The order is not optional
Warehouse, then governed master data, then automated planning, then agents. I have never seen that order successfully rearranged, and I have seen it attempted often, usually because the agent is the part that gets funded.
The reason is not purity. It is that each layer removes a category of question the agent would otherwise have to answer at runtime, and an agent answering questions about its own inputs is an agent producing decisions nobody can defend.
The three layers, and what each one is actually for
A warehouse, so there is one place to be wrong
The value of a central warehouse is not speed. It is that when a number is wrong, there is exactly one place to go and fix it, and the fix propagates to everything. Before that exists, every correction is a negotiation between spreadsheets, and nobody can tell you which version the decision was based on.
A useful sign that this layer is missing: two people bring different numbers for the same thing to the same meeting, and the meeting is about which number is right rather than what to do.
Governed master data, so the agent and the operator mean the same thing
This is the layer that gets skipped most often and causes the most damage. Master data is the shared definition of what a store is, what a product is, what a category is, and who owns each of those definitions.
Governance here means something narrow and concrete: one owner per attribute, one process for changing it, and a record of changes. Not a committee, not a policy document. If nobody can tell you who is allowed to change a product's category, the agent has no stable ground to stand on and its decisions will drift for reasons no one can reconstruct.
Automated planning, so the decision has somewhere to land
An agent that decides something has to hand that decision to a process that executes it. If the plan is assembled by hand each cycle, the agent's output becomes an input to a manual step, which means it becomes a suggestion, which means it will be overridden whenever the week is busy.
Automating planning first is what makes the difference between an agent that decides and an agent that advises. The distinction is not a matter of the model.
The one test
There is a single question that tells you which layer you are actually on. Ask the same question of the data twice, a week apart, and see whether you get the same answer.
If the number changes and nobody can explain why, you are below the warehouse layer. If the number is stable but two departments disagree about what it counts, you are below the master data layer. If everyone agrees on the number but acting on it requires a person to rebuild a plan, you are below the planning layer. If none of those happen, you are ready for agents, and you are also in a much smaller group than the industry conversation suggests.
I have done this sequence twice in two industries. In airline operations the same groundwork turned a twelve-hour analysis into a one-hour one and surfaced $700K a year in fuel savings, which came from validating how fuel was actually used rather than from any model. The method did not change when the industry did.
Why the boring part decides everything
Foundation work is hard to fund because it produces no demo. There is no meeting where governed master data is impressive. The payoff arrives later, disguised as the absence of problems: decisions that do not have to be checked by hand, results that reconcile the first time, an agent whose output nobody argues with.
The honest way to fund it is to stop presenting it as infrastructure and start presenting it as the reason the last three initiatives did not survive contact with the operation. That framing is usually accurate, and it is usually already in the room.
What to do once the foundation exists is note 01. How to prove that any of it worked is note 02.