DXP or Best-of-Breed? AI changes the debate
)
AI shifts the Best-of-Breed vs. DXP debate
Why integrated DXPs are gaining ground in the middle market
Until a few years ago, I experienced many conversations about the architectures of Martech stacks like this: Best-of-breed architectures stood for agility and the desire to modernize or adapt the stack step by step. DXPs stood for ready-made integrations, unified user interfaces, and shared data, but also came with stronger vendor lock-in and dependency on the product roadmap and priorities of the manufacturer.
AI changes this consideration significantly from my perspective. Many discussions today are less about feature comparisons of individual components. Increasingly, the central question is: How do I ensure that fragmented AI functions pursue a common goal across all touchpoints and share insights with each other?
My thesis is therefore: Integrated DXPs with a shared AI layer, shared context, and shared knowledge base have an architectural advantage over fragmented best-of-breed stacks.
AI transforms tools into decision systems
In classic (hybrid) stacks, many components primarily function as executors. CMSs deliver content. Marketing automation systems control the sending of emails and journeys. Commerce systems manage product logic, checkout, prices, and offers. CDPs consolidate profiles. Analytics solutions measure. All these components now include AI functions, which extend their role. Systems can now make decisions: prioritize content, generate variants, derive target groups, schedule contacts, create recommendations. This can increasingly occur automatically and without human intervention. On paper, this sounds like increased speed. In practice, a new challenge arises: decisions are distributed across different components, goals often remain implicit. The individual components learn for themselves. Insights stay within the respective tool, are interpreted differently, or do not systematically integrate into other decision logics. This shifts the architecture question. Best-of-Breed has long been a good model because components with the best features could be integrated. With AI, it's more about whether there is a common goal system and learning logic.
When every component has its own AI
Hybrid stacks today almost inevitably come with multiple AI functions. This can be useful, for example, for content drafts in the CMS, recommendations in commerce, optimizations in marketing automation, or identifying user groups with similar behavior patterns. The challenge lies in the interplay. Part-AIs work with the context they see. This context is often complete from the tool's perspective, rarely from the company's perspective, leading to local optima. In digital marketing, this becomes evident quickly, even without technical depth: one system measures success through clicks, another through conversion, a third through openings. One system knows cross-channel contact history, another only within its channel. One system uses profile information under certain consent conditions; another interprets it differently or doesn't have the information at all. As a result, performance suffers. The output across various channels is often less consistent, less aligned with the target audience or goals, or varies between channels. This directly affects conversion, pipeline quality, and efficiency. Optimization quality suffers because each system amplifies its signals, and no overarching goal system resolves conflicts. Instead of a company-wide learning curve, multiple parallel learning loops emerge. Insights remain local, aren't consolidated, and contribute only limitedly to a joint optimization strategy.
What I mean by "shared AI layer"
That's why it's architecturally much more helpful to have a central AI layer in the martech stack that has access to all relevant data and a common goal system. From my perspective, this AI layer should include at least the following three aspects:
A unified identity and consent model. This provides a central view of identity, account and buying groups, interaction history, funnel status, preferences, and contact rules. In B2B marketing, this is the difference between point personalization and control that considers account dynamics. This includes consent and purpose as part of this context. Without these guidelines, personalization becomes either unused or legally risky.
A structured content and asset base, which includes content, product data, taxonomies, metadata, terminology, brand voice, approved statements, visual language, and assets. This is crucial for generative AI. Without this base, variations that may seem formally plausible but don't fit content-wise or ignore brand guidelines can be created. Then review processes end up rejecting them repeatedly, and ultimately, time-to-market doesn't really improve.
A common decision and optimization logic that bundles goals, priorities, rules, and measurement logic across components. This includes conflict resolution: What takes precedence when multiple systems want to execute at the same time? This includes frequency control: How often do I communicate through which channel at which stage? It includes a definition of impact that can do more than just "increase clicks."
Why this architectural point is so relevant for mid-sized companies
If you think about it architecturally, the ideal state would surely be a best-of-breed stack with its own, central AI layer. In such a model, CMSs, commerce, automation, and CDP systems remain specialized execution systems. The intelligence that defines goals, consolidates insights, and resolves conflicts lies in a central platform. Insights from one channel are structured into others. Optimization doesn't occur locally but along a common goal system. For very large companies, this model is feasible. They have platform teams, data governance, and financial and personnel capacities. Sometimes, they also have an experimental mindset that allows building their own AI and decision-making layer. For many medium-sized companies, however, this path is barely feasible. Building and operating such an AI layer requires continuous investments in skills, governance, infrastructure, and know-how. The complexity is enormous, and often not all necessary experts are available. This is exactly where integrated DXPs have a structural advantage, bringing a shared AI and context layer as part of the platform architecture. Providers like Adobe, Salesforce, Sitecore, and Optimizely are pursuing exactly this approach and target larger companies. Ibexa follows a very European approach, focusing on digital sovereignty and trust, targeting the upper mid-market.
Two limiting aspects we should talk about
These two points apply regardless of architecture. They affect integrated DXPs as well as best-of-breed stacks. Without them, neither a central AI layer nor a distributed AI landscape will function sustainably.
Data quality and measurement logic remain the entry ticket. A common goal system only works with good data quality, clean event logic, and clear KPIs and metrics. If "lead," "MQL," and "opportunity" are interpreted differently across systems, if datasets are not complete and shared between systems, AI amplifies this ambiguity. But exactly this is what we see in many customer projects: incomplete customer data sets and, above all, no feedback on the quality of generated leads from sales. The manual evaluation and feedback of results is a crucial prerequisite for effective optimization. This cannot be provided by an AI system alone.
Consent, purpose, and governance determine acceptance in customer communication. Especially in Europe and the DACH region, a clean handling of consent signals, purpose limitation, permissions, and documentation is an imperative requirement. The best personalization options are worthless if customers develop the feeling that data is being collected improperly. Trust in the proper and careful handling of data must guide actions and must be technically and organizationally ensured. And here, too, we see a great need for improvement in practice. Many companies are currently unable to properly document the consent and source of their customer and lead data. What we often see: Some data is so incomplete due to numerous tool migrations in the past that it is no longer clearly documented when it was first collected.
Conclusion
AI provides great new opportunities in Martech tools. However, AI also changes the logic of control in Martech stacks. Once multiple systems optimize independently, it is no longer the best feature that decides the performance but rather the common target system. Fragmentation is the enemy of performance. Those who use fragmented sub-AIs without central coordination risk inconsistent decisions and suboptimal outcomes. A best-of-breed approach with its own central AI and decision layer can deliver this target system. For large companies with the necessary resources, competencies, and governance, this is a feasible path. For many medium-sized organizations, however, this self-build is hardly achievable organizationally and financially. For them, an integrated DXP architecture is the more pragmatic way to strategically and controllably use AI. Because these anchor context, knowledge base, optimization logic, and executing components in a common architecture, thereby laying the foundation for consistent, controllable performance.
Share this post