How do you choose which AI use cases to start with?

AI use case selection is the process by which an organization decides which project to tackle first, weighing three axes together: the value it creates for the business, the readiness of the data it needs and the organizational feasibility, meaning whether an internal owner exists, whether the process can change and whether anyone will use the output produced. It is not a brainstorming exercise or a list of ideas ranked by enthusiasm: it is an explicit filter meant mainly to reject candidates, not just to pick one. Most enterprise AI projects fail not because of a model limitation but because of a selection made badly at the start, a case chosen because it sounded good on a slide rather than because it answered all three criteria together. A use case that wins on a single axis and loses on the other two should be dropped, not patched along the way.

See also: AI maturity assessment · The assessment of how ready a company is for AI across dimensions: data, skills, processes, governance, measurement.

The three axes to cross

The first axis is business value: how much is it worth, in time, money or reduced risk, to solve that problem. The second is data readiness: does the data already exist, is it accessible and clean enough to feed a model, or does it need to be collected first, which changes the project's timeline and cost. The third, often the most overlooked, is organizational feasibility: is there an owner accountable for the result, can the process actually change, and will people use the output instead of ignoring it. A candidate strong on a single axis stays weak overall: the matrix exists to surface those imbalances before a budget gets committed.

The criteria for rejecting, before those for choosing

The most useful part of this method is knowing when to say no. Drop a use case when the data does not exist and nobody intends to collect it. Drop it when the estimated savings are lower than the ongoing maintenance cost, because a model needs monitoring, retraining and correction, not a one-time build. Drop it when it requires an accuracy today's technology cannot reliably guarantee. Drop it when nobody will change how they work as a result of the output, because an ignored result creates no value even when the model is technically correct. And drop it when its value depends on an integration outside your control, a third-party system or external data whose quality nobody guarantees.

An enterprise example

An industrial group with multiple plants compared ten AI candidates using this matrix. It dropped five: two because maintenance data existed only on paper and nobody was willing to digitize it, one because the projected savings were lower than the cost of running the model in its first year, one because it depended on an integration with an external vendor's system outside its control, and one because the proposing department had no intention of changing its decision process, it only wanted a dashboard to show in meetings. Of the five that remained, it started with the smallest, not the most ambitious.

Why it matters for decision makers

Without this filter, enterprise AI becomes a collection of pilots chosen for enthusiasm or political urgency, which quietly die in demo because nobody checked the minimum conditions for success beforehand. An organization that runs every candidate through the three axes ends up with a smaller but real portfolio. It is the same work I do with the AI Maturity Assessment: first measure how ready the organization is, then choose what to do with that readiness.

The practical advice on the first case

The first use case to tackle is not the most ambitious one on the list, it is the one that teaches the organization the most at the lowest cost. A small but well-chosen project builds internal trust and competence in the process, while an ambitious project that fails on the first attempt makes every later proposal harder to sell internally. The same logic applies to the decision intelligence angle: a first case that makes visible how a decision changes thanks to data is worth more, in the first twelve months, than a case promising huge impact but requiring a transformation the company is not yet ready to sustain. It pays to keep the Pareto principle in view while selecting: a few variants almost always cover most real cases, and the project that digs in on the long tail is the one that never reaches production.