The board-engineer gap is the story of the year
The headline finding from a recent closed engineering retreat is not new, but the framing is sharper than before. Boards are chasing AI productivity. Engineers are increasingly worried about what citizen developers are shipping. The mismatch is becoming structural.
The retelling that nailed it: a company trained ML models to optimize air filter replacement on field equipment. The models had been trained on equipment in the desert, where the issue is dust. The company operated in the arctic, where the issue is mosquitoes. The model saved them 100 million.
You don't need AI to make that mistake. Humans have done it forever. But the speed at which AI scales a wrong assumption is the new variable.
Vibe coding as shadow IT 2.0
The same dynamic shows up in software. Citizen developers are now vibe-coding internal apps at a pace that traditional IT can't keep up with. The pattern is familiar: spreadsheets run the business, nobody governs them, and one day a CFO makes a decision on a model that a junior analyst built in 2019. Vibe coding amplifies the same risk with a much wider blast radius.
The companies handling this best are the ones that put vibe-coded apps in separate infrastructure with deterministic controls on data access, not the ones that try to ban the practice. Some are running threat modeling sessions with board members to make the risks tangible. The ones doing worst are the ones that encouraged vibe coding, got a surge of useful tools, and now have a shadow IT problem they can't see.
A $50 million productivity assumption
Part of the problem is that LLMs are genuinely good at the kind of work managers consume. Summarizing reports, drafting memos, preparing slide content. The work looks competent, so managers extrapolate: if LLMs are good at this, they must be good at programming too. The gap between those two conclusions is where a lot of enterprise AI strategy goes wrong.
The legal department tends to disagree. Lawyers are early to notice that LLMs are confidently wrong in ways that matter. If your organization has a general counsel who has used an LLM for contract review, they probably have a more realistic view of the technology than your CIO does.
A bubble, but how big and how long
Almost everyone at the retreat agreed this is a bubble in the technical sense: the gap between deployed value and market valuation is wider than the underlying technology will support. The dotcom comparison is unavoidable. The historical fact is that the bubble was clearly recognizable as such by 1995, three years before it actually popped. The current cycle has produced fewer genuinely new products, not more, which suggests the value at the bottom will be more concentrated than the 1990s version.
Agent code is fragile, in specific ways
Two patterns keep showing up when agents write code. First, they insert features nobody asked for. A team spent three days tracing an unrequested feature before realizing the agent had hallucinated a requirement. Second, agents estimate incident resolution as a linear process and quietly miss all the parts that aren't. Agents are useful for repeated incidents where the data already exists. They are dangerous for novel ones where the data is the answer.
LLMs are weirdly good at DSLs
A line of work worth watching: domain-specific languages designed to be LLM-friendly. Token efficiency, hard security boundaries, and a deterministic compiler that translates high-level intent into safe code. Spender Nelson's recent piece on this captures the shape of what's coming. The biggest barrier to building a DSL has always been the parser and the tooling. LLMs eat that barrier for breakfast. The semantic model still has to be designed by a human, but the surface that humans have to learn is getting thinner.
The LLM-speak problem is getting worse
Reader confession: I have started physically recoiling from prose that smells like an LLM. It's not just the usual tells ("in today's fast-paced landscape"). It's a miasma. A syntactic flatness. The LLM has a house style and that style is now everywhere. I find myself abandoning articles I would have finished a year ago, because my brain flags them as AI before I get to the argument.
The implication is professional. If your reader can detect LLM-speak in your writing, your writing loses credibility before you get to make your point. Read your drafts out loud. Speech patterns are closer to your real self than your typing patterns. That gap is the only defense right now.