Quick answer
Write every AI idea as a job a person already does. Score it on value, effort, and risk. Build the top two far enough to fail on your real data. Drop anything you cannot prototype inside the sprint. That cut is the job of AI strategy, not a deck of maybes.
Key takeaways
- Value is money or hours you can name. “Feels modern” is not a score.
- Effort is weeks to a working slice, not a vendor slide.
- Risk is what happens when the model is slightly wrong. Recoverable stays. Customer money and HR do not.
- We publish $2,000 a week. The scoring sprint is two to four weeks, so $4,000 to $8,000 before a build.
The three numbers
- Value
- Hours back, or revenue you can point at. If you cannot name the number, the idea is a vibe.
- Effort
- Calendar weeks to a slice on real data. If it needs a six-month data clean first, that is the project, not the model.
- Risk
- Cost of a confident wrong answer. A digest you can ignore is fine. A refund you cannot unwind is not.
High value, low effort, recoverable risk: do that first. High value, high risk: keep a human on the path. Low value: drop it even if it demos well.
“If we cannot put a working slice in front of you in the sprint, it is not a bet yet. It is a research topic. We will say so.”
Chris NutbeenNuttifox
What we drop on the first pass
- Anything that needs a clean warehouse before week one
- A chatbot that “is the product” with no owner for the wrong answer
- Fully automated outreach. It burns the list.
- Strategy, firing, and negotiation. Those stay with people.
The leftover list is usually two or three jobs. That is enough. Then pick the seat — chat, workflow node, or product API — in where to put the model. Do not buy four licences to look busy.
What the sprint is for
A score without a prototype is theatre. We take the top two bets onto your data and stop when the slice works or when it does not. Claude Pro is $20 a month ($17 annual) on the Sep 2026 list. That seat writes the first draft. It does not score the bet. The engagement is the people in the room.
You want a 40-page strategy PDF and no build. Or you already know the bet and only need hands. Then skip the score and go straight to AI automation or a product slice.
FAQ
Do we need clean data first?
No. Part of the score is what the data allows. If the slice dies on messy rows, that is the finding.How many ideas should we bring?
All of them. We will cut. A list of twenty is normal. A list of two that both fail the prototype test is also a result.What if nothing scores well?
Then we stop. Paying for a build you do not believe in is worse than a short no.Is this the same as an AI agent?
No. An agent is one shape a winning bet might take. Score the job first. See AI agents for small business.Have something to build? Let’s ship it.
Book a call