Everyone agrees agentic AI is necessary. Nobody agrees on anything else.

I went looking for one number this week: how many enterprises actually have an AI agent running in production in 2026. I expected a range. What I found was chaos.

One report says 11% of enterprises run agents in production. Another says 51%. A third says 80%. A fourth splits the difference and says 31%, except in banking, where it’s supposedly 47%. These aren’t old numbers from different years catching up to each other — they’re all dated 2026, all claiming to describe the same industry, the same moment.

That’s not noise around a real number. That’s five different definitions of “production” wearing the same word.

The part everyone actually agrees on

Here’s what’s strange: underneath that chaos, there’s real consensus on the thing that matters more than the headline number. Nobody serious argues anymore about whether agentic AI needs to be trustworthy before it touches anything real. Two years ago, an agent that impressed in a demo was the whole pitch. Now the pitch is different: can this thing run unattended, fail safely, and let a human step in before it does something that can’t be undone?

That’s real progress, and it’s not marketing. It shows up in the architecture. Frameworks like LangGraph forced developers to stop letting the model freelance through an entire task and instead define the actual flow — the routing, the retries, the state — as ordinary code, with the model doing reasoning only at specific, bounded points. It shows up in what a “production-ready” agent is now expected to have: a way to save its state and resume after a crash, a full trace of every decision it made so a human can find where it went wrong, and an explicit pause built in before anything consequential — sending an email, moving money — happens without a person saying yes.

None of that is flashy. None of it demos well. And that’s exactly the point — it’s infrastructure, not a feature, and infrastructure is what “necessary” actually looks like once the flashy phase is over.

So why can’t anyone agree on the numbers?

My guess: “in production” got useful as a marketing claim before it got useful as a measurement. A vendor counting “at least one agent embedded somewhere in a shipped feature” and a CIO counting “an agent I’d trust to run at 2 AM with nobody watching” are answering two completely different questions with the same three words. The first number is almost certainly the 80% one. The second is probably a lot closer to the 11%.

I’ve spent the last year hands-on with this — fine-tuning models, building agent pipelines, watching where they actually break versus where the demo made them look solid. The gap between “it worked in the demo” and “I’d trust it unsupervised” is the real story right now, and it’s a gap no survey percentage captures cleanly, because half the industry is still answering the easy version of the question.

If you’re evaluating a vendor’s “production-ready” claim this year, ask them which version of that question they’re actually answering. It’s usually not the hard one.

This is the first entry in The Long View*‘s Technology section — ongoing essays on what’s real versus what’s marketing, as I keep digging into this.*