Jonas BullNavigating the frontier between AI and human
The instrument
Four questions, in order.
Pass or fail, cheapest kill first, no partial credit. I run this on other people's AI projects for a living, and roughly four in ten survive it. Most of the rest die on question two. Killing a project in week two is cheap. Killing it in month six is a loan you are still paying on.
Before you start
Two people read this page differently.
If you own the business, this is a focus instrument: what to build next, and what Monday looks like. If you lend to one or invest in one, it is a diligence instrument: is the AI story value or narrative. Same engine, different output.
Bring one task you already suspect is bleeding money. Nothing you type here leaves your browser unless you ask me to send it.
Stage one
The gates.
A strength on one gate does not offset a fail on another. If you fail one, stop, fix that, come back. Then, and only then, argue about what the thing is worth.
Gate one
Granularity
Unanswered
The unit of AI value is the task. Not the job, and not "operations." Name the task, name who does it, and put an hourly cost on that person. Then name what that person does with the hour you give back. Miss any of it and there is no project yet.
One repeated task. Same shape every time, different contents.
How often it actually happens, not how often it should.
Door to door, including the part where they go find the answer.
What that person costs you per hour, not what they take home.
Nobody gets cut in a small shop. An hour with no destination turns into slack, and slack never shows up on the books as savings.
Gate two
Verification economics
Locked
Every AI system has an error rate, and it is never zero no matter what the demo says. So the question is what a wrong output costs you, and whether checking it is cheaper than producing it. Output includes actions: an agent that emails the wrong customer, or files a document where it should not go, made an error with a price like any other. Leaks and compliance breaches sit at the top of that list. Price them here, not in a security review that shows up after the decision. Most cases die right here.
Write your failure price down before you talk to anybody who sells software. Count what it could send or expose, not just what it gets wrong.
How long it takes a competent person to tell whether the answer is right.
Gate three
Durability
Locked
Build where your data or your workflow is the thing that makes you different. Buy where it is a commodity. Most small businesses have no moat here, and that is fine, but then you are buying operating leverage and you should price it like opex. Agentic systems do not cost what the demo cost. They cost what the usage costs, and usage climbs. If your advisor never says "just buy that," get a different advisor.
Is the data or the workflow behind this task something a competitor could not easily copy?
Does a product already on the market do most of this?
List price plus whatever the one integration costs you monthly.
Is the honest answer that the technology is not there yet?
Model calls, hosting, licenses, at real usage. Not the demo, and not the first month.
Parked is not dead. Parked does not get money either.
If the trigger has not fired by then, it dies on its own.
Gate four
Proof of correct
Locked
Before you build, say in writing what a correct output is, and have a fixed set of inputs where you already know the right answer. Without that you cannot tell whether a change helped, you cannot swap to a cheaper model when one lands next quarter, and you cannot defend the spend to the person who signed the check. That set is also the only gate artifact still working after launch. No eval set, no build. No rerun, no renewal.
Can you write down, in a paragraph, what a correct output looks like?
Do you have, or could you assemble inside two days, real inputs with known correct answers?
Two days of your best person's time. You keep it no matter which vendor you end up with.
Will you rerun it on a schedule, and again whenever the model, the vendor, or the process changes?
No score resurrects a kill. No valuation, however pretty, waives a gate.
Stage two opens when all four gates pass
Stage two
Now discount it.
It survived, so it gets valued, and like any good valuation it gets discounted. These four are where otherwise good projects quietly bleed out. A bad decay score is an instruction to redesign, not to abandon. Each decay gets a countermeasure and a name beside it, or it is not handled. Redesign, re-score, and re-run the gates only if the redesign changed a gate answer.
Readiness
Leave it blank if nobody has quoted yet.
The page
What you walk out with.
The output is never a yes or no on "should we do AI." It is a short ranked list of the things that survived, with owners and dollar figures on them, and a longer list of the things you are done pretending about.
Next
Bring one task and the second meeting is short.
One task with an hourly cost on it, the cost of getting it wrong, and the honest answer about where your data lives. That is the whole intake. Book time directly, or send me the run and I will read it before we talk.
Sent. I read these myself.
There is no model behind this page. It is arithmetic you can check. Seemed like the best way to build a tool whose whole job is killing AI projects.