In my meanderings with Claude, I've started using a term I'm increasingly convinced describes where most organizations actually are with AI. I don't know if I invented it. Maybe I just put words to an idea I'd been circling since the Bard days.
A lot of digital ink has been spilt on where this is going. The Frontier is amazing, and gets press it deserves. But where are the rest of us?
I think we are in what I call “The Dangerous Middle”.
People say it’s good to fail fast. With AI, it’s only good to fail fast if you realize you failed.
I built a small tool to score and improve the prompts and context files I share — a library with a quality rubric and an improve function running on a local model. During testing I fed it a vague status document and asked for an improved version. What came back was better in every visible way: cleaner structure, tighter language, more specific. It was also wrong. The model had invented the specifics. The test-case I was running was a blank… yet the AI just inserted a good answer.
I caught it in seconds, but only because I knew the content of the original and knew what wasn't in it. Take away that proximity and the output just passes through. It was plausible, well-worded, and formatted like something a careful person would produce.
That gap has a name in my notes: the dangerous middle.
Picture the two ends first. On one end sits the pre-AI enterprise: slow, manual, and full of swivel-chair work — but with a person attached to every number. On the other sit validated AI systems: constrained, scoped, and built with verification. The dangerous middle is everything in between: general-purpose AI answering enterprise questions with no validation between the model's output and the decisions it drives.
The middle is where most organizations are today because it's the easiest place to reach. A chat window takes minutes. A validation layer takes real work.
What makes the middle dangerous is the false authority of natural language. A wrong number in a spreadsheet at least looks like a number — it invites checking. A wrong answer in fluent prose looks like an answer. We spent entire careers using fluency as a proxy for competence because, for humans, it mostly worked.
Language models break that heuristic the same way they break time as a measure of effort — completely, instantly, and without announcing it. The output isn't wrong in a way that LOOKS wrong.
My own answer so far doesn't scale, and I want to be honest about that too. I do the last mile by hand. Anything AI-assisted that leaves my desk gets walked, checked, and owned before it moves — and that costs me real hours every week. I'm not describing a virtue. I'm describing a tax I can afford because the volume is still low, and one that collapses the moment a thousand people are pasting fluent outputs into decision decks. Personal diligence is not an operating model.
How to build the NEXT process is the operator's actual job: putting a layer between the model and the decision.
- Constrain the tools — a purpose-built app with defined inputs and outputs has fewer ways to be wrong, and the ways that remain are visible.
- Repoint the reporting — dashboards stop being where work happens and become where you verify the work happened correctly.
- Ask where it came from — make "where did this come from?" a routine question before the number matters, not after it fails.
None of this bans the middle, because you can't. People will use the chat window. The job is making sure the path from its answer to a decision runs through something that checks.
The middle isn't a phase that passes on its own. Better models reduce the number of mistakes. They don't eliminate the need to know when you're looking at one. Someone has to build the checking into the system. That someone sits between the strategy and the tools, close enough to the work to know what wrong looks like. If you're in that seat, this is the part of the job nobody is going to hand you a mandate for. Take it anyway.