Europe/US Roundtable: March 26, 2025

Guest: Vince Jeffs, Senior Director Product Strategy, Pegasystems

Topic: Buying vs Building Your Own Solutions with Open Source and Other Tools

• Trend among big enterprises to build things themselves
• Ultimately, it’s all about use cases.  Building a data warehouse with no specific reason in mind didn’t work because users didn’t find a reason to use them.
• CDPs became important because marketers wanted a say in how their data was organized and used, to do specific things such as acquisition, retention, cross sell, improved customer service, etc.  So, always about use cases. 
• CDP/martech use cases span entire customer lifecycle from awareness and acquisition, to increasing spend, upsell & cross sell, to providing service, retaining customers, collecting debts, and winning back past customers.  Always comes down to increasing revenue or decreasing costs.
• CDPs must integrate with activation in general, real time interaction management in particular.  Some CDP vendors do all of these, others integrate with third party products.
• Core strengths of CDP including data collection, identity management, insights, real-time data feeds with integrations to email systems, marketing clouds, journey orchestration engines, etc.
• Range of CDP options: packaged CDP, hybrid/composable, build from scratch. 
  • Huge enterprises with lots of experience lean more to composable or building, while smaller organizations get more from buying packaged CDP.  Some verticals also lean more toward building, especially in highly regulated/governed industries such as telecom, banking, insurance, healthcare, which have more experience with data management.  Large companies with less B2C marketing experience lean more towards packaged apps.
  • Composable CDP vs packaged CDP: composable requires more configuration, usually combines importing some data into SaaS application and reading other directly e.g. from BigQuery; generally, requires more technical work than end user easily can do.
  • CDP functions are different from DMP functions for advertising management; DMP is adjacent world to CDP; must consider use cases: CRM vs paid media are different purposes, could try to find platform that does both but must decide which is more important. 
  • If have invested a lot in data warehouse, more likely to build than replicate work in packaged CDP
  • For use cases to help drive decisions: must be detailed, not just ‘improve cross sell in website’.  Must be specific about what’s needed for use case, and have one or two year plan for how will use solution.  Need to define what kind of offers will serve and where, goals for use cases, and data that will drive use cases.  May not need complete list of data but need to understand most of it so don’t find yourself waiting for one or two years to assemble it.  Data doesn’t need to be perfect but make sure it’s adequate; don’t want to jump ahead and then find you are sending based on bad data.
  • Data is how much you know your customer: PII, actions, purchases, etc. comes from owned channels, less from acquisition programs or acquired data.  Due to privacy limits, advertisers are becoming more dependent on first party (owned) data.  
  • Raw data isn’t enough: good CDP vendor will help company confines, aggregates, and organize data to highlight signals, key behaviors, etc..  Those features should be part of data model provided by CDP vendor.
  • Use case should identify data sources, data elements, data model, processing needed, etc., so can then compare to current state and effort needed to close gaps with different solutions.
  • Use cases also need to include ‘so what’, which shows the business value of use case.
  • Need to look at use case specifically for your industry,
  • Need highly rationalized set of data for each customer, not just scattered among many places.  Needed to make understandable by users and to make easily and quickly available to systems, typically in a denormalized row (single record). 
  • Use case needs to assess performance requirements e.g. response time, performance at scale, cost of achieving necessary performance, impact on related systems such as web page
• Build requirements:
  • Consider technical effort to design, construct, test, and roll out, for all tasks from data & event assembly to analytics, decision management, action and treatment selection, channel integration, and strategy optimization.  Need IT, subcontracts, and tools. 
  • Low code/no code tools help but still takes technical skills.  Lots of “free-ish”, open source tools; but still must connect all these even if they’re free.  Will also need paid tools for ETL, data integration, data management services.
  • Many of these tools will already be in place at large organizations, so makes IT more confident they can do the work.  Doesn’t mean that IT knows exactly how to use them for CDP tasks, but does explain why they lean towards building things rather than buying a packaged CDP.  This impacts how you estimate costs in time and money.
• Build challenges:
  • Data integration complexity, real time processing and analytics, developing sophisticated algorithms, multi-channel orchestration, talent acquisition and retention, cost and time to value, need to stay ahead of curve over time
  • Vendors provide many of these skills and resources; must provide for yourself if building on your own
• Build misconceptions:
  • Overlooking hidden costs, underestimating complexity and scope by looking only at first use cases/minimum viable product, unrealistic time expectations, neglecting importance of data quality, underestimating value of external expertise
  • Challenges continue beyond initial deployment.  One result is that people buy new solutions for each problem, so end up with ‘frankenstack’ of different point solutions. 
• Resource gaps:
  • Business users need access to tools that let them do things for themselves, without relying on IT for support.  This includes support groups within business unit such as marketing operations people, which also need tooling. 
  • IT is shared resource for all business units.  Often underestimate need for on-going support and maintenance.
  • Marketing needs people who understand data, is another resource gap.  More than just learning an application.  Technical resource must be embedded in business.  Is separate from training end-users.
  • Evaluation criteria for martech vendors should include support they can provide with data management, since can’t expect users to become experts in a few weeks.  Need to consider both on-going support from vendor, as well as support during implementation.  Depth of vendor experience in your industry is very important and will be highly apparent.
• AI: distinguish predictive vs adaptive modeling. 
  • Predictive requires experts to monitor and adjust models over time;
  • adaptive models maintain themselves so take less support
  • Key aspects of AI that impact build vs buy decision: how much AI is off-the-shelf (determines how much skilled staff support is needed). 
  • With genAI, must examine closely how much manual labor is involved with checking for quality, regulatory compliance, brand standards; vendors often overstate how automated the flow will be.  Most people in big organizations recognize there is promise but genAI isn’t fully mature for complex use cases.  Are seeing applied to low risk, fairly simple use cases e.g. summarizing call records.
  • Distinguish genAI from other AI; don’t assume genAI can do everything; genAI isn’t good at forecasting, predicting, statistical work.
  • AI for data management: newer LLM models are good at classification, so can be good at data tagging, identifying PII, sentiment analysis, intent detection.  If that is well integrated to vendor system, is good for CDP.

APAC Roundtable: March 28, 2025

Guest: Subra Krishnan, Co-Founder and Chief Executive Officer, Lemnisk

Topic: Enterprise Challenges in Adopting AI on Top of CDP 

• Deployment: haven’t seen any enterprise that deployed genAI at scale; partly because people don’t yet understand how to deploy the models, how to overcome other obstacles such as brand regulations
  • smaller start-up may only need to worry about efficiency, so is more able to deploy
• Challenge: Ability to do anything with AI starts with data.  Issues include:
  • Fragmented data sets, across business units, channels, geography, products
  • Data quality and integration complexity; e.g. agent’s email used as customer ID
  • Legacy systems, including old technologies, home-built systems, difficult to extract data
  • Restrictions on access to most important data, e.g. transactions; often are only allowed to access on-premise or in private cloud
  • For AI in particular, data needs consistency & greater precision (so AI can easily interpret);
    • may need new data types, depending on use cases.  Main use case groups are: find right buyers for a product, next best action or product, content creation, right channel/time combination; often transaction data is very particular
• Challenge: Deployment at scale:
  • real-time processing may create bottleneck during peak periods (which are the most valuable)
  • cloud could ease bottlenecks, but often regulated industries require on-premise processing
    • may ease this by tokenizing (anonymizing) data, so can run on the cloud and then brought back on-premise for execution, using combination of on-premise warehouse and cloud-based CDP
• Challenge: compliance and governance
  • Must comply with brand guidelines, which is hard to ensure with AI models; companies often require manual intervention to ensure compliance before releasing results; these problems will go away soon
  • Data privacy, ethical use, regulatory requirements and workflows will not go away
• Challenge: organizational readiness
  • Lack of AI expertise (different skills for genAI vs other methods)
    • Deployment success often depends on organizational experience;
      • need basic skills in sourcing, cleaning, importing data, need ability to link AI to value creation;
      • but are different skills to build model with LLM than traditional methods
  • Organizational change management, e.g. drive change throughout organization during deployment, e.g. in customer service where are concerns about job loss
  • vendor lock-ins: may be hard to get data in and out of existing systems, especially in real time; may have systems that worked well in the past but are not adequate for AI
    • Seeing some solutions, e.g. consent management frameworks to deal with regulatory issues, but these often can’t overcome vendor limits
    • Stand-alone CDPs can help, since enable switching to other vendors using CDP data
    • Having some technologies within CDP can also help, e.g. consent management either built in or integrating with external products
    • But different organizational groups often have preferred products (e.g. consent manager) and don’t want to adopt CDP solution instead, especially if CDP solution is less advanced
    • Big, global enterprises in particular have complex requirements, e.g. consent management accommodating different rules for different countries, these are best found in specialist systems
    • suites often can provide adequate features across many functions, but many enterprises don’t want to commit to a single suite; in multi-vendor situation, CDP can be helpful in gluing the systems together
• Recommendations
  • Start with culture and organizational alignment: be clear about where jobs will be lost, training is needed, what use cases
  • Pick a flexible and open CDP, especially if not committed to a single suite
  • Centralize all model executions on top of a CDP, and then deploy AI models on top of that (including external models)
  • Tokenize for scale: bring data into on-premise CDP, convert customer ID to tokens and push to cloud, run your operations in the cloud, then bring it back and re-identify
  • Use cases can be very narrow; should look at larger categories such as CX or marketing automation, which combine many use cases
  • Often, people want AI to fit into existing workflows without changing them, so use cases are not a critical consideration because people are looking to drive efficiency into existing processes, not create new processes (which is where use cases are most helpful)
• What should CDP vendors do, given the growing role of Snowflake and other big cloud providers?
  • Those firms are a new set of competitors for traditional CDPs, making it harder for them to grow.  But they appeal to customers with different needs than independent CDPs, so are not necessarily going to cost them a lot of business they would have won otherwise