Data Warehouse + lineage graph (foundation)

A warehouse that lands raw events from billing, CRM, support, helpdesk, telephony, and 1C-style accounting plus a lineage graph that lets agents join entities across those sources. A separate "golden record" mapping reconciles merchant/customer IDs across systems so an agent asked "is account X churning?" can stitch transactions + tickets + login telemetry without manual joining.

A large payments processor sized this as the single foundation that unlocks churn prediction, support copilots, and merchant self-service — none of which can be built reliably without it. Cost: low-six-figure USD equivalent. Without it, every downstream AI initiative reduces to a demo that breaks in production because the agent can't resolve identities.

1. Inventory every source. Billing, CRM, HelpDesk, telephony, 1C, product DB, ad platforms.
2. CDC each source into the warehouse with no transformation (Fivetran / Airbyte / custom).
3. Build a golden_record table that maps source IDs → canonical entity (merchant_id, customer_id, deal_id).
4. Author a lineage graph as a YAML or markdown file: which downstream table depends on which source, with column-level mapping for the load-bearing fields.
5. Stand up the semantic / metric layer on top (see #142).
6. Define a "source freshness" SLA per table; surface in a dashboard the agent reads before answering.
7. Ship Q1 MVP for one critical question (e.g., churn). Q2 add SSoT. Q3-Q4 extend coverage.

• Vendor estimates run 3–4× inflated; one team challenged a vendor quote that was inflated ~4× over the independently-rebuilt cost. Always re-cost AI infra independently before signing.