You are running your business on numbers that stopped being true six months ago.
The accounting partner whose capacity model says a senior can do thirty corporate returns between March 1 and April 15 — that number is five years old. Rock solid back when it was true. She staffs off it. Bonuses off it. Hires off it. She has never asked “what if the right senior with the right tools can do seventy-five?” — because she has never watched one do it. The number in her spreadsheet is a ghost. It’s the number from the world before.
Same shape, different shop. The home-services owner budgeting his next CSR hire off “it takes a good one six months to get productive.” That number was true in 2019. In 2026, with the right AI onboarding on day one, the same hire is holding their own in six weeks. But he doesn’t know that, because his last hire was pre-shift. Six months is the ghost in his head.
Your old baselines are gone. What you replace them with is the next competitive advantage.
Why this matters more than any AI tool review you’re going to read
Every AI pitch you’ve heard in the last year has been about cost and speed. “Your team will be ten times faster.” “This tool saves four hours a week.” “This one drafts proposals in under a minute.” All of that might even be true. None of it matters if the number anchoring your plan is a ghost.
Here’s the thing. One of the most experienced engineers working today recently admitted — flatly — that his own calibrated sense of how long things take, twenty-five years of pattern recognition, is completely gone. Not degraded. Not needs-recalibrating. Gone. If a 25-year expert can’t estimate a timeline anymore, what does that say about the three-year strategic plan your board approved in January? The five-year revenue model your banker is reviewing? The capacity spreadsheet your controller builds your bonus pool from?
Every one of those numbers was calibrated in the world before. How long a proposal takes to write. How many hours a month your bookkeeper needs. How many tickets a support rep can close in a week. How long a senior CPA takes to turn around a corporate return. Every one of them is in motion. And most operators are still running off the pre-shift numbers, making pre-shift decisions, and wondering why the plan keeps missing.
The ingredient that replaces the ghost
Our AI Onboarding Framework — the seven-ingredient Professional Recipe — has a specific ingredient for exactly this. It’s Ingredient #6: Measurement.
Data over feelings. Three metrics that map to value. That’s the discipline. The Chef doesn’t get to run the kitchen on the argument “I think we’re faster now.” The Chef measures. Specifically — three numbers that each tell you something concrete, stay the same long enough to compare against themselves week over week, and map to something a customer actually cares about.
Most operators I work with have spent the last year telling themselves they’re “more productive with AI” without ever writing down a single number. “My team saves like four hours a week on emails now.” Maybe they do. Maybe they spend four more hours checking AI-generated emails that went to the wrong clients. Nobody measured. So nobody knows. And in a year when the first six months of productivity gains could be real or imaginary depending on who you ask, the shops that measured are about to run away from the shops that didn’t.
Ingredient #6 is the part of the Professional Recipe most shops skip because it feels like overhead. It isn’t overhead. It’s the difference between “I think we got faster” and “we went from thirty returns per senior to fifty-four, and our error rate held” — which are two completely different sentences with two completely different consequences when you’re sitting across from your banker, your board, or the partner who’s about to buy in.
The ceiling isn’t the machine. It’s you.
Same engineer, different sentence worth pulling out: he can fire up four AI agents in parallel and have them work on four different problems at once. And by 11:00 a.m. — he is wiped out.
Read that again. One of the most productive engineers on the planet, running a machine-powered four-stream workflow, cashed out by lunch.
That’s the punchline most AI pitches won’t say out loud. The machine is no longer the bottleneck. The bottleneck is the human keeping up with the machine. If you put your most senior person on three AI-assisted drafts in parallel, she’s doing the cognitive work of three of her, even though the typing is free. That’s not 3x the work. It’s 3x the decisions. And decisions exhaust people in a way typing never did.
The premium human skill is no longer doing the work faster. It’s staying clear and decisive while a machine hoses output at you. That’s a different muscle. Nobody trained your best people for it. And the shops that figure out how to pace that muscle — when to go to four streams, when to pull back to two, when to stop — get more compounding value out of the same headcount than anyone else in their market.
Which is, again, a measurement question. How many parallel streams before quality drops? How many before your senior person’s error rate starts creeping? How many before the exhaustion compounds into the following day’s output? If you’re not measuring, you’re guessing. And the shops guessing are about to start losing to the shops measuring.
Free to build, not free to run
One more quote from the same conversation, worth stopping a planning meeting for: there’s very cool software the engineer built that he’s never used — because it was quicker to build it than to actually try and use it.
Think about that for a second. Every operator I talk to now has a backlog of “AI we could build.” An HVAC owner tells me his team is about to build a proposal generator, a call-review tool, a dispatch optimizer, and a marketing assistant this quarter. He’s excited. He should be. But building them is now the easy part. Using them — actually changing how your team works, what they stop doing, how the output flows into the next step — that part is exactly as hard as it ever was. Build four things that no workflow absorbs and you’ve got four cool tools gathering digital dust and four sunk investments.
Ingredient #6 catches this too. Before you build a thing, pre-commit to the use. Who runs it. When. What gets retired so there’s time in the week to run it. What good output looks like. What wrong output looks like. And — critically — what number tells you whether it’s working three weeks in. If you can’t answer those on one page, you’re building another tool you’ll never use.
The measurement discipline turns a build list into a proof list. That’s the upgrade.
The week in one sentence
Three pieces from me this week. Same worldview in different costumes.
Tuesday — the hard part of AI isn’t the tool, it’s specification. Writing down what you actually want the machine to do. The gap between 50x speed and 2x business results is specification work you skipped.
Thursday — the discipline isn’t doing more with AI. It’s doing less. Output is cheap now; the skill worth anything is knowing what not to ship. That’s Ingredient #5.
Today — your old baselines are ghosts. Every number in your plan is in motion. The operators who replace feelings with fresh measurement — Ingredient #6 — are the ones who find out which of the last six months of “productivity gains” are real. Everyone else is guessing in public.
One worldview, three costumes. Specify what done looks like. Don’t ship what the machine happily produces. Measure the gap between feeling and fact. The AI is finally doing its job. Now we find out whether you can do yours.
The Monday Move — run your own measurement experiment
This week, pick one task you’d normally block a full day for. Proposal. Engagement letter. Quarterly marketing plan. Quote package. Whatever your shop’s full-day task is.
Give it to AI with a clear spec — what good looks like, what’s non-negotiable, three examples from your own archive of work that shipped clean.
Time the result. Then write down three numbers.
- How long you thought it would take.
- How long it actually took.
- How good the output was, on a 1-to-10, compared to what you’d have produced yourself.
The gap between line one and line two is your personal inflection point. The number on line three tells you whether you still have specification work to do — Tuesday’s piece — or whether you’re ready to scale.
Run that on one task this week. A different one next week. Inside a month, you’ve got real numbers from your own shop replacing the ghosts in your planning model. That’s the upgrade. Not a new tool. A new measurement.
So.
Your old baselines are gone. What you replace them with is the next competitive advantage.
Your capacity model is a ghost. Your hiring spreadsheet is a ghost. Your three-year plan is anchored to ghosts. And every day you wait to measure fresh, your competitors who are measuring gain one more day of compounding advantage over you.
Go measure.
Framework spine: The Professional Recipe, Ingredient #6 — Measurement. Read the full framework
Companion pieces: The Specification Bottleneck (April 21) — the management skill every AI deployment actually needs. The Laziness Problem (April 23) — Ingredient #5 and why most shops ship the wrong kind of output.
Source influence: a long-form conversation with Simon Willison — 25 years in software, now writing 95% of his code without typing it — on Lenny Rachitsky’s podcast, April 2026. His numbers anchor the “gone baselines” claim and the “wiped out by 11am” observation. Distilled and operator-translated.
~ source material · The Professional Recipe, Ingredient #6: Measurement
