Specialist hiring is slow and expensive, and a full roster of outside hires still would not carry the knowledge of how your business runs. Most of the capability has to come from people already inside it.
Most programs try to solve the people problem by hiring. Specialist hiring is slow and expensive, and even a complete roster leaves out what decides whether a system gets used, which is knowing how the business runs. Some hiring is necessary. It is rarely the largest source of the people a program needs.
People come from three places and they suit different roles. Hire for the work that has to be strong from the first release and is slow to build internally: platform engineering, ML engineering, MLOps. Upskill people already in the business for the product owner, analyst, and domain roles, where knowing the process is the harder half and the tooling can be taught. Use a partner for demand you cannot staff permanently, or for a specialism you need twice a year. Most programs put nearly all their attention on the first, which is what leaves them exposed when those hires leave.
It is usually faster to teach an experienced insider the data and modeling skills than to teach an outside specialist your business. The insider already has the part that takes years to acquire, and the technical part has a curriculum.
They are not hard to find, though they are rarely in a formal analytics role: the person in finance who maintains the reporting the rest of the function depends on, the process engineer who has automated part of their own job in Python, the planner keeping a separate model in a spreadsheet because the system does not answer the question they need answered. Identify them in the first few weeks. They become the first training cohort, and they explain the program to their own function more credibly than the program can explain itself.
The people worth training are usually already doing analytical work outside any analytical role.
Individual upskilling does not scale on its own.
Run it in cohorts rather than one person at a time, so people have peers on the same material and the curriculum can attach to live use cases instead of exercises. Give capable non-specialists a supported route to do their own analysis: approved data sets, a defined tool set, and a review step before anything gets published or used in a decision. Cover the executives who fund and challenge the work, far enough that they can read an evaluation result, ask what the system does when it is wrong, and tell a hard constraint from a preference.
The embedded domain expert on each delivery team comes from the same place. Training your own is the alternative to borrowing one for a single use case and losing what they learned when the engagement ends.
A program staffed entirely by outside hires and contractors stops when they leave, and the reasoning behind how the systems were built leaves with them. A program that has trained its own people keeps running, because the capability sits with people whose careers are already there.
The measure worth tracking is how much of the next use case your own people can run without outside help, from the value case through to supporting it after release.
Tell us the process or system you are trying to improve, and we will tell you what it would take to change it.