The CDP Is Disappearing Into the Data Cloud
- 4 hours ago
- 23 min read
Databricks’ CustomerLake and the Coming Battle for the Decisioning Layer
The Customer Data Platform, or CDP, has spent the better part of the past decade fighting to become the center of the marketing technology stack. It may finally get there—just in time to disappear.
In June, Databricks announced CustomerLake, or what it's calling an “Agentic CDP,” embedded directly into the Databricks platform. The offering combines customer profiles, identity resolution, audience building, campaign automation, activation, AI models, and agents on the same data foundation enterprises already use for analytics and machine learning. On the surface, CustomerLake looks like Databricks officially entering an already crowded CDP market. But I believe something much more significant is happening. Databricks is not simply launching another CDP. It is instead making a direct play to absorb the CDP into the enterprise data and AI platform.
This is the natural continuation of a shift I wrote about recently in “New Center of Gravity for Customer Data: The Data Cloud.” For decades, the CRM was treated as the system of record for the customer. Then digital behavior exploded, customer journeys fragmented, and organizations began generating far more useful customer signals outside the CRM than inside it. As those signals accumulated, the center of gravity began moving toward cloud data platforms capable of storing, processing, governing, and analyzing enormous amounts of behavioral, transactional, operational, and identity data.
Databricks’ announcement suggests the next stage of that transition has officially arrived. The data cloud has already made a strong case to become the system of record. Now it wants to become the system of intelligence—and ultimately the system of action for marketing. If you ask me, this is the real story behind the Agentic CDP.
The data cloud has already made a strong case to become the system of record. Now it wants to become the system of intelligence—and ultimately the system of action.
From Data Gravity to Decision Gravity
For the past several years, much of the CDP debate has centered on data gravity: Where should customer data actually live? Packaged CDPs offered their own customer-data stores, which went something like this: Enterprises copied data into those platforms, created unified profiles, built audiences, and sent those audiences to marketing and advertising channels.
Emerged around 2022 as a response to the rigid, siloed nature of traditional, packaged CDPs, Composable CDPs challenged this model. If an organization’s customer data already lived inside Snowflake, Databricks, Google BigQuery, or another cloud platform, why copy it into a second system? Why not leave the data where it was and add the identity, segmentation, and activation capabilities needed to make it useful to marketers?
The composable model was based on a simple idea: the warehouse or lakehouse should remain the source of truth. As of 2026, it's safe to say this argument has largely been won. Most sophisticated enterprises no longer want customer data scattered across dozens of disconnected marketing platforms, each maintaining its own incomplete and occasionally contradictory version of the customer. They want data consolidated, governed, and controlled within their enterprise data foundation.
But settling the question of where data lives creates a much bigger question: Where will decisions be made? This is the emerging battle over what I would call "decision gravity." It is one thing for Databricks to store customer transactions, behavioral events, product usage, service interactions, consent signals, and identity data. It is another thing entirely for Databricks to interpret that information, decide what should happen next, activate an experience, measure the response, and continuously improve the next decision.
Data gravity asks where customer data lives. Decision gravity asks which platform interprets that data, chooses what happens next, executes the decision, and learns from the outcome.
CustomerLake is Databricks’ attempt to do exactly that. The platform includes profile agents that help prepare customer data and resolve identities. Campaign agents allow marketers to describe goals using natural language and receive recommendations for audiences, messaging, timing, and channels. Its decisioning capabilities are designed to identify the next-best action (NBA) for each customer and continuously adjust based on new behavior and outcomes.
Databricks calls this model “Infinity Campaigns”—continuous, agent-driven engagement loops that replace the conventional process of building and launching one campaign at a time. In other words, Databricks does not want to be another source feeding the marketing stack. It wants to become the intelligence layer deciding what the rest of the stack should do. This is a profound expansion of the data cloud’s hitherto role.
In this paradigm, the data platform stops being passive infrastructure sitting underneath marketing. It evolves to begin planning campaigns, constructing audiences, selecting offers, personalizing experiences, choosing channels, applying guardrails, evaluating results, and learning from every interaction. This is important because the platform that controls these decisions will increasingly control the customer experience itself.
Moreover, if identity, segmentation, decisioning, orchestration, and activation all happen inside the enterprise data cloud, we have to ask an uncomfortable question: What exactly is left for the standalone CDP to do?
Databricks says CustomerLake is designed to replace one-off campaigns with continuous agentic loops capable of delivering personalized customer experiences “a billion times a day.”That is not simply campaign automation. It is an attempt to turn marketing into an always-on decisioning system.
From the Golden Record to Golden Context
For years, the CDP industry promised marketers a “Golden Record”—a single, unified profile containing everything the enterprise knows about a customer. In all fairness, this was a huge improvement over the fragmented customer data that preceded it. Instead of maintaining separate versions of the same person across CRM, ecommerce, loyalty, mobile, service, and advertising platforms, organizations could now resolve those identities and create a more complete Customer 360 (C360). Emphasizing the Golden Record is a big reason why the CDP became an important MarTech category.
But a unified customer profile primarily answers one question: Who is this customer? The issue is an agent needs to answer a much harder one: What should we do for this customer right now, and why?
Knowing that someone is a high-value customer, opened three emails, purchased twice in the past year, and belongs to a particular audience segment is useful. But it is not enough to make a good decision.
An intelligent system also needs to understand what the customer is doing at this moment, what the business is trying to accomplish, what actions have already been taken, which offers have been presented, how the customer responded, which channels are available, and what the organization is permitted to do next.
Databricks calls this expanded understanding “Golden Context.” This distinction matters big time. A Golden Record is largely historical. Golden Context, by contrast, is historical, situational, operational, and strategic. Imagine an airline customer whose profile shows that she flies eight times a year and has elite status. A traditional CDP might place her in a high-value traveler segment and trigger a generic loyalty campaign.
An agent operating with richer context could know that her flight was just delayed, she is traveling with three children, the airport lounge currently has capacity, and she previously complained about a delay that was never properly resolved. Instead of sending another promotional email, the agent could offer lounge access, provide a meal credit, or proactively rebook the family. What s world of difference.
The difference is not simply better personalization. One system recognizes the customer, while the other understands the situation. This is also where the distinction between storing data and using intelligence becomes much clearer. Traditional CDPs were primarily built to unify profiles, create segments, and move audiences into downstream channels. Humans still designed the campaign, established the rules, selected the audience, chose the message, and determined when the campaign would run. The CDP supplied the data, and the marketer supplied the judgment.
An "Agentic CDP" begins to blur that boundary. It does not merely provide information to a marketer. It interprets customer and business context, recommends an action, executes that action within defined guardrails, observes the result, and adjusts what happens next. This represents a fundamental change in the role of the platform. The CDP is no longer just maintaining a record of the customer. It is maintaining a memory of the relationship—including what the organization has attempted, why it made those decisions, and whether they worked.
GOLDEN RECORD Who is this customer? What have they done? GOLDEN CONTEXT What is happening right now? What is the business trying to accomplish? What have we already tried? What are we permitted to do next?
Source: Databricks introduced “Golden Context” as the combination of the customer profile, current business context, and the history and outcomes of previous decisions. Databricks’ Agentic CDP overview
Decision history may prove to be one of the most important and overlooked forms of customer data in the AI era. Without it, agents risk behaving like disconnected marketing, sales, and customer support teams do today. One agent offers a discount while another sends a full-price promotion. A service agent apologizes for a problem while a marketing agent attempts to upsell the customer. A customer declines an offer, only for another channel to present the same offer five minutes later. Each action may appear reasonable when viewed in isolation.
Together, they create an experience that feels confused, repetitive, and occasionally hostile.
Consumers already encounter this problem with today’s marketing automation and siloed teams. Agents could make it dramatically worse because they operate faster, across more channels, and at much greater scale. Golden Context is therefore not simply about giving AI more data. It is about giving the system an institutional memory: what the organization knows, what it wants to accomplish, what it has already done, what the customer allowed it to do, and what happened as a result.
This connects directly to the argument I made in “The AI Race Isn’t About Models Anymore. It’s About Data.” The model may provide the reasoning capability, but the quality of the decision still depends on the data and context surrounding it.
AI is still “smart dumb.” It can make an extraordinarily sophisticated decision based on incomplete, outdated, or incorrect information. It can then execute that bad decision instantly and present it with absolute confidence. The Agentic CDP does not make the customer-data foundation less important.
It makes that foundation impossible to ignore.
The Campaign May Be the Next Thing to Disappear
The marketing campaign is one of those concepts that feels permanent because it has existed for so long. A team develops a strategy, identifies an audience, builds creative, chooses channels, establishes a budget, launches the campaign, measures the results, and eventually starts the process again.
The tools have changed dramatically, but the underlying operating model has remained surprisingly consistent. Campaigns are still planned around calendars, budgets, product launches, seasonal moments, and organizational workflows.
They begin, run, and end. Customer behavior, of course, does none of those things. Customers do not organize their lives around campaign flights. They research products, visit stores, abandon carts, contact service teams, compare competitors, change preferences, revoke consent, and make purchases continuously. Their needs evolve in real time, while most marketing still operates according to plans established weeks or months earlier.
That mismatch has always existed. Until recently, we lacked the technology to do much about it.
Databricks’ concept of “Infinity Campaigns” proposes a different model. Instead of marketers designing a fixed sequence of interactions for a predefined audience, agents continuously evaluate customer signals and determine the next appropriate action based on current context and business objectives. The system does not wait for the next campaign to begin because it is always running.
A customer browses a product but does not purchase. New inventory becomes available. A service case is opened. A subscription approaches renewal. The customer enters a store, redeems an offer, changes a preference, or signals interest in a different category. Each event changes the context. The agent evaluates that change, selects an action, activates it through the appropriate channel, observes the outcome, and feeds the result back into the next decision—collect, understand, decide, act, and learn. Rinse and repeat.
This is less like running a campaign and more like managing an intelligent, continuous relationship.
It may sounds like a subtle distinction, but it is not. Marketing automation platforms generally require humans to anticipate customer behavior in advance, so marketers construct journey maps filled with triggers, branches, delays, exclusions, and decision rules. If a customer does X, send message Y. If they do not respond within three days, move them to path Z.
These workflows can become incredibly sophisticated, but they are still largely deterministic. Someone must imagine the possible scenarios, encode the rules, produce the necessary content, and maintain the journey as conditions change.
Agents shift part of that responsibility from workflow construction to goal-directed decisioning.
Rather than specifying every step, a marketer could define an objective, budget, eligible audience, brand constraints, consent requirements, channel limitations, and performance thresholds. The agent would then determine how best to pursue that objective for each customer, continually adjusting based on response.
In this future, humans would still set the strategy, which would amount to defining the boundaries, approving sensitive actions, monitoring performance, and intervening when the system behaves unexpectedly. But they would no longer need to manually choreograph every interaction. This is not simply a more efficient way to build campaigns. I would argue it changes the basic unit of marketing from the "campaign" to the "decision."
Instead of asking, “How did this campaign perform?” marketers may increasingly ask:
Did the system make the right decision for this customer?
What data and context influenced that decision?
Was the action permitted and consistent with our brand?
Did it improve the customer relationship?
What did the agent learn from the outcome?
That shift also changes measurement. Campaign reporting traditionally aggregates performance across audiences, channels, creative variations, and predetermined time periods. Continuous agentic engagement requires organizations to evaluate millions of individual decisions and understand how those decisions collectively affect conversion rates, customer lifetime value, satisfaction, retention, and brand equity.
The campaign may remain the container marketers recognize. But the decision—not the campaign—will increasingly become the fundamental unit of marketing.
The campaign will not disappear tomorrow. Brands will still launch products, sponsor events, create major creative platforms, and organize marketing around important commercial moments. Nor are humans are not going to stop developing strategy or big ideas. But beneath those visible campaigns, an increasingly autonomous decisioning layer could continuously determine how each customer experiences the brand. The campaign may remain the container marketers recognize, but the agent will decide what actually happens inside it.
The CDP Market Is About to Feel Pain
Databricks is entering a CDP market that is already fragmented, crowded, and increasingly difficult to define. Packaged CDPs built the original category by combining data collection, identity resolution, customer profiles, segmentation, and activation within a single platform. Composable CDPs then challenged that architecture by leaving data inside the enterprise warehouse or lakehouse and adding the tools marketers needed to use it.
Application suites such as Salesforce and Adobe incorporated CDP capabilities into their broader marketing clouds, and customer engagement platforms like Braze moved deeper into data and decisioning. Meanwhile, cloud data platforms like Snowflake and Databricks continued expanding upward into analytics, machine learning, governance, and increasingly marketing.
CustomerLake brings these competitive lines together. If Databricks can provide customer profiles, identity resolution, marketer-friendly audience creation, AI decisioning, campaign orchestration, and activation directly on top of the lakehouse, it places pressure on almost every layer of the existing customer-data ecosystem.
The most obvious pressure falls on standalone packaged CDPs. Their core proposition has always been that marketers need a purpose-built platform to unify customer data and make it accessible. But if the same capabilities can be embedded directly into the enterprise data foundation—without copying customer data into another proprietary store—that separate platform becomes much harder to justify.
Why pay to maintain another customer database when the governed data already exists somewhere else?
Why reconcile another version of the customer? Why create another place where identity, consent, and data quality can drift out of sync? Packaged CDPs will still have advantages. Many have mature marketer interfaces, proven real-time capabilities, established identity frameworks, consent integrations, journey orchestration, and years of experience handling complicated enterprise deployments.
CustomerLake is still in private preview. A product announcement is not the same thing as a production-ready enterprise platform.
But the strategic direction is difficult to ignore. Data duplication, implementation complexity, and overlapping functionality have long been weaknesses of the packaged-CDP model. Databricks is attacking all three. Composable CDPs face a more complicated threat.
Companies such as Hightouch helped establish the argument that the warehouse or lakehouse should be the source of truth. They allowed organizations to build audiences and activate data without copying it into a separate CDP. In many respects, their success helped validate the architecture Databricks is now embracing. Hightouch already offers audience management, identity resolution, reverse ETL, event collection, and AI decisioning on top of Databricks. It can send data to more than 300 destinations and has built marketer-friendly capabilities that the underlying data platforms historically lacked. The relationship is therefore both collaborative and competitive.
Databricks still needs a broad ecosystem of activation, identity, engagement, and advertising partners. No enterprise data platform is going to replace every specialized marketing application, nor should it try.
At the same time, every native capability Databricks adds reduces the amount of functionality an organization needs to purchase elsewhere. This is the central tension of the platform economy: partners help make the platform more valuable while the platform steadily expands into the territory those partners created.
Customer engagement platforms are not going away either. Brands will still need technology to deliver emails, text messages, mobile notifications, advertisements, web experiences, and customer-service interactions. Specialized platforms often have deep execution capabilities that would be expensive and unnecessary for Databricks to reproduce. But their strategic position could certainly change. The engagement platform may deliver the message while Databricks determines the audience, timing, offer, channel, and objective. In that scenario, execution remains distributed, but intelligence consolidates inside the data cloud. That is why decision gravity matters so much.
The most strategically valuable platform will not necessarily be the one that sends the email or displays the offer. Instead, it will be the one that determines why that action should occur, authorizes it, measures the outcome, and decides what happens next.
The underlying economic model makes this potentially even more disruptive. According to CDP.com’s analysis of CustomerLake, Databricks can monetize the product through the compute and storage consumed underneath it rather than depending on a traditional standalone platform fee. In other words, CustomerLake does not necessarily need to become a large independent software business. It needs to make the Databricks platform more valuable and drive greater use of the underlying infrastructure.
Standalone CDP providers do not have that luxury. Their CDP must pay for its own engineering, sales, support, infrastructure, and product development.
Databricks can potentially treat the CDP layer as an accelerant for its much larger data and AI business.
That creates a structural pricing asymmetry, not merely another feature comparison. CustomerLake does not need to be better than every CDP in every category on day one. It only needs to be sufficiently capable for enterprises already committed to Databricks—and improve quickly enough that buying another major platform becomes difficult to defend. This is what should make the CDP market uncomfortable.
Databricks is not necessarily competing for a place in the marketing stack. Instead, it is betting that the marketing stack will increasingly be built around Databricks.
Agents Make Trusted Data More Important
There is a temptation whenever a new AI product arrives to focus on what the agent can do.
Can it build an audience, select an offer, resolve an identity, choose a channel, or personalize a message? Can it optimize the next interaction? These are important questions, but they skip over a more fundamental one: Should the agent trust the data it has been given?
An Agentic CDP can only be as intelligent as the customer foundation beneath it. If identities are incorrectly matched, consent records are outdated, behavioral signals arrive late, or customer profiles contain conflicting information, the agent will not magically repair the enterprise. Worse, it will automate its mistakes.
This is the central risk in moving from data management to autonomous decisioning. Bad customer data has always produced wasted media, irrelevant communications, inaccurate reporting, and poor experiences. Garbage in, garbage out. But historically, these problems moved through relatively slow, human-operated workflows, so errors could often (but not always) be caught before any damage was done. Agents change the speed and scale of the consequences.
In the past, a flawed audience might once have affected a single campaign with a few touchpoints at most. An agent operating continuously could use the same faulty identity or incorrect preference across email, mobile, advertising, web personalization, customer service, and other channels—all within seconds. The system could even interpret the customer’s negative response as a new signal and make another bad decision based on it, compounding the problem. That is how a data-quality problem easily becomes a self-reinforcing customer-experience problem. Identity resolution is therefore not merely a CDP feature, but a prerequisite for safe decisioning.
CustomerLake includes what Databricks calls Agentic Identity Resolution, combining deterministic and probabilistic matching with AI-assisted workflows. Agents can identify ambiguous records, surface edge cases, recommend matches, and use feedback to improve resolution logic over time. There is real promise in this approach. Enterprise identity resolution still involves substantial manual tuning, inconsistent rules, and difficult trade-offs between reach and precision. Agents could soon be helping data teams identify problems faster and continuously improve the quality of customer profiles.
But identity resolution is not a problem where confidence should be confused with correctness.
An agent may determine that two records probably belong to the same person. The business still needs to understand how that conclusion was reached, which data supported it, how confident the system is, and what happens if the match is wrong. Merging two website sessions incorrectly is one thing. Combining healthcare, financial, location, or purchase data belonging to two different people is something else entirely. Think an embarrassing event versus a lawsuit.
An Agentic CDP without trusted identity, current data, enforceable consent, and clear governance is not intelligent marketing. It is automated data malpractice.
The more sensitive the data and consequential the decision, the more important explainability, review, and reversibility become. Consent adds another layer of complexity. In my recent article, “The AI Race Isn’t About Models Anymore. It’s About Data,” I argued that consent is becoming AI’s permission layer. It determines not only what data an organization possesses, but what an AI system is allowed to know, infer, recommend, and do in a particular situation.
An Agentic CDP turns this idea into an operational requirement. Before an agent activates an audience, personalizes an experience, selects an offer, or contacts a customer, it must know whether the organization is permitted to use the underlying data for that particular purpose. This decision may depend on jurisdiction, channel, data category, customer preference, contractual restrictions, and how the data was originally collected.
Consent cannot remain a static record stored somewhere in the compliance stack. It must become part of every decision. The same is true of governance. Organizations will need to determine which actions agents can take autonomously, which require human approval, and which should never be allowed. They will need limits around budgets, channels, frequency, pricing, offers, customer segments, sensitive attributes, and the types of inferences agents can make.
They will also need auditability. If an agent decides to suppress a customer, extend a discount, change an audience, or select one offer over another, the organization must be able to reconstruct that decision. This capability seems non-negotiable to me. What information did the agent use? Which objective was it pursuing? Which rules applied? Was consent valid? What alternatives did it consider? Who approved the governing policy? “AI made the decision” will not be an acceptable explanation to a customer, a regulator, or an executive team.
Real-time data presents another challenge. Databricks correctly argues that agents cannot operate on yesterday’s customer profile. A next-best-action engine is not particularly useful if it recommends a product the customer purchased an hour ago or sends a promotional message while an unresolved service complaint is escalating.
But real-time activation requires more than streaming events into the lakehouse. Identity, consent, inventory, service status, interaction history, and business constraints must all be updated and available quickly enough to influence the decision. The system has to be current, not merely comprehensive.
This is where the idea of Golden Context becomes both powerful and demanding. Creating a unified customer profile is already difficult. Maintaining a live, governed, permission-aware understanding of the customer and every decision surrounding that customer is substantially harder.
Big picture, the Agentic CDP does not necessarily eliminate any of these foundational problems.
Instead, it raises the cost of failing to solve them. An Agentic CDP without trusted identity, current data, enforceable consent, and clear governance is not intelligent marketing. It is something more akin to automated data malpractice. No bueno.
What Remains Unproven
CustomerLake is an ambitious product and a strong statement about where Databricks believes marketing technology is heading. But it is still in private preview, and this distinction matters. The architecture is compelling. Keeping customer data within the governed enterprise data foundation sure makes sense. Fee will argue that giving marketers easier access to that data is a bad idea sense. And using agents to reduce manual work, surface insights, and improve decisioning also makes sense.
Of course none of that guarantees enterprises are ready to hand meaningful control of customer relationships to autonomous agents. The first question is whether CustomerLake will actually work for marketers, which is the point. Cloud data platforms have traditionally been designed for data engineers, analysts, data scientists, and technical teams. Their power comes from flexibility, openness, and the ability to work with enormous volumes of complex data.
Anyone who has worked with marketers will confirm that marketers generally need something different. They need intuitive interfaces, understandable workflows, reusable audiences, visual journey tools, clear performance reporting, and the ability to move quickly without submitting a request to the data team. Databricks plans to bridge this gap through natural-language interfaces and campaign agents. A marketer should be able to describe an objective, ask questions about an audience, and instruct the system to construct and activate a campaign without writing SQL.
All in all, it's an exciting vision. But it's a vision that needs to be proven with actual marketers under actual deadlines, dealing with actual constraints. Natural language can simplify a complex interface, but it does not necessarily simplify the decisions underneath it. A marketer still needs to understand what audience was created, which filters were applied, what data was excluded, how identity was resolved, whether consent was enforced, and why the agent recommended a particular action. A conversational interface cannot become a black box with a friendly text field.
The second unanswered question is whether the platform can deliver genuine real-time customer decisioning at enterprise scale. “Real time” is among the most overused (and abused) terms in marketing technology. It can mean milliseconds, seconds, minutes, or simply faster than whatever process existed before. The appropriate threshold also varies by use case. Personalizing a web session, suppressing a recent purchaser from paid media, resolving an identity, and adjusting a loyalty offer all operate on different timelines.
Real-time definitions aside, CustomerLake will need to demonstrate that it can ingest signals, update profiles, evaluate consent, make a decision, and activate that decision quickly enough for the interaction at hand. A real-time interface sitting on top of delayed data is still a batch system wearing a nicer outfit—putting lipstick on a pig, if you will.
The third challenge is orchestration across the broader marketing ecosystem. Large enterprises do not operate through a single platform. They use CRMs, service systems, commerce platforms, consent-management tools, email providers, mobile-engagement platforms, advertising technology, content systems, analytics tools, identity partners, and dozens of other applications.
CustomerLake includes connectors and an ecosystem of partners, but coordinating decisions across those systems is much harder than passing audiences between them. What happens when two agents pursue conflicting objectives? Which platform has authority when a service agent wants to protect the relationship while a marketing agent is trying to increase revenue? How does the organization prevent one channel from undermining another? Who maintains the shared memory of what has already happened? The Agentic CDP may become the orchestration layer, but claiming that role and successfully performing it are two very different things.
Measurement presents another unresolved issue. An agent may make millions of individualized decisions across customers and channels. Determining whether those decisions created incremental value will require more than conventional campaign reporting. Did the agent generate a purchase that would not otherwise have occurred? Did a discount sacrifice margin unnecessarily? Did repeated short-term optimizations reduce long-term customer value? Did improved conversion come at the expense of brand perception, customer satisfaction, or channel efficiency?
Agents will optimize whatever objective they are given. The difficult part is defining the right objective and ensuring the system does not exploit an unintended shortcut to achieve it. An agent tasked with increasing conversions may distribute discounts too aggressively. One instructed to reduce churn may overcommunicate with customers. A system optimizing engagement may learn that irritating or sensational content produces more clicks. Marketing has spent years learning that what is measurable is not always what is valuable. Bluster aside, agentic decisioning will not make that lesson obsolete. If anything, it will make the consequences of forgetting it arrive much faster.
Then there is the human operating model. Databricks is clear that humans will remain involved by setting goals, establishing guardrails, reviewing outcomes, and approving sensitive decisions. But most organizations have not yet defined who owns those responsibilities. Is it marketing, data and analytics, IT, legal, privacy, or customer experience? How about a new AI governance team? Moreover, who ultimately decides which actions an agent can take? Who investigates when it makes a mistake? Who is accountable when the technology works exactly as designed but produces an outcome the company never intended?
End of the day, the hardest part of deploying an Agentic CDP may have very little to do with the CDP itself. Enterprises will need new governance structures, approval models, measurement frameworks, and ways of working across teams that have historically operated in silos.
Finally, there is the broader issue of platform concentration. Bringing data, models, governance, identity, decisioning, and activation together can reduce complexity. But it also places an extraordinary amount of control inside a single platform. Organizations should thus be careful not to solve today’s MarTech fragmentation by creating tomorrow’s data-cloud dependency and vendor lock-in.
The winning architecture will need to remain open enough for enterprises to use different models, identity providers, activation platforms, measurement approaches, and agents. Otherwise, the industry may simply replace one generation of closed marketing suites with a more powerful generation built around cloud data infrastructure.
Based on what I've read, I believe Databricks has identified the direction correctly. Customer data, AI, and decisioning are converging, and the traditional boundaries between the CDP and the data cloud will continue to collapse.
But direction is not the same as destination. CustomerLake has given us a credible picture of what the Agentic CDP could become. Now Databricks has to prove that enterprises—and their customers—are ready to live with it.
From System of Record to System of Action
For years, enterprise technology companies competed to become the system of record. CRM platforms wanted to own the customer record. Marketing clouds wanted to own customer engagement. CDPs wanted to unify customer data. Data warehouses and lakehouses wanted to become the governed foundation beneath everything else.
The rise of the data cloud has started to settle part of that debate. Customer data increasingly belongs in an enterprise-controlled foundation where it can be collected, governed, enriched, analyzed, and made available across the organization. Copying that data into another application every time a team needs to use it creates more cost, more complexity, and yet another version of the truth—not to mention thorny legal and compliance issues. This is fundamentally the argument behind the composable CDP, and it is the direction many sophisticated enterprises were already moving.
CustomerLake takes the argument one step further. If customer data already lives in Databricks, the models run there, the governance exists there, and the organization’s agents operate there, why should customer decisions occur somewhere else? This is the transition from system of record to system of action.
The winning platform will not be about maintaining the most complete customer profile. I would argue it will center on interpreting that profile alongside current behavior, business objectives, consent, interaction history, and operational context, and deciding what should happen next, execute that decision through the appropriate channel, measure the result, and apply what it learns to the next interaction.
That is a much larger prize and bigger TAM than the CDP market. It is the enterprise customer-intelligence layer. Databricks is not alone in pursuing it. Snowflake, Salesforce, Adobe, Tealium, Hightouch, and many others are moving toward some combination of unified data, embedded AI, identity, decisioning, and activation. Each approaches the opportunity from a different starting point, but they are increasingly converging on the same territory.
The competition will not be decided by which company uses the term “agentic” most aggressively.
Instead, it will be decided by which platform can provide the most trusted customer context, make the best decisions, activate those decisions across an open ecosystem, and prove that the outcomes create real incremental value.
This is why I do not believe enterprises should rush out and purchase an Agentic CDP simply because the category has a new name. Before they purchase anything, they should start asking harder questions like:
Where does our customer data actually live?
Can we trust it?
Are identities resolved accurately?
Is consent current and enforceable at the moment of decision?
Can customer context be updated quickly enough to support real-time action?
Which decisions can an agent make autonomously?
Which require approval?
Can we explain why an action occurred?
Can we measure whether it created value?
Without satisfactory answers, adding agents will not transform the customer experience. It will only serve to accelerate the organization’s existing problems. The companies that benefit most from the Agentic CDP will therefore not necessarily be the first to buy one. They will be the ones that have built the organizational and technical foundation required to use it responsibly: trusted data, strong identity, dynamic consent, clear governance, reliable measurement, and teams prepared to manage autonomous decisioning.
This connects directly to the arguments I have been making recently. The center of gravity for customer data is shifting toward the data cloud. At the same time, the AI race is increasingly about trusted, governed, real-time data rather than access to a particular model. CustomerLake shows us what happens when those two trends collide. The data cloud stops being somewhere customer information sits, and starts becoming where customer strategy is executed.
Let's be clear, I do not expect the CDP category to disappear overnight. Enterprises have enormous existing investments, and specialist platforms retain valuable capabilities. For the foreseeable future, most organizations will operate hybrid architectures combining data clouds, CDPs, engagement platforms, identity providers, consent tools, and activation partners. But the direction is becoming clearer. The capabilities we once purchased as a separate CDP will increasingly become embedded within the enterprise data and AI foundation. The CDP label may not survive ultimately survive, but the functionality will remain essential, and the standalone platform will become harder to distinguish from the infrastructure surrounding it.
This is why the most important thing about CustomerLake is not whether Databricks has created the best CDP. Better stated, Databricks may have shown us what comes after the CDP. The data cloud has largely won the battle over where customer data lives. Now the battle begins over who gets to decide what happens next.
------------------------------------------------
Rio is an executive with 20+ years at the intersection of strategy consulting, AdTech, data, and media. He's a trusted advisor on customer experience, digital strategy, and marketing transformation. He's a partner at Credera, Omnicom's consulting arm. He's also a podcast host, writer, and public speaker focused on the future of advertising and AI-driven infrastructure.
Comments