Serious AI with Chris Duff
September 22, 2026, updated October 12, 2026
These four questions are our test for whether a thing is worth building at all, given what AI can and cannot reproduce. Run them before starting a business, before adding a service line, and before pointing AI at any process inside a company that already exists. Same four questions, and the first use is the one that matters most.
They are below in full, including the version to run on your own operation. Serious AI Partners is Chris Duff’s AI implementation practice, working with operators who carry real financial consequence for a decision.
Read them as warning lights rather than as a score. What you are looking for is which ones carry weight and which ones are empty, because that is what determines whether there is anything here to defend.
The familiar half (whatever sits on the internet can be reproduced) is, I think, the less useful half.
The load-bearing half is the inverse. Fei-Fei Li built ImageNet and is a professor of computer science at Stanford. On the Huberman Lab podcast on August 10, 2026, agreeing with Huberman, she said that human cognitive behaviors that were never captured were never uploaded to the internet, and “today’s AI doesn’t have a way to do that.” The internet, she also said, is “the biggest collection of human behavior in multimodal forms,” and that collection is why AI is so powerful.
My read follows from hers: anything captured is now effectively reproducible, and anything never captured is out of reach for today’s AI. What decides the outcome is whether the data was ever recorded. Model quality is the wrong variable to watch.
1a. Could the output be linked on the public internet? Not whether it is, whether it could be. Published, indexed, scraped, sitting in a forum answer from 2017.
1b. If it could be, can access to it be controlled? Captured does not automatically mean exposed. An artifact can be written down, complete, and still sit behind an agreement, a licence, a login, or a client relationship that governs who reads it.
That second gate is the difference between a defence that works and one that does not. The defence is artifact control, rather than secrecy. Secrecy fails the moment one person writes the thing down. Control survives it, because control is about the terms of access rather than the existence of the document.
Applied to us: our methodology has been forced into codified form, so it can be written down. Which, the way I see it, makes it controllable, a very different thing from secret.
The fast curve is captured domains, anything with dense recorded outcomes.
The slow curve is uncaptured and liability-bearing domains. Hardware is one reason a curve runs slow, and Li’s timeline for robots shows how slow. Asked on the same episode where robots helping people would show up first, she called 12 months “a little bit too fast for robotics,” and when Huberman offered two or three years, she said to zoom out 30 years, because technology that involves hardware takes longer to manifest.
Do not read the slow curve as permanently slow. I’d watch for captured domains creeping into the work that services uncaptured ones. “Slow” describes the act, rather than the job built around the act.
Run it on your own business: the part of your process that is written down somewhere is already commoditized. The part that lives in one person’s head is the asset, and it is also your bottleneck.
The way I see it, AI can’t be the party that answers for a result: it can’t hold a licence, carry insurance, sign its name, or lose its own money.
AI can produce the analysis. It cannot be wrong.
Where a real person or entity has to own the outcome, there is a floor under that work that no model release moves.
Run it on your own business: find the moment in your process where a human has to own the result. That moment is staying. Everything built around it should be designed to protect it rather than to route past it.
Of the four, this is the one I think is easiest to fool yourself on.
A loop is powered by usage. It compounds whether or not you keep sprinting, and it survives you taking a month off.
A velocity lead is powered by the operator. You improve faster than a copier can follow. It evaporates the moment you stop, cannot be sold separately from you, and is a treadmill by construction.
Both are real defence and they fail differently. Having data is not a loop. The trap is describing a velocity lead in the language of a loop, then planning as though it were passive.
Run it on your own business, and this is the sharpest question in the set: is the AI thing you built getting better because people use it, or only when somebody tends it?
The follow-up matters more than the answer. What would convert it? Look for the improvement signal inside your own review process first. Converting it means making usage produce that signal directly. Small engineering, large effect.
Hands and hearts are two separate tests, and a business can pass one and fail the other.
AI cannot show up. The site visit, the install, the inspection, the thing that has to be physically moved or physically signed for. This is the sturdiest of the four, because it is bounded by hardware rather than by data.
Run it on your own business: separate the work that requires a body from the work that only requires a decision. My read: much of the second can be automated today and the first can’t, and treating them as one thing is how an automation project succeeds in the middle and dies at the last mile. Then find where the physical step sits in the sequence, because a physical step at the end constrains everything upstream of it.
Trust transfers person to person, and it is inheritable, which is the part that gets under-scored. A product inherits trust from whoever ships it: a track record, a public body of work, a reputation for showing real numbers. Hand a competitor identical code and I’d bet they still don’t get the sale, because the buyer is underwriting the operator as much as the artifact.
That comes with a ceiling worth naming. Inherited trust means the thing cannot yet be sold detached from the person, which limits delegation, acquisition, and any plan that assumes it sells itself.
Run it on your own business. Sort every customer interaction into two piles.
Transactional: where is my order, what does my policy cover, send me the certificate, fill in this form. Nobody’s trust is at stake. Automating these is pure upside, faster for the customer and cheaper for you. Start here when you look for recoverable hours.
Relational: the conversation where a customer decides whether to stay, escalate, or buy more. That interaction is a brand asset and it is carried by a person.
Automate the first hard. Instrument the second so a human arrives informed rather than replaced. The failure to avoid is skipping the sort, so the transactional work stays manual while an agent gets pointed at the one conversation that needed a person.
Answer each one yes, partly or no, on evidence rather than hope. Then look at the pattern.
Strong on none of them. The output is already reachable and nothing on two, three or four carries weight. That is a graveyard business. Arguing with this result for a year or two is the expensive way to arrive at it.
Strong on one. Workable, and narrow. It tells you what the offer has to be built around, and it comes with a ceiling attached. Strong only on handcuffs and the offer is the liability rather than the method. Strong only on hearts and it works as far as one person’s presence reaches. Name that ceiling before you plan around growth.
Strong on two or three. A sound result, and plenty to build on.
Strong on all four. Do not treat it as the bar.
One way is treating 1b like 1a. If the output could be linked but access to it can be controlled, that is a control problem. Control problems get solved with engineering and paperwork. Read as a verdict, it talks people out of something fixable.
The other way is failing yourself on 1a because your inputs are public.
Take land underwriting, which is where I learned this. As I see it, inputs like comps, parcel records and zoning are public, so the instinct is that a model can already do the job.
But what every deal turned on was never the inputs. It was a number on a specific parcel, with a position taken behind it, and whether that number turned out to be right. Nobody published that. The only person holding it was the one with money on the outcome.
Public inputs, unrecorded outputs. Check which one you are looking at before you score yourself a fail.
It does not tell you where to point AI next inside a business that already works. That’s a different diagnosis, and there the most expensive mistake is automating work that was never the bottleneck: a faster version of something that wasn’t holding the business back, plus a larger bill.
Say the four questions come back strong and the work is worth building. The next question comes from whoever signs the budget, and we would expect it to be about money: what pays for it, and how anyone will know it worked. Three answers to have ready, sized for a company of 50 to 500 people.
One number per project, today’s value written down, and one named person who owns it. Days to close the books. Hours from quote request to quote sent. First-response time on support tickets. Rework rate on invoices.
If nobody can name the number, the project is not ready for the budget, however good the demo was. My guess is it’s the step most projects skip. Skip it, and the AI bill can climb while nobody can point to a number that moved.
Some of what a first AI project needs may already be going out the door. Four places to look:
Add them up and set the total beside the ask. The closer those two numbers are, the shorter the budget conversation.
Ask for a first step rather than the full year. Each step unlocks when the number from step one moves by an amount agreed before work starts. One process for the first quarter. The second quarter unlocks when that process hits its target. The third unlocks when the result holds without someone tending it, which is question three, the loop, showing up in the budget.
A ramp is an easier yes than a lump sum. It also means nobody keeps paying for something that stopped working in March.
In a medical practice we would look in four different places first: billing and denial rework, prior authorization hours, front-office turnover, and tools bought to patch the EHR. Where Does a Medical Practice Find the Money for AI in 2027? walks through each one, citing sources from 2025 and 2026.
The value was never the framework. It is running it against a specific company’s specific constraint, and then staying accountable for what happens next. Anyone can produce the analysis now. Somebody still has to own the result.
Serious means underwritten, and we stake the result.
Every Monday, how enterprise AI actually gets built, what it cost to get right, and the parts that did not work. Every issue carries a receipt: a real number, a real decision, or a real thing that failed. Join 850+ operators, or read every past issue free first.
If something in your business is breaking right now, tell us what it is.
Fill out a short form, and we read every answer ourselves. If there is a fit we will reach out and find a time that works for you, and if there is not, we will say so and point you toward someone better suited. Either way, the moat question is worth answering now, in a market AI has already reset.