Organize the evidence before optimizing the action
Bronze retains source records and history; Silver validates, reconciles and integrates shared entities; Gold packages datasets for a business purpose. These are logical responsibilities, not three mandatory servers. A governed Silver dataset may also support ML. Agree what each consumer needs rather than copying every record through every layer.
A shared definition is an operating agreement
Two systems can both be correct and still disagree: sales may count ordered units while the warehouse counts available stock after reservations. A useful data product states its grain, definitions, refresh deadline and owner. Teams agree how cancellations, returns and late updates change the record. A freshness alert should reach someone who can resolve it, and downstream decisions need a defined response while the issue remains open.
Treat the decision interface as a contract
A clean dataset does not decide whether to accept an urgent order. The decision model still needs contribution assumptions, capacity limits, service commitments and the authority to approve changes. Record the exact input snapshot used, retain the proposed alternatives and compare execution with the approved plan. Start with a shadow workflow: review recommendations alongside existing practice, then expand automation only within an agreed operating scope.
A worked example
Illustrative assumptions for discussion, not a customer case or a performance claim.
Suppose ERP lists 120 units on hand, WMS shows 90 available after reservations, and a planner is considering an urgent order for 40. Bronze preserves both source snapshots and their timestamps. Silver aligns the SKU and stock definitions and identifies the 30 reserved units. A Gold availability product publishes 90 units with its refresh time. The Decision Engine still checks material use, capacity and existing commitments before proposing an action. If the WMS snapshot exceeds the agreed age, the policy may hold the recommendation for review rather than treat 90 as current.
Questions to take into a pilot
Which business decision justifies the first data product? Who resolves conflicting definitions and stale sources? Can the team reproduce a recommendation from its data and model versions? Agree quality thresholds, recovery exercises and the response to missing inputs. Evaluate one workflow before expanding the lakehouse; a larger dataset is useful only when its operating purpose and ownership are clear.