APAC Roundtable, May 22, 2025

Guest: Vlad Silik, Chief Technology Officer, carbonix

Topic: CDP as Modern ‘Fuel Tank’ for AI Initiatives

• CDP solves problem of assembling customer data, AI consumes it at scale
  • AI needs data beyond the customer profile for complete context (e.g. product data, etc.)
  • AI changes CDP task from providing single data model (golden record, single customer view) to providing contextual data fragment suitable for the process that’s being executed
  • different processes need different data; invalidates traditional view of single data model available to all users
  • should CDP reformat its data or provide a complete profile and leave reformatting to the applications?
  • CDP includes raw data (e.g. purchases) and derived data (e.g. product categories of interest to this customer)
  • may be harder to extract application-specific data views that is usually realized
  • Less costly to provide carefully chosen fragments rather than full data set
• CDP needs to limit what’s shown to AI so it can understand the context efficiently; requires selecting only relevant data for particular use case
  • Giving AI access to only fragments reduces security and privacy risks, which could be violated by AI if give full access to everything and let it choose what it wants
  • Guardrails exist to prevent AI from unauthorized access (but has been research showing that AI sometimes try to get around the guardrails)
  • Still need humans to define metadata, privacy rules, etc.
• Will always have CRM, ERP, etc. as systems of record/truth/knowledge; but contextual view of customer is fluid
  • Companies have multiple systems of record; role of CDP is to unify these
  • Helpful to have terms such as systems of engagement, of interest, etc. to help non-technical users understand how pieces fit together, how to build use cases
• Scope of CDP may include applications such as marketing automation
• AI plays two roles in CDP
  • AI within CDP should be used to design, construct, configure the tools; check for quality, etc.
  • AI can also be used to create temporary data model to build extracts for specific purposes e.g. customer service
  • This is separate from AI that is applied to manage personalized interactions or do analytics
• Single AI tool that does everything would be very complicated;
  • better to have smaller specialized tools e.g. one to select data for a purpose, another to make recommendations
• Question: in the future, will AI be able to ingest all raw customer data directly and work with it, making CDP unnecessary? 
  • Definitely not now
  • Concern that AI may make mistakes if left on its own; would supplement, using AI agents for particular functions
  • Might be AI within CDP, talking to AI in marketing automation, etc.
  • Data changes quickly, so pulling out fragment as context may result in AI using out-of-date information
  • Could be too costly to load all raw data into AI for training, without putting into CDP/warehouse
  • Would also need CDP for non-AI systems that need clean data
  • Need to assess the value of investments; may be things that are technically possible but not financially worthwhile
  • What sets CDP apart is that it interprets and transforms data, making it usable for other platforms
• May not need CDP if don’t have real-time use cases

Europe/U.S. Roundtable, May 22, 2025

Guest: Lee Hammond, Principal, KLH Consulting

Topic: Practical Experiences Building Composable CDPs

• Important to work across teams
• Lee’s experience at Universal Music Group: had modern data stack and engineers, but the data team didn’t want to build connectors to CRM, advertising, and other systems; pushed marketing to buy CDP; implemented CDP that met requirements but didn’t support nuances such as local privacy rules; later discovered that Reverse ETL might have met his needs but sounded to marketers like an engineering tool, so recommended ‘composable CDP’ to Hightouch as something that would attract marketers
• Defines ‘composable CDP’ as separate of data infrastructure from marketing application (audience building, customer journey orchestration, synchronization to downstream marketing, sales, and operational systems)
• Have seen traditional CDP vendors adding connectors to access warehouse data via sharing without copying
• Some traction in the industry to limiting data warehouse to data collection and doing computation elsewhere (vs. in the warehouse, as in original ‘composable CDP’ descriptions)
• Best architecture depends on your company’s existing data stack: e.g., company that has Braze might want to do more in warehouse because Braze has limited data capabilities, while company with MessageGears could do more almost everything in MessageGears because it can; also depends on skills and capabilities of different teams within the organization
• During 2022-2025 have seen rise of data warehouse as primary data source; David Chan of Deloitte defined ‘dual zone’ model:
  • ‘zone one’ data ecosystem (data assembly, identity resolution, cloud data warehouse, enrichment and insights) and
  • ‘zone two’ marketing ecosystem (CDP, journey orchestration, personalization & experimentation)
• Dual zone model maps to organizational structure (zone 1=data team vs zone2=marketing team); reflects Conways Law (quoted by Scott Brinker) that software design reflects organizational structure
• Thinks composable CDP matches structure of most organizations: CDP is really a big data application which requires a data engineer, who must still be responsible to marketing, whether they work in the marketing department or the data department, but most organizations don’t have that
• CDP gives marketers a tool to fill the gap between what data team can deliver, and what marketing needs
• Data warehouses already give IT to deal with data residency requirements, whereas SaaS products (including CDPs) often need a separate instance for each country or region to keep the data separate
• Consent management platforms such as OneTrust are critical inputs to data warehouse (zone one)
• Large companies (Snowflake clients that Lee worked with at Hakkoda)
  • In zone one: have complex environments (multi-brand, multi-region, multiple revenue channels); give great deal of autonomy to individual operations units, which don’t want to give up their own systems; usually have large, successful, ambitious data teams, which prefer build over buy but follow rigid processes, and concern themselves only with IT/compute costs, not marketing system costs
  • In zone two: disappointment with martech results causes reluctance to spend more; often are accessing warehouse data informally, but have had bad experience with IT building projects in the past.  Lee argues that composable CDP means past IT issues are less relevant because is more assembling than building.  Most CDP projects originally by owned media teams, who wanted automated data flows and journey orchestration (not by advertising teams).  Marketing teams struggle with measurement, limited experimentation, limited content library.  Companies get stuck on post-delivery issues, not deployment.
• Enterprise deployment of AI is moving more slowly than AI capabilities are evolving
• Marketers see ‘center’ of their stack as whatever team they work with most often; primarily CRM, marketing automation, customer engagement platform; had been some movement to making that CDP but seems to have stalled or reversed
• Two ways to meet CDP requirements (in addition to conventional stand-alone CDP): composable CDP inside data warehouse, or embedded CDP inside marketing/sales applications (CRM, MAP, etc.)
• Data ecosystem components: data warehouses are best at collecting many types of data; data science systems make choices such as offer selection; CDPs are widely used to manage customer journey; have separate Content Management Systems; separate paid marketing systems
• Different teams in marketing often have their own tools
• Once have delivered CDP, often get questions about reducing duplication within the stack
• Scott Brinker categorization: systems of truth (warehouse, digital asset mgt, product information manager) vs systems of context; could be integration of warehouse, digital asset mgt, composable CDP, although it hasn’t happened yet)
• Warehouse will continue to support AI by providing aggregated data; 60% of companies already integrate AI martech (Zone Two) with warehouse (in Chiefmartec & Martech Tribe survey)
• Also expect AI to be active in Zone One data management tools (ingest, transform, identity resolution)
• Had previously been limit that ‘composable CDP’ can’t do real time but that is largely resolved by ability to connect warehouse with data streams such as Kafka