Salesforce Data 360 Agentic Enterprise Reference Architecture

Most customer data platforms stop at the unified profile: harmonize the records, publish a segment, hand it to a campaign tool. Data 360 is built to go one step further — the same profile that powers a segment is also what an Agentforce agent queries mid-conversation, which is why identity resolution and governance sit underneath everything rather than beside it.

How it works

Data enters from Salesforce apps, external systems and MuleSoft, or through zero-copy connectors that read data in place rather than duplicating models and pipelines someone already built. Everything lands as Data Lake Objects in its native schema, then maps onto Data Model Objects on the C360 schema — over 300 industry-agnostic objects, extendable for a specific domain. Unstructured content takes a parallel path: chunking, indexing and embedding into UDLOs so it’s searchable alongside structured records.

Harmonization is where the platform earns the name. Match rules group records that likely belong to the same person — fuzzy, exact or normalized, weighted by recency or source trust — and reconciliation rules pick the best value from each group. The result is a gold-standard unified profile, with link tables preserving the trail back to every source record rather than collapsing them.

Everything above the unified profile reads from it: waterfall and nested segments, calculated insights, Einstein Studio predictions and bring-your-own models, and queries over SOQL, SQL or vector search. Query and Act is the layer that turns a unified record into a decision — an Agentforce topic, an activation to an ad platform or Marketing Cloud, a triggered flow, or a zero-copy data share back to a partner system.

The top layer is what makes this an enterprise architecture rather than a single-org one: a Data 360 Home Org shares data spaces with Companion Orgs and exchanges metadata with other Home Orgs and external LLMs, so the unified profile isn’t trapped in the org that first ingested it. Running the full height beside all of it, the foundational rail — governance, the Einstein Trust Layer, sandboxes, Data Cloud One — is what keeps a fast-moving agent from also being an ungoverned one.

The original design

The diagrams above are built from the Data 360 architecture concept below, designed by Jordan Seeley at Salesforce.

Salesforce Data 360 Agentic Enterprise Reference Architecture, July 2026 — design credit: Jordan Seeley, Salesforce. View in Figma.

What to get right

• Reconciliation rules decide what “true” means. Two source systems that disagree on a customer’s email or phone number resolve however the reconciliation rule says to — get the trust ranking wrong and every layer above inherits a wrong answer with high confidence.
• Don’t copy what zero-copy can federate. The 270+ connectors exist so data already modeled in a partner platform doesn’t need a second pipeline and a second place to go stale.
• Unstructured data needs the same rigor as structured. Chunking and embedding choices for UDLOs are a modeling decision, not a default to accept — they determine what vector search can actually retrieve.
• Design the companion-org boundary before you need it. Shared data spaces and companion connections are how a Home Org scales past one org; bolting that on after launch means re-litigating access to data already in production.
• Governance is not a layer you visit later. Tagging, masking and classification set at ingestion are what make every downstream segment, agent and data share provably compliant instead of merely convenient.