The Specification Bottleneck

Jamie runs dispatch at an HVAC shop outside Greeley. Twelve years in the chair. Knows which technician likes which truck. Knows which ZIP codes on the east side of town are actually in the county — the ones that look residential but get routed to commercial crews. Knows that Tuesdays are commercial restock day, and that’s why you never, ever book a residential call for a Tuesday.

The owner bought her an AI agent six weeks ago. For after-hours lead intake. Picks up the phone when she’s off the clock. Takes the customer’s name and address and books them a slot.

And every Monday morning, Jamie spends the first hour of her day unbooking the Tuesday appointments the bot booked over the weekend.

Yeah. That’s the piece.

Nate B Jones keeps saying it and nobody’s listening

Nate B Jones — one of the more sober voices on AI agents right now — has been hammering the same line for a couple of weeks. Specification remains a 40-hour problem. Meaning: you can install an agent in 40 minutes, but getting it to actually do work that holds up takes 40 hours of defining what you want it to do, the constraints it has to respect, the edge cases you need it to catch.

And in the same run of content, he landed the companion insight. Your AI is 50x faster than a human at the task. Your business is getting 2x the results. The gap between 50 and 2 isn’t a technology gap. It’s a human gap. The tool is fast. Your specification is slow.

Here’s the thing — the specification bottleneck isn’t an AI problem. It’s the same problem that was always there. It’s called management. You’re just meeting it in a new costume.

What the agent can’t know is what you forgot to tell it

Back to Jamie. Why does the bot keep booking Tuesdays?

Because nobody told it not to. Nobody wrote down “we don’t run residential on Tuesdays — that’s restock day.” It lives in Jamie’s head. Has for twelve years. The bot doesn’t know because Jamie doesn’t know she knows it. She just does it.

The east-side ZIP problem is the same shape. There are a handful of ZIPs where the residential-looking addresses are actually outside the delivery radius, or are commercial strips that need a different crew. Jamie catches those in two seconds. She sees the address and reroutes. The bot can’t see that, because the rule has never been written anywhere. It’s shop folklore.

And that’s the pattern you’re going to see over and over in your AI agent that isn’t quite working. The bot confidently does the wrong thing because it has no access to the thing you never wrote down. “It doesn’t know that,” Jamie told me last month. “I know that. Nobody told it.”

That’s the whole diagnosis. You bought the fastest employee you’ve ever had — and then you handed it the thinnest onboarding packet any employee has ever gotten.

Specification was always the job

Here’s where it lands if you let it. The thing the AI agent needs from you — a clear statement of what work looks like, what the constraints are, what counts as done — is the thing your best manager has always had to produce. This is what we do. This is what we don’t do. Here’s why. Here’s what to do when X. Here’s what to do when Y.

Most small and mid-sized businesses have never had that written down. Not because owners are lazy. Because it wasn’t necessary. Jamie has been running dispatch since before the owner’s kids were born. The Tuesday rule lived in Jamie. The ZIP rule lived in Jamie. The “never route the new guy to Mrs. Alvarez’s house because she only lets Gabriel in” rule lived in Jamie.

That’s tribal knowledge. And tribal knowledge is fine — great, actually — right up until the moment you try to give that work to an AI agent. Then the gap becomes a cliff. Tribal knowledge can’t be read by a machine. It can’t even be read by a new hire, which is why the last person you onboarded was a six-month disaster until they’d absorbed enough of it by osmosis.

The agent has no osmosis. It has what you wrote down. And if you wrote down 40 minutes of “pick up the phone and book the customer,” you got 40 minutes of competence and the rest is folklore-shaped holes.

The 50 and the 2

Right? This is the 50x faster, 2x results thing. The tool is fast. You didn’t tell it where to go. So it goes fast — in the wrong direction.

I had another client try an AI agent on email responses to lease renewal questions. Insurance agency. The bot cranked out 80 replies a day where the office manager used to do 20. Owner was thrilled at the volume. Office manager was miserable because she was now rewriting 60 of the 80 — the bot was giving the wrong policy reference, wrong state cap, wrong renewal window. Same shape as Jamie. Bot fast. Specification thin. Human cleanup hidden in the shadows.

The owner’s first instinct was the AI is no good. His second instinct was we need a better model. His third instinct — the one that actually worked — was oh. We haven’t actually written down what a good reply looks like.

Forty hours of specification work later, the bot was writing 70 of 80 replies that shipped clean. Same model. Same tool. Same “AI.” Different specification.

So when you read the headline — 50x faster, 2x results — the gap isn’t the AI. The gap is the 38 hours of specification work you skipped.

Why you skipped it

Here’s the thing. You skipped the specification because specification is hard and nobody’s ever paid you extra to do it. It’s the work that makes every other piece of work go better, which means it’s invisible when it’s done and catastrophic when it’s missing. It’s the kind of work that gets deprioritized every single week.

And there’s a second reason that sits underneath the first one. Writing down what you do forces you to admit you don’t fully know what you do. You’ve been running on instinct and pattern-matching for years. Getting it out of your head and onto paper feels like you’re being quizzed on something you shouldn’t have to justify.

Every owner I’ve sat with has hit this wall. The owner of Jamie’s HVAC shop hit it last week. He asked me why the bot wasn’t working. I asked him if he’d documented the dispatch rules anywhere. He stared at me for a beat and then said — “I’ve been meaning to for years.”

Yeah. That’s the whole business of AI deployment right there in one sentence. I’ve been meaning to for years.

The good news hiding inside this

Here’s the upside, and it’s a big one. If you’ve run a business for any length of time, you already have the knowledge. Every rule the AI needs is in somebody’s head — yours, Jamie’s, the office manager’s, whoever. The raw material exists. You don’t need to go learn how to do something new. You need to extract what you already know.

That’s not a technology skill. That’s a management skill. It’s the same skill that lets a good manager write a training doc for a new hire. It’s the same skill that lets a good operator make a decision tree for a tricky situation. It’s the same skill that makes a company still function the week Jamie is on vacation.

If you’re somebody who’s been in the business long enough to know where all the bodies are buried, you are literally the person the AI needs. You’re the only person who can tell it the Tuesday rule. You’re the only person who can tell it the ZIP rule. You’re the only person who can say “never route the new guy to Mrs. Alvarez’s house.”

The ops manager who’s been told for three years that AI is going to replace her is, in fact, the exact person who has the skill AI needs most. Funny how the narrative has that exactly backwards.

The Monday Test — pick one process, write it for a new hire

This week, do one thing. Pick one process that an AI agent in your business is currently running — or one you’ve been thinking about giving to an agent. Not the whole business. One process.

Write it down step-by-step as if you’re handing it to a brand-new hire who’s never worked in your industry. Not as if you’re handing it to a machine. As if you’re handing it to a human who’s smart, motivated, and completely ignorant of your specific operation.

Time yourself.

If you finish in under an hour, you’re in rare company. Most people can’t.

If you can’t finish in an hour, congratulations — you just found the exact gap the AI agent has been running into. Every place the documentation got hard to write is a place the agent is currently making the wrong decision.

Does that make sense? The test isn’t whether the AI can do the work. The test is whether the work has ever been defined well enough for anyone — human or machine — to do it consistently. If the answer is no, the AI isn’t failing. The specification is.

Which, for the record, is a fixable problem. The Tuesday rule takes ten minutes to write down once Jamie sits to do it. The ZIP rule takes twenty. Every rule she’s been carrying for twelve years can be out of her head and into a document in about a week of concentrated effort.

A week that turns the bot from broken to functional, without touching the bot at all.

So.

The AI agent isn’t failing a technology test. It’s failing a management test. And the management test is one you’ve been putting off for years because there’s always something more urgent.

There isn’t, actually. Anymore. The specification work is the leverage. The person on your team who knows where the rules live is the person who can unlock the 50x speed for real, instead of watching it produce 2x results while somebody quietly cleans up the other 48.

You’re blaming the agent for a test you never wrote. Write the test. Watch the agent pass.

Source: Nate B Jones — “The Real Problem With AI Agents Nobody’s Talking About” and “Your AI Is 50x Faster. You’re Getting 2x. You’re Fixing the Wrong Thing.” April 2026. Credit where the thinking lives.

Companion piece: The Recursive Pattern (April 30) — the same diagnosis applied to how you delegate to humans, not just agents.