Databricks CustomerLake now runs a campaign agent on a copy of your data. What does the copy leave out?
Build the agent that acts in your workflow inside your product, and keep the warehouse sync for what your customers report on and build audiences from. In October 2026 Databricks put CustomerLake Campaign Agents into Beta and Snowflake put Horizon Context into public preview, so your customer's data platform now runs agents and a context layer on the copy of your data that your product syncs there. That copy carries rows. The workflow state, the rules on your write path, the outcome of each action and your users' corrections stay in your product.
- 100
Audiences Wine.com syncs daily to Braze from CustomerLake, customer-reported.
Source: Databricks, 19 Oct 2026 - 18
Metadata connectors in Snowflake Horizon Context, reaching tools and platforms beyond Snowflake.
Source: Snowflake, 12 Oct 2026 - 80%
Net-new enterprise CDP deployments Gartner expects to be embedded in or composable with data platforms by 2030.
Source: CMSWire reporting Gartner, 20 Oct 2026
What did Databricks and Snowflake ship in October 2026?
Databricks moved two parts of CustomerLake forward on 19 October 2026: Campaign Agents entered Beta and Profile Agents entered Private Preview, and a customer can adopt either one on its own. Access runs through Databricks account teams. The offering itself was announced on 16 June at Data + AI Summit. On the Snowflake side, Horizon Context, a governed context layer within Horizon Catalog, entered preview on 12 October 2026.
The results Databricks published with the 19 October beta announcement are customer-reported. Wine.com ran production email campaigns 23 days after kickoff, with 100 audiences syncing daily to Braze across 6.5 million profiles. HP reports its data was ready for audience building within 2 hours, against 2 to 3 months on a traditional CDP.
HP reports data ready for audience building in 2 hours, against 2 to 3 months on a traditional CDP
Time until data is ready for audience building, in days, as reported by HP
Snowflake moved on the context side. Its 12 October release note introduced Horizon Context in preview, extending Horizon Catalog beyond Snowflake with 18 metadata connectors that reach Power BI, Tableau, dbt, Databricks, BigQuery and others, and column-level lineage across platforms. Snowflake's 20 October blog post put those connectors into public preview and describes Horizon Context as a governed context layer within Horizon Catalog that records what data means and where it came from, readable by people, BI tools and AI agents. Our reading is that both platforms now want to own the agent and the context it reads, inside the customer's own account.
Who in your customer's company now decides where agents run?
Your customer's data and platform lead decides, because that is the person who buys from the data platform. CustomerLake is reached through Databricks account teams, governed by Unity Catalog, and priced on consumption rather than as a software licence, as Databricks put it in June. CMSWire reported on 20 October that no list price or general availability date had been announced.
Analysts see control moving the same way. David Raab of the CDP Institute told CMSWire the change hands IT more control. Gartner, as reported by CMSWire on 20 October, expects 80% of net-new enterprise CDP deployments to be embedded in or composable with data platforms by 2030.
For a vertical SaaS company, this changes who asks the question. The person asking where your agent should run now holds the data platform budget and expects an architectural answer: which data the agent reads, where it acts, and what crosses between the two systems.
Which of your data does the platform's agent read, and how did it get there?
The platform's agent reads what has been synced into your customer's account, and for a marketing or customer engagement product much of that is your own export: contacts, events and campaign results. When Databricks announced CustomerLake in June it described two-way pipelines to the marketing and advertising stack, and named Braze, Bloomreach and Iterable among its partners, so data moves in both directions.
The published results show the cadence. Wine.com's audiences sync daily, and Databricks says Profile Agents update profiles in micro-batches. As CMSWire reports, Treasure AI argues that real-time activation will decide who owns orchestration. That leaves the question of what your product holds that the rows do not.
What does a warehouse copy of your data leave out?
A warehouse copy carries rows, and four things an agent acting inside your product learns from do not arrive as rows:
- Workflow state. A send halfway through approval, or a campaign flow paused for one segment while someone checks the list.
- The write path and its rules. Consent, frequency caps, the customer's own suppression logic, and the reason a send was held back.
- The outcome of each action as your product saw it. What went out, to whom, and what happened next.
- Corrections. The edit a marketer made to the agent's draft before approving it, and the subject line they rejected.
Snowflake is precise about how far meaning travels in its own layer. Its 20 October post notes that description propagation "stops when values are transformed, combined or aggregated." An export is a transformation, so the meaning your product gives a field does not travel with the export by default, and the customer's catalog sees the column without the rule behind it.
The platforms treat human correction as a requirement too. Databricks describes Profile Agents as taking "manual data steward overrides", which is the same need expressed on the profile side.
What the sync carries, and what stays in your product
An agent learns from the edit a marketer makes before approving its draft, so it has to run where that edit happens.
An agent improves through a loop of proposal, rule, correction and outcome. In a marketing product each step of that loop happens at the moment of the write, inside the product that performs the write, and a warehouse copy receives the result some hours later as rows.
What should keep flowing to the customer's data platform, and what should stay in your product?
Keep sending everything your customer reports on and builds audiences from. They own that data, and the sync is a feature they rely on. Keep the write path, its rules and the agent's own loop inside the product, and treat anything that comes back through a two-way pipeline the way you treat a user's action: it passes the same consent, cap and suppression checks before anything is sent.
Ali Ghodsi, Databricks' chief executive, put the direction well in June: "Marketing stops being a series of campaigns and becomes a continuous loop." We agree, and the feedback in that loop lands in the product where the actions happen.
Snowflake's design offers a pattern worth copying. Horizon Context separates seeing metadata from querying data. A vendor can apply the same split to its own share: publish descriptions and lineage for every exported table, so the customer's catalog reads your columns correctly, while the workflow internals stay where the agent works.
| Data | Send to the customer's platform? | Where the agent reads it |
|---|---|---|
| Contacts and profile attributes | Yes, the customer owns them | Either; the copy suits audience work |
| Engagement eventsSends, opens, clicks, deliveries | Yes, for reporting | Product for the latest state, copy for history |
| Campaign results | Yes | Either |
| Audience definitions | Yes, both ways | Product, after the return passes your rules |
| Workflow state in flight | As a snapshot if asked | Product, live |
| Write-path rules and why a send was held | Descriptions and lineage, as metadata | Product, at the moment of the write |
| Agent proposals, edits and approvals | As a log if the customer asks | Product, where the correction is made |
Our recommendation, to adapt to your own export schema.
If you own the AI roadmap
Pull your warehouse export schema, or the reverse ETL connector spec your customers use. Mark each table as reporting and audiences, or as write path, rules, outcomes and corrections. Then pick the first agent feature that reads the second group and take it to your CPO for this quarter's roadmap.
Where does the AI agent in your product read customer data today?
Anonymous. One vote per browser.
Where does SynOS fit?
SynOS runs inside your product and your cloud, white-labelled, and builds a Brain for each customer from that customer's data and procedures. The agent works on your write path, and every write it proposes is recorded rather than executed until someone approves it. Each run is traced and scored, so the corrections and outcomes accumulate on your own infrastructure in the shape a training set is built from, while your sync to the customer's data platform keeps running as it does today.
Questions about CustomerLake and your warehouse sync
What is Databricks CustomerLake?
CustomerLake is a customer data offering from Databricks, announced on 16 June 2026. Its Campaign Agents entered Beta on 19 October 2026 and its Profile Agents entered Private Preview. It is governed by Unity Catalog and priced on consumption, and as of 20 October no general availability date or list price had been announced.
What is Snowflake Horizon Context?
Horizon Context is a governed context layer within Snowflake's Horizon Catalog, in public preview since October 2026. It adds business meaning and column-level lineage across platforms to metadata, through 18 metadata connectors, for people, BI tools and AI agents, and it keeps permission to see metadata separate from permission to query data.
Is the CustomerLake sync two-way?
Yes. When Databricks announced CustomerLake on 16 June 2026 it described two-way pipelines to the marketing and advertising stack, naming Braze, Bloomreach and Iterable among its partners, and in results Databricks published on 19 October, Wine.com reports 100 audiences syncing daily to Braze. Treat anything that comes back into your product, through that pipeline or a reverse ETL job, the way you treat a user's action.
Should we stop syncing our data to customers' warehouses?
No. Your customer owns that data and uses the sync for reporting and audiences, so keep it running. Decide instead which agent features read the warehouse copy and which read live product state, and route anything written back to your product through the same rules a user's action passes.
What does an agent inside our product need that the warehouse copy lacks?
Four things: workflow state, the write path and its rules, the outcome of each action, and the corrections users make to the agent's work. All four are produced inside your product at the moment of the write, and they change faster than a daily or micro-batch sync carries them.
Sources
- CustomerLake enters Beta, Databricks blog (19 October 2026; Campaign Agents in Beta and Profile Agents in Private Preview, adoptable independently; customer-reported results from Wine.com and HP; micro-batch profile updates; manual data steward overrides)
- Databricks CustomerLake enters Beta as enterprises go live with agentic CDP, CMSWire (20 October 2026; access through account teams, no list price or general availability date, consumption pricing wording from June, Unity Catalog, David Raab on IT gaining control, Gartner 2030 projection, Treasure AI's view that real-time activation decides who owns orchestration)
- Databricks enters the marketing industry with CustomerLake, Databricks press release (16 June 2026; announcement at Data + AI Summit, two-way pipelines to the marketing and advertising stack, Braze, Bloomreach and Iterable in the partner list, consumption pricing, Ali Ghodsi quote)
- Horizon Context (Preview), Snowflake release notes (12 October 2026; Horizon Context in preview, 18 metadata connectors, column-level lineage across platforms)
- Horizon Context: business knowledge for AI, Snowflake blog (Shinji Kim with Danielle Kucera) (20 October 2026; public preview of the metadata connectors, steward permissions and description propagation, governed context layer for people, BI tools and AI agents, metadata visibility separate from data access, description propagation stops when values are transformed, combined or aggregated)
Results for Wine.com and HP are reported by those customers and published by Databricks. All sources checked 3 November 2026.