How to choose your first AI project

The first project decides whether there is a second

Most companies have more AI ideas than they can fund. Every department has a list, vendors have demonstrations, and the board wants to see something working. Which idea goes first is the most consequential choice in the program, because the first project sets the budget, the credibility and the appetite for everything after it. A first project that stalls makes the second one harder to approve.

Stalling is common. In July 2024 Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value. Value is concentrated, too. In McKinsey's 2026 State of AI survey, 37 percent of respondents attribute at least some EBIT impact to AI use, while the high performers, who attribute at least 5 percent of EBIT to AI and describe its impact as significant, are about 6 percent of all respondents. Most of the reasons a project fails are visible before it starts. Choosing a first project well means looking for them.

Score every idea on what it is worth

Worth is what the business gains if the system works: money saved, decisions made faster, fewer errors, customers better served. A worth estimate is only as good as the number it starts from. How many of these cases arrive in a month, how long each one takes, how often one goes wrong and what a wrong one costs. If nobody can say how the work performs today, nobody will be able to show that the AI improved it, and the project will end in an argument rather than a decision.

Worth also needs an owner: a named person who runs the process, wants the improvement and will sign the pass mark the system has to reach. An idea that belongs to everyone, or only to the technology team, has no one to accept it when it is done.

Score it on how easy it is to deliver

Ease has four parts. The first is data: the documents, records and history the system needs exist, are reasonably clean and are yours to use. In a February 2025 Q&A, Gartner predicted that through 2026, organizations will abandon 60 percent of AI projects unsupported by AI-ready data. The second is integration: the system can read and write what it needs through interfaces your IT team already supports, without a migration first.

The third is risk: the system proposes and a person decides, the use is not one a regulator treats as high-risk, and a wrong answer is caught before it reaches a customer. The fourth is adoption: the system fits into the tool and the moment where the work already happens, instead of asking people to open something new. Adoption decides more of the value than most plans allow for. McKinsey found that nearly three-quarters of its high performers report fundamentally redesigning workflows, a reminder that the work around the model matters as much as the model.

Four kinds of idea, and a list that gets shorter

Put every idea on the two scores and four groups appear. Ideas that are worth more and easier to deliver are where to start. Ideas that are valuable but hard belong on the plan for later, with the work that would make them easier written down: the data that needs cleaning, the system that needs an interface, the review that has to happen first. A well-chosen first project often builds some of that groundwork.

Ideas that are easy but worth little are quick wins: take them when they are cheap, and buy rather than build when a product already does the job, as we argued in build versus buy for enterprise AI. Ideas that are worth little and hard to deliver should stop, however popular they are.

The scoring only helps if it removes ideas. If almost everything survives, the scores were too generous, and the program ends up funding a little of everything. Gartner's June 2026 advice to CFOs building finance AI roadmaps puts it in one line: "If everything is a priority, nothing gets funded." It recommends prioritizing use cases on net business value, feasibility and scalability, and piloting three to five at a time.

The usual mistakes

Starting with the most visible idea. A customer-facing assistant is easy to demonstrate and hard to deliver well, because every answer reaches a customer and every mistake is public. An internal process where a person reviews the output is usually the better first step, and it builds the controls the customer-facing version will need later.

Starting where the system acts on its own. A first system should propose and let a person decide, so that it earns trust from a record of good recommendations before it is given authority. Starting without a baseline, which turns the result into an opinion. And reaching for AI where a rule would do: if the decision can be written down as a clear rule, ordinary software is cheaper, faster and easier to audit.

What a good first project looks like

A good first project is narrow: one process, one owner, one measurable outcome. The system reads and proposes, and a person decides. The pass mark is written into an evaluation suite before the build, the controls a reviewer will ask to see are designed in from the start, and the system runs alongside the existing process until the results agree. When it works, it leaves more behind than one improvement: a data path, a habit of evaluation and a relationship with your reviewers that the second project can reuse.

Choosing that project, before a line of code is written, is our strategy work: a scored list of ideas, including the ones we recommend stopping, an architecture under the survivors and a costed plan. Most work with us starts small: one fixed-scope piece of work that proves the system before anything larger is committed. If you have a list of AI ideas and no agreed first step, book a conversation with a founding partner.