I review a lot of AI roadmaps. Companies bring us in to help prioritize and execute their AI strategy, and the first thing they show us is a list of twenty or thirty potential projects. The list is always too long. The ambition always exceeds the capacity. And nobody has done the hard work of deciding what not to do.
The kill list is the most valuable artifact in any AI strategy. It is the list of projects that you have evaluated, understood the potential of, and consciously decided to not pursue. Not because they are bad ideas, but because they are not the best use of limited resources. Every hour spent on a mediocre AI project is an hour not spent on the one that could transform the business.
We help clients build their kill list using three criteria. First: is the data available and clean enough to support this? If the answer is no, and the data cleanup would take more than three months, it goes on the kill list. Not forever — just until the data situation improves. Many AI projects fail because the team assumed the data would be there and it was not. Second: does this solve a problem that someone is actively complaining about? If the project was suggested by someone who does not do the work, it is suspect. Third: can we measure success in dollars or hours saved? If the answer is vague, the project is not ready.
I will give you a real example. A client wanted to build an AI system to predict which of their customers were likely to churn. Classic AI use case. Sounds great in a strategy deck. When we dug in, we found that they had no historical churn data worth using, their customer interaction data was spread across four systems that did not talk to each other, and even if the model worked perfectly, the team that would act on the predictions was already stretched thin. We killed it. Instead, we automated their invoice processing, which saved them 200 hours per month. Less exciting. More impactful.
The organizational politics around the kill list are tricky. Every project on the list has a sponsor. Telling a VP that their pet AI project is not worth pursuing requires diplomacy and evidence. We have found that framing it as sequencing rather than killing helps. Your project is not dead; it is phase three. And phase one is going to generate the ROI that funds phase three. This is usually true and always easier to hear.
One common category on the kill list: AI projects that are really data engineering projects in disguise. The AI model is the easy part. The real work is building a data pipeline that nobody wants to fund because it is not as exciting as an AI demo. If 80% of the project is data infrastructure, call it a data infrastructure project and evaluate it on those terms. Do not use AI as a Trojan horse to get data engineering funded — it leads to misaligned expectations and disappointed stakeholders.
Another category: projects where the AI is solving a problem that a simple rules engine could handle. Not everything needs machine learning. If you can write the logic as a flowchart, you probably do not need a neural network. We have talked clients out of AI projects and into rule-based automation more times than I can count. It is less impressive on a slide. It ships faster, costs less, and is easier to maintain.
The kill list should be reviewed quarterly. Projects that were killed six months ago might be viable now because the data improved, the team grew, or the business need intensified. The kill list is not a permanent graveyard. It is a waiting room. But the waiting room needs a bouncer, because the default organizational tendency is to let everything back in.
My experience is that companies with a ruthless kill list ship two to three times more value from their AI investments than companies that try to do everything. Focus is a force multiplier. The most successful AI teams I have worked with are the ones that say no more often than they say yes.