HP and Databricks Partner to Build an AI-Ready Customer Data Platform

At the Data + AI Summit 2026, Luis Alonso, HP's Global Head of Data Strategy and Engineering, shared how HP used Databricks to build an AI-ready customer data platform (CDP) called CustomerLake. This case reveals a typical path for traditional enterprises restructuring their data infrastructure in the AI era, and offers a replicable paradigm for organizations facing similar challenges.

Why Does HP Need a New CDP?

As a global PC and printing giant, HP has a massive number of customer touchpoints: its official website, e-commerce, support services, channel partners, and various marketing campaigns. Data generated by these touchpoints has long been scattered across different systems, forming de facto data silos. Traditional customer data platforms tend to focus on marketing automation, but when AI applications require high-quality, real-time, unified customer views, the old architecture falls short.

In his talk, Luis Alonso emphasized that HP's goal was not simply to upgrade its existing CDP, but to build an "AI-ready" data foundation—one capable of supporting a range of intelligent applications from customer segmentation and personalized recommendations to predictive services. This means the platform must reach a new level in data governance, real-time processing, and machine learning integration.

CustomerLake's Architectural Choice: Why Databricks?

HP ultimately chose Databricks as the core engine of CustomerLake. Databricks' lakehouse architecture combines the flexibility of a data lake with the structured query capabilities of a data warehouse, making it naturally suited to handling the diversity of customer data—from structured transaction records to semi-structured clickstream logs to unstructured customer service conversation text.

More importantly, Databricks natively supports MLflow and Spark, enabling data engineering and data science to collaborate on the same platform. HP's data team can therefore train and deploy models directly on customer data without moving data between multiple systems. This "move compute, not data" model significantly reduces data redundancy and compliance risk.

According to Alonso, CustomerLake was not built overnight but delivered in phases: first unifying customer identity, then building real-time behavioral data pipelines, and finally opening it up to business lines for AI application development. This incremental strategy reduced the risk of large-scale migration.

Lakehouse architecture is an open architecture that fuses data lake and data warehouse capabilities. Data lakes excel at storing raw data in any format at low cost but lack transactional guarantees and efficient SQL querying; data warehouses are rigorously structured and fast to query but have closed formats and high costs. A lakehouse introduces open table formats (such as Delta Lake) and a metadata layer on top of object storage, while providing ACID transactions, schema evolution, time travel, and BI-grade query performance, so that the same data can be shared by data engineering, BI, and machine learning workloads, avoiding repeated movement between lake and warehouse.

MLflow is an open-source machine learning lifecycle management platform from Databricks (now under the Linux Foundation), covering experiment tracking, model registry, packaging, and deployment. It lets data scientists record the code, parameters, metrics, and artifacts of each training run, and publish approved model versions for engineering teams to call via API, thus closing the "last mile" from experiment to production. In HP's scenario, MLflow enables customer behavior models to iterate in a closed loop with real-time pipelines on the same platform.

Structured Streaming is Spark's structured stream processing engine, which treats streaming data as an unbounded table that is continuously appended to, performs incremental computation using the same DataFrame/SQL API as batch processing, and ensures fault tolerance through checkpoints and exactly-once semantics. Compared with the earlier DStream, it greatly lowers the development and operations barrier for streaming pipelines, making it suitable for customer behavior data such as website clickstreams and order events that arrive continuously and require low-latency processing.

The Key to AI Readiness: Unified Identity and Real-Time Pipelines

In HP's practice, unifying customer identity was listed as the highest-priority task. HP has many business lines, and the same customer may appear under different identifiers in different systems. CustomerLake built a unified customer master record by combining deterministic identity resolution rules with probabilistic matching. This step may seem basic, but it directly determines the accuracy of subsequent AI models.

Real-time data pipelines are another pillar. HP used Databricks' Structured Streaming capabilities to bring website behavior, order events, and customer service interactions into CustomerLake in near real time. Marketing teams can therefore trigger personalized outreach based on the latest behavior rather than relying on T+1 batch data.

You may not have noticed that HP emphasized front-loading data quality monitoring in pipeline design—validating and cleaning data when it enters the lakehouse rather than remediating afterward. This design choice reduces the data preparation burden on downstream AI applications.

Governance and Security: The Bottom Line That Enterprise CDPs Cannot Ignore

As a multinational enterprise, HP must comply with multiple data privacy regulations such as GDPR and CCPA. CustomerLake built fine-grained access control and data lineage tracking on top of Databricks' Unity Catalog. Alonso mentioned that governance is not a compliance check added after the fact but part of platform design—every data asset has a clear owner and classification tags, and the training data sources of AI models are traceable.

This governance-first approach allows HP to protect customer privacy while still providing data science teams with enough room to explore. For example, through dynamic data masking, analysts can perform aggregate analysis without accessing personally identifiable information.

Unity Catalog is Databricks' unified data governance layer, which centrally authorizes tables, views, volumes, models, and functions across workspaces and clouds, supports role-based fine-grained permissions (catalog, table, row, and column level), and automatically captures data lineage—that is, the complete flow of data from source tables through processing to downstream models or reports. Lineage is especially critical in compliance audits and troubleshooting: when regulators ask about the training data source of an AI model, lineage can be traced back to the specific upstream assets and their owners.

Dynamic data masking refers to masking or hashing sensitive fields (such as names, ID numbers, and contact information) in real time at query time based on the caller's identity and policies, while the underlying storage retains the original values. Unlike static masking, it does not produce multiple copies, and policy changes take effect immediately. Therefore, it allows analysts to perform aggregate statistics and feature exploration without exposing personally identifiable information (PII), balancing privacy compliance and data usability.

Implications for Enterprises: CDPs Are Evolving into AI Data Foundations

HP's case is not an isolated one. Judging from the trends revealed at the Data + AI Summit 2026, more and more enterprises are repositioning CDPs from marketing tools to AI data foundations. There are two driving forces behind this: first, generative AI applications require high-quality customer context data; second, lakehouse technology has lowered the cost of building unified data platforms.

For enterprises planning similar projects, HP's experience offers several references: first, make identity unification the starting point rather than an ancillary feature; second, choose a platform that supports open table formats (such as Delta Lake) to avoid vendor lock-in; finally, governance and security must be incorporated into the design from day one rather than patched in before launch.

CustomerLake's evolution continues. Alonso revealed that HP's next step will be to explore the integration of real-time feature engineering and model serving, further shortening the path from data to intelligence. This direction also represents the future competitive focus of CDPs—not only storing and querying customer data, but continuously producing AI-usable data products.

(This article is based on and analyzes HP's presentation at the Data + AI Summit 2026.)

One-line Summary

HP built an AI-ready CDP with Databricks lakehouse, with unified identity and real-time pipelines at its core.

Paragraph Summary

Based on the presentation by Luis Alonso, HP's Global Head of Data Strategy and Engineering, at the Data + AI Summit 2026, this article analyzes how HP used Databricks to build an AI-ready customer data platform called CustomerLake. Faced with scattered touchpoint data and the difficulty traditional CDPs have in supporting AI applications, HP chose a lakehouse architecture so that data engineering and data science can collaborate on the same platform, avoiding the redundancy and compliance risks caused by moving data. The platform was delivered in phases: first unifying customer identity resolution, then building real-time behavioral pipelines based on Structured Streaming, and finally opening AI application development to business lines. Governance was built into the design from the start, relying on Unity Catalog for fine-grained access control, data lineage, and dynamic masking to meet requirements such as GDPR and CCPA. The article concludes that CDPs are evolving from marketing tools into AI data foundations, and offers practical recommendations such as prioritizing identity unification, adopting open table formats, and incorporating governance into the design from day one.