The instinct when building an "AI team" is to hire data scientists and machine learning engineers first. This isn't wrong, but it's incomplete in a way that causes a specific, recurring failure pattern: technically excellent models that never get deployed into anything that actually changes how the organization works, because the roles that would have made deployment and adoption succeed were never staffed.
The roles that get skipped, and what goes wrong without them
Someone who deeply understands the business process being changed, not just the data involved in it. Technical staff frequently understand data structures without understanding why the underlying process works the way it currently does (including the informal exceptions and edge cases that never got documented anywhere). Without this role, models get built against a clean theoretical version of a process that doesn't match its messy operational reality, and the gap surfaces only after deployment, when it's expensive to fix.
Someone accountable for change management, not just technical delivery. A technically successful model that changes how people are expected to work needs the same change-management attention any other significant process change requires: communication, training, addressing the specific concerns of people whose jobs are affected. AI initiatives without this role tend to produce technically working systems that see low real adoption, because the human side of the transition was never actively managed.
Someone with genuine authority to make deployment decisions, not just technical review authority. A model that's technically ready but lacks a clear organizational path to actual deployment (nobody with the authority to say "yes, put this into production and change the associated workflow") stalls indefinitely in a "successful pilot" limbo that never converts into real operational change.
Someone specifically responsible for ongoing monitoring after deployment, not just development. AI systems can degrade in accuracy over time as real-world conditions shift away from the conditions the model was trained on, a problem invisible without deliberate monitoring, and one that pure development teams, once the "interesting" building work is done, are often not staffed or incentivized to keep watching for.
The actual team shape
A genuinely effective AI initiative usually needs, in addition to technical AI/ML expertise: someone with deep operational knowledge of the specific process being changed, someone accountable for the change-management side of adoption, a clear decision-maker with real authority over deployment, and an ongoing monitoring function that doesn't end when the initial project does. Organizations that staff only the technical role and expect it to also cover process understanding, change management, deployment authority, and monitoring are asking one function to do the job of several distinct ones, which is a large part of why so many technically sound AI projects never produce real organizational impact.