Buying happens in milliseconds now, or it will soon enough that the distinction stops mattering. Consumers already ask a large language model to compare prices and features before they purchase anything. The next step, already underway at several frontier labs, is letting that same model finish the transaction on the buyer's behalf. Marketing organizations that segment an audience on Monday and launch a campaign on Wednesday have no way to be present for a decision an agent closes in the time it takes to load a product page.

Databricks's answer arrived at the Data + AI Summit in San Francisco this month: CustomerLake, a full customer data platform, the shorthand for the category is CDP, built natively inside the Databricks lakehouse rather than sold as a separate system. Identity resolution, audience segmentation, campaign activation and personalization all happen in the same governed environment where the company's artificial intelligence models already run. The architecture argument is sound. Whether marketing organizations will buy a CDP from a data infrastructure company is the harder question, and it is one Databricks has never had to answer before.

Agents Don't Wait for a Campaign Calendar

Tasso Argyros, the Databricks vice president of engineering who led the CustomerLake build, has been candid about the premise behind it in recent interviews. His view: the CDP, as middleware, is going away, because agents collapse the layers it used to bridge. The agentic buying shift becomes the forcing function. If a buyer's agent operates in milliseconds, marketing that takes weeks to plan and execute is invisible by the time it arrives.

Databricks calls its proposed alternative an infinity campaign: a continuous engagement loop that adjusts in real time instead of a static journey built for a segment of a million people. The campaign agent inside CustomerLake is meant to analyze, personalize and act at the same speed the buyer's agent makes its decision. One marketer, one campaign agent, multiplied across every customer at once.

That premise has not been pressure-tested anywhere yet.

"I think the CDP, as middleware, is going to go away."

Tasso Argyros, VP, Engineering, Databricks

Middleware Is What Agents Erase

The part of this launch that rarely makes the coverage is architectural, not conceptual. Traditional CDPs sit between the data warehouse and the execution tools, functioning as middleware. Data comes in, profiles get assembled, audiences get pushed out, and every hop between those systems adds latency that an autonomous agent cannot tolerate.

CustomerLake collapses that chain. The CDP lives where the data already lives. Identity resolution, enrichment, segmentation and activation happen inside the same governed environment, with no data movement, no duplication of personally identifiable information, and no separate governance layer to reconcile against the warehouse's own rules.

For an enterprise already running analytics and machine learning workloads on Databricks, this removes an entire category of integration work. For the data teams who serve those enterprises, it eliminates the handoff tax between the data being ready and the campaign going live.

Marketers Don't Log Into a Lakehouse

Rick Schultz, Databricks's chief marketing officer since 2017, walked through how his own team uses the platform internally during a session at the summit, reverse engineering the keynote demo for a room of marketing practitioners. The session existed for a reason. Databricks needs proof that CustomerLake is not just infrastructure wearing a marketing costume.

Marketers pay close attention to marketing companies. They have recently started paying attention to large language models. They are not, as a rule, opening a lakehouse interface before lunch.

The marketers who do tend to work inside enterprises with large data teams, places where the marketing organization has already learned to lean on what the data organization builds. That population is real but narrow, and CustomerLake's growth depends on reaching past it, toward marketers who think in campaigns and audiences rather than notebooks and schemas.

CustomerLake's answer is to put agents at the interface layer instead of asking marketers to learn the platform underneath. Profile agents turn raw data into business-ready customer profiles. Campaign agents build audiences, recommend next actions and activate across channels, while the marketer stays in the strategist's seat. Whether that abstraction holds up under a busy marketing calendar, with people who never wanted to think about data governance, is the open question.

Acxiom and Bloomreach Cover What Databricks Cannot Build Alone

CustomerLake launched with more than twenty ecosystem partners, but two of them say more about what Databricks knows it lacks than the announcement itself does.

Acxiom supplies the identity layer. Its Real ID product, built on a graph the company describes as covering 260 million addressable individuals in the United States, now runs as a native application inside Databricks. No personal data has to leave the workspace to use it. The resolved identity becomes the reference point every campaign agent works from, and Acxiom is a name enterprise marketers already trust, which gives them a familiar anchor inside an unfamiliar environment.

Bloomreach supplies the execution layer. Its Loomi platform, trained on what the company describes as more than twelve years of ecommerce interaction data, connects to CustomerLake to push personalized campaigns out across email, the web, text messaging and other channels. Bloomreach closes the last mile Databricks does not own: the channel-specific delivery and tuning where a marketer sees whether a campaign worked.

The division of labor is clean enough to read like a diagram. Databricks owns the data foundation and the orchestration layer sitting on top of it, while identity and execution come from elsewhere. Acxiom resolves who the customer is. The message itself, what channel it goes out on and when, belongs to Bloomreach. None of the three pieces substitutes for the other two, which is the point: a marketer running a campaign through this stack never technically has to leave the Databricks ecosystem to do it.

Personalization Was Always a Data Problem Before It Was an AI Problem

The best use of artificial intelligence in marketing was never going to be a sharper subject line. It was always going to be getting the right message to the right person at the moment that message is relevant, which depends entirely on whether the underlying data is right in the first place. Identity has to be resolved. Context has to go beyond knowing which static segment a customer was assigned to six weeks ago.

Most marketing tools have spent the past two years bolting AI onto data foundations that were never built to support it. Databricks has the foundation. What it has never had is the marketer's attention.

CustomerLake is a bet that the foundation wins, that the platform already holding the data, the models and the governance becomes the platform where marketing runs, and that CDPs built apart from that foundation become a layer enterprises eventually compress out. One industry forecast already puts a number on the shift: by 2030, eighty percent of newly deployed enterprise customer data platforms are expected to be embedded in or composable with the underlying data platform, rather than sold as standalone software.

Nobody Has Run This in Production Yet

CustomerLake shipped in private preview. The early customers Databricks has named, HP, Circle K, AB InBev and Getnet by Santander, are still standing the product up. The infinity campaign concept reads as compelling on a keynote stage. Whether it performs inside a real marketing organization, with its existing tools, agency relationships, team politics and budget cycles already in motion, is unproven.

The message also has to land with chief marketing officers, not just chief data officers. If CustomerLake stays a product that data teams champion and marketing teams merely tolerate, the platform will not reach the scale Databricks is describing. It needs to be explained in the language marketers already use, acquisition cost, lifetime value, campaign velocity, not in the language of lakehouse architecture.

Databricks appears to know this already. The partner strategy and the agent-first interface design both say as much. Whether execution catches up to that intent is the story worth tracking over the next year.