How To Choose Your First AI Project, And What To Refuse
Most AI projects fail before a line of code is written. They fail in the decision about which problem to solve. This guide is the framework I use, including the four dimensions I score every candidate against and the three types of project I refuse outright.
theaiceo.orgYears ago I took a project to build an email application. I was excited, because it was the first big check. I told the client three weeks. It took seven months.
We had not researched it properly. We underestimated our own capacity and we underestimated the integration, and I did not have the money to do the research that would have told me the truth before I quoted. I did not fail at engineering. I failed at scoping, and scoping is a leadership decision.
Later, running a product of my own, I modeled AI credit costs against what a light user consumed. Real customers are not light users. We lost more than fifty dollars a month on every client we signed. We were growing and bleeding at the same time, and growth was the problem.
No vendor lied to me in either case. I got both wrong myself. Everything in this guide exists because of one of those two failures, or because I have since watched a client walk toward the same cliff.
Trent T. Daniel · CEO, BotMakers, Inc.
A division of publicly traded BioQuest, Inc. (BQST) · hello@theaiceo.org
The highest-value problem in your organization is almost always the wrong place to start.
Every leadership team I work with wants to begin with the biggest number. The eight-million-dollar inefficiency. The thing that would transform the business. It is the right instinct and the wrong first move.
That project is too visible, too complex and too slow. It touches six departments, three systems and a compliance question nobody wants to open. By the time it half works, the sponsor has been promoted, the budget cycle has turned over, and the organization has learned that AI does not work here.
Your first AI project has one job: earn the right to do the second one. That means it has to finish, show a number, and do it before anyone loses interest.
If you cannot imagine reporting a result on this project within one quarter, it is not your first project. It may still be your third.
Score every candidate out of twenty. Below fourteen is not your first project.
This is the framework. It is deliberately crude, because a crude framework everyone applies consistently beats a sophisticated one that becomes an argument about weighting.
How often does this actually happen?
Volume is what turns a small per-instance saving into a number your CFO can see. Low-volume work is rarely worth automating first, however annoying it is.
Can you state the success number before you start?
If you cannot measure the before, you will never prove the after. Teams skip this because measuring is boring, and then spend a year arguing about whether the project worked.
What happens when it is wrong?
AI is probabilistic. It will be wrong sometimes. Choose a first project where being wrong costs an apology rather than a filing.
Who wants this to exist?
This is the dimension teams score most dishonestly, and the one that predicts failure best. If you had to assign it, you have already lost.
Print this page. Score every AI idea currently circulating in your organization. Anything below fourteen is not your first project.
Saying no to three things is a real decision. Saying yes to everything is not.
I build custom AI for a living and I turn down work regularly. Not out of principle, but because some problems are genuinely worse with AI in them. These are the three I decline every time.
If one wrong answer is a regulatory event, a safety incident or an irreversible financial action, a human stays in the loop. AI can draft, sort, retrieve and suggest inside that workflow. It does not get the final decision. Anyone who tells you otherwise is selling.
If your own experts disagree about the correct answer, there is nothing for the system to learn. This is a process problem wearing a technology costume. Fix the disagreement first, and you may find you no longer need the AI.
A project with no named owner who personally benefits will end the day the initial enthusiasm does. Not because the technology failed, but because nobody was ever going to fight for it in week seven.
The license price is the part you can see. It is rarely the largest part.
Software pricing trained everyone to think in seats. AI pricing works on consumption, which means the number scales with how successful you are. That is a genuinely different shape and it catches experienced buyers.
Here is what I lost to it. We priced a product against what a light user consumed. Real customers used it properly and we lost more than fifty dollars a month on every client. We fixed it by re-architecting rather than repricing: cheaper models for the easy work, the expensive model only where it changed the answer, aggressive caching, and cutting everything we were sending that never needed to go.
It took real engineering skill and some judgment to cut consumption without the product getting worse. That is the work nobody prices in, including the people selling to you.
Consumption at real volume
Model the usage you expect in year two, not month one. Ask the vendor for the pricing curve, not the entry point.
Integration
Frequently a separate statement of work. Establish who does it and whether it is in this price before you are excited.
Data preparation
Somebody has to clean it. Establish who, and at whose cost.
Human review
The review step you will still need is a permanent operating cost, not a temporary one.
Retraining and drift
Performance degrades as your business changes. Ask what maintaining it costs annually.
The exit
What does leaving cost, and do you keep your own data? Almost nobody asks this before signing.
Size the first project so that cancelling it is not embarrassing. That is the real constraint, and it is the reason most first projects should cost less than one senior hire.
Three phases, one owner, one number, and a review date that exists before you start.
A plan without an owner and a review date is theatre. Everything below assumes you have both. If you do not, stop and get them; the rest of this is wasted effort without them.
Decide today what result by the review date would make you stop. Write it down while everyone is calm. Later, when the project is struggling and reputations are attached, the decision becomes arithmetic that was agreed in advance rather than a judgment on anybody.
This guide is the short version of a working session I run with executive teams. In half a day a leadership group surfaces every AI opportunity in the organization, scores them against this framework, refuses three, assigns owners to the top three, and drafts the ninety-day plan before anyone leaves the room.