Any Data Cloud figure on a CRM record page is missing a caption: “as of some point in the last several hours.” Lifetime spend on a Contact is the usual example.
That gap is the most useful thing to understand about Data Cloud before you buy it or build on it. The number was computed in a separate store, after matching and aggregation that each run on a schedule, and yet another job wrote it back onto the Contact. Every hop has its own clock and its own meter.
A second store with its own model
The product began as a CDP, was relaunched as Genie, and has been called Data Cloud since last year. You reach it from the same login, through its own app and Setup pages, so it’s easy to assume it’s part of the org’s database.
It’s a separate system. Data Cloud keeps its own copy of the data in its own storage, shaped by its own model. Your Contact rows are ingested through a data stream into a data lake object, then mapped onto data model objects such as Individual and Contact Point Email. The org’s Contact and Data Cloud’s Individual are different records in different systems, kept in step by a schedule.
Three flows, three clocks
I plan a Data Cloud project as three separate flows, because they fail differently and cost differently.
Ingest comes first. Data streams pull from the CRM org and from external sources. Each stream either refreshes on a schedule or takes events as they arrive. The CRM connector is the scheduled kind, so a field changed in the org isn’t in Data Cloud until the next run.
Unify is next. Mapping puts every source onto the shared model. Identity resolution then applies match rules and reconciliation rules to produce unified individuals. Calculated insights aggregate over those profiles on their own schedule. None of this is visible to a rep yet.
Activate is the return path, and there are several. A segment published to a marketing channel. A Copy Field enrichment, which writes a Data Cloud value onto a Contact or Lead field. A Related List enrichment that shows Data Cloud rows on the record page without copying them. A data action that fires a platform event or a webhook when a condition is met.
It’s tempting to spend all the design detail on the first two. But ingest and unify only produce a tidy profile inside Data Cloud, which nobody in sales or service will ever open. The return path is where a user sees the result, so it sets the latency they experience. It sets much of the bill too, because a fresher value means more frequent runs, and every run reprocesses rows whether or not anyone looked.
So I size by the return path. Pick the use case, decide how fresh the value has to be on the far end, and work backward to the stream schedule and insight schedule that freshness requires. Then price that.
Five million profiles, two use cases
Larkspur Outdoor is a retailer of my own invention, with 1.8 million contacts in the CRM, 2.6 million web shopper profiles and 640,000 loyalty members. That’s a little over 5 million source profiles. Matching on normalized email and normalized phone resolves them to 3.1 million unified individuals, so roughly 38 percent of the source records turned out to be someone already known.
The first use case is lifetime spend on the Contact page. It’s a calculated insight summing orders per unified individual, copied to a Contact field. The CRM stream runs hourly and the insight refreshes every 6 hours, the shortest schedule a calculated insight offers. A rep can therefore be looking at a figure up to 7 hours old, before the copy job adds its own delay. For lifetime spend, that’s fine. Nobody changes a call plan because the number is one order behind.
The second use case is abandoned carts. Web events arrive by streaming ingestion, and a data action on the cart event raises a platform event that a flow in the org subscribes to. The flow does the waiting and the check for an order. From cart event to flow start is about 3 minutes.
The two use cases share a platform and a set of profiles, and their latencies are more than a hundred times apart. That’s why I won’t give one number for “how real-time is Data Cloud.”
A data action that publishes a platform event also inherits the usual subscriber problems, the ones in four CRM event anti-patterns.
The flow inventory I fill in before sizing
One row per use case. If a row can’t be completed, the use case isn’t ready to estimate.
| Use case | Sources | Stream schedule | Match rule | Insight | Return path | Expected latency | What drives consumption |
|---|---|---|---|---|---|---|---|
| Lifetime spend on Contact | CRM contacts, order history, loyalty members | CRM hourly, orders daily | Normalized email, or normalized phone | Sum of order value per unified individual, refreshed every 6 hours | Copy Field enrichment to a Contact field | Up to 7 hours, plus the copy job | Rows ingested per stream run, profiles processed per identity resolution run, rows read by each insight refresh |
| Abandoned cart follow-up | Web events, CRM contacts | Web streaming, CRM hourly | Normalized email | None, the condition is on the cart event | Data action raising a platform event, consumed by a flow | About 3 minutes | Streaming events ingested, data action evaluations and events sent |
Argue the consumption column with whoever holds the budget. Had I started that insight on a 12 hour schedule, moving it to 6 would roughly double what it consumes and make the rep’s number six hours fresher. Make that trade on purpose, row by row.
What Data Cloud is not
- A transactional store. Don’t point an order entry screen at it or expect a write to be readable a second later.
- A replacement for the warehouse. Finance models, long history and cross-domain reporting still belong there. What to bring from there into the CRM is its own decision, covered in virtualize or copy.
I’d add one test to any Data Cloud proposal. For each use case, name the record page, flow or segment where the unified data lands, and the age of the data when it gets there. A use case with no landing spot is a data project without a user.