deep institutional change. In short, there’s still time to figure out AI. You’re not behind.

Cue the engineering boom

I keep hearing that “finance is cautious” or “regulated industries are behind” or “everyone is building with agents.” None of that is quite true. Some financial firms are moving aggressively. Some aren’t. Some teams inside the same firm are doing both at once. Deloitte’s 2026 enterprise AI research makes the same point from another angle. Only 25% of respondents said they had moved 40% or more of their AI pilots into production. Just 34% say they’re using AI to deeply transform their businesses (a number I suspect is more aspirational than actual), while 37% are still using it at a surface level with little or no change to core processes. That sounds a lot less like a tidal wave and a lot more like a messy, uneven organizational test.

Same as it ever was, right?

And that, in turn, is why I think a lot of the “AI will wipe out software jobs” talk is wrong and misses the point. The interesting thing about AI coding tools isn’t that they make software cheaper to produce. It’s what companies do with that lower cost. Box CEO Aaron Levie recently invoked Jevons paradox to explain exactly this dynamic: When a capability becomes cheaper and easier to consume, demand for it often rises rather than falls. That’s not a law of nature, but it is a pretty good description of what technology has done for…ever. Cloud computing didn’t lead companies to need less compute. It made them build more things that consumed compute. AI-assisted coding may be doing something similar for software itself.

This is where the data on engineering jobs gets interesting. Lenny Rachitsky recently highlighted that engineering openings are at their highest levels in more than three years. The underlying TrueUp data shows 67,665 open engineering jobs as of March 2026, up 78.2% from the recent low. More importantly, this isn’t just concentrated at the very top end of the market. TrueUp’s breakdown shows 44.6% of posted engineering roles within tech companies are entry and mid-level, versus 38.3% at senior level and 13.8% at senior-plus. So no, the data doesn’t say AI is eliminating roles for junior developers; rather, it says companies still want a lot of engineers, even as AI tools spread throughout the profession.

There’s a cleaner way to understand what’s happening. AI isn’t killing the need for engineers. It’s changing what enterprises want from engineers.

Stack Overflow’s 2025 survey found that 84% of respondents are using or planning to use AI tools in development, and just over half of professional developers use them daily. McKinsey’s software development research found that the highest-performing AI-driven software organizations are seeing 16% to 30% improvements in productivity, customer experience, and time to market, along with 31% to 45% improvements in software quality. But McKinsey’s crucial point is that these gains don’t come from sprinkling copilots over an unchanged process. They come from reworking roles, workflows, and the full product development system. That’s a much harder organizational challenge than buying licenses for a coding assistant.

Software engineering is alive and well

Let’s go back to my conversations in London. The hedge fund leader may be an early glimpse of where parts of enterprise engineering are headed. Less time hand-authoring code, more time specifying, reviewing, steering, and orchestrating systems that increasingly generate code for you. But that does not mean the retail bank division is irrationally lagging. In a heavily regulated environment, code generation is not the hard part. Governance is. Deloitte reports that only 21% of surveyed companies currently have a mature governance model for autonomous agents (and those 21% are probably kidding themselves). At the same time, 73% cite data privacy and security as a top risk, and 46% cite governance capabilities and oversight. That’s not bureaucracy for its own sake. It’s a recognition that plugging non-deterministic systems into deterministic, compliance-heavy environments gets messy fast.

Still, caution isn’t free. Every quarter a team spends in pilot mode is a quarter in which more aggressive peers are building operational muscle. OpenAI’s enterprise usage data is useful here because it shows how uneven that muscle-building already is. Frontier workers, defined as the 95th percentile of adoption intensity, send six times more messages than the median worker. Frontier firms send twice as many messages per seat. OpenAI says the primary constraints are no longer model performance or tools, but rather organizational readiness and implementation.

This rings true to me. In my experience, the real divide is increasingly not between companies that have access to AI and those that don’t. It’s between teams that have learned how to integrate AI into repeatable work and teams that are still treating it as a promising but dangerous sideshow, as I’ve written.

This is also why I think the distinction of task versus job matters. Writing a chunk of boilerplate code is a task. Engineering is a job. Jobs bundle judgment, trade-offs, accountability, architecture, security, integration, testing, and the ugly reality of operating systems in the real world. AI can automate more tasks, but it hasn’t eliminated the need for jobs, especially in environments where bad software decisions carry real operational or regulatory consequences. In fact, McKinsey’s broader AI survey found that most organizations are still navigating the transition from experimentation to scaled deployment, and that high performers stand out precisely because they redesign workflows and treat AI as a catalyst for innovation and growth, not just efficiency. That is a very different thing from saying, “We gave everyone a chatbot and now we need fewer people.” (By the way, that would be a very naive statement.)

So no, AI isn’t plodding (or rocketing) toward one uniform enterprise future in which software engineers quietly fade away. Instead AI is splitting enterprises into fast-learning and slow-learning teams and is rewarding organizations that redesign work, govern risk, and turn lower software costs into more software, not less. The code may be getting cheaper, but the ability to decide what should be built, how it should fit together, and how to keep it from breaking the business keeps increasing in value.

That’s not the death of software engineering. It’s the repricing of it, and every company and every team is paying different prices.