CDPs are drowning in LLM hype. Here's how in-stream decisioning — not generation — unlocks the customer data you already have.
Your CDP is probably sitting on enough behavioural signal to prevent your next three churn events. The problem isn’t the data — it’s that the AI layer on top of it is built to generate text, not make decisions.
The Generation Trap Inside Your Customer Data Platform
There’s a structural bias baked into most AI-augmented CDPs right now. Because large language model technology is dominant and accessible, vendors have bolted decoder-based architectures onto data pipelines that were never designed for open-ended generation. The result: your platform produces outputs — a personalised email subject line, a next-best-offer recommendation — when what the moment actually requires is a decision.
Towards Data Science contributor Lambert Leong puts this cleanly: generation is not always a decision. When a customer opens a support ticket, revisits your pricing page, and then searches your cancellation flow inside 48 hours, that sequence isn’t a content brief. It’s a triage signal. Feeding it to a decoder to produce a retention email is the wrong tool for the job — and in Southeast Asian markets where LINE, WhatsApp Business, and in-app messaging require near-instant response windows, the latency cost of generating prose compounds the problem.
The architectural fix isn’t wholesale replacement. It’s knowing where in your data pipeline generation belongs and where classification or rule-augmented decisioning should take over.
In-Stream Intelligence: Acting Before the Signal Goes Cold
Tealium’s newly announced Jev model inside Tealium Functions is a useful reference point for what in-stream intelligence looks like in practice. Rather than batching events for downstream analysis, Jev sits inside the data stream itself — evaluating the combination of signals arriving from a session in real time and routing toward a decisioning output rather than a generated response.
The scenario Tealium describes is familiar to anyone who’s managed a SaaS or e-commerce book of business in this region: support ticket opened, pricing page revisited, cancellation intent detected. Each event alone is noise. Together, they’re a churn probability spike that demands a specific action — not a personalised email, but a priority queue flag for a human retention agent, or an immediate in-app message with a concrete resolution path.
For brands running on Shopee Mall or Lazada storefronts alongside their own D2C properties, this kind of cross-surface signal stitching is particularly valuable. A customer who abandons a cart on your D2C site and then browses a competitor’s equivalent product on Shopee within the same session window is telling you something that a batch-processed audience segment will surface 12 hours too late.
Implementation note: Jev-style in-stream processing requires your event schema to be clean and consistently tagged at the point of collection — not retrospectively normalised. If your web SDK and mobile SDK are firing different property names for the same behavioural event, no decisioning model recovers that gap gracefully.
Debugging the Middle Layer Before You Scale
One underappreciated failure mode when embedding AI decisioning into a CDP: the logic that sits between the raw event stream and the final action is largely invisible. Monte Carlo’s Ruslan Sultanov makes this point in the context of LLM agents — the document retrieval, tool calls, and prompt construction that happen between input and output receive a fraction of the observability investment that goes into monitoring the endpoints.
The same blind spot exists in CDP decisioning pipelines. Teams instrument the data collection layer and the activation layer, but the transformation logic in the middle — the enrichment steps, the identity resolution joins, the audience qualification rules — operates without adequate tracing. When a decisioning model fires the wrong action on a high-value account, diagnosing why requires reconstructing a chain of events that nobody logged in sufficient detail.
For marketing operations teams in Southeast Asia managing multilingual profiles across Thai, Bahasa Indonesia, Vietnamese, and English touchpoints, this problem is acute. Declared data (language preference, location) can conflict with behavioural data (the language the customer actually browses in) in ways that silently corrupt segment membership. LLM trace tooling — logging intermediate states, not just inputs and outputs — gives your data team a recovery path when decisioning goes sideways.
Practically: before enabling any real-time AI decisioning layer on your CDP, instrument your transformation pipeline with event-level logging. Define what a “decision audit” looks like for your team, and build that observability before you scale activation volume.
Earning the Licence Fee: What Good Decisioning Architecture Looks Like
CDPs are expensive. The platforms that earn their contract renewal are the ones where the unified customer profile is genuinely powering faster, more accurate decisions — not just feeding a richer batch of attributes into the same campaign workflows that existed before.
The architecture that justifies the investment combines three things: clean, consistently structured event data at the collection layer; a decisioning model that distinguishes between moments requiring classification (act on this signal now) and moments requiring generation (craft this communication thoughtfully); and observability tooling that makes the middle layer auditable when decisions misfire.
None of this requires replacing your current CDP vendor. It requires being deliberate about where AI sits in your pipeline and what job you’re asking it to do. Generation has a role. It’s just a smaller one than the current vendor marketing cycle suggests.
Key Takeaways
- Deploy in-stream decisioning for high-urgency behavioural signals (churn intent, cart abandonment sequences) rather than routing them through generative AI layers that introduce unnecessary latency.
- Standardise your event schema across all collection surfaces before enabling real-time AI decisioning — inconsistent property naming is the silent killer of cross-surface intelligence.
- Instrument your CDP’s transformation pipeline with trace-level logging before scaling activation volume, so decision failures are diagnosable rather than mysterious.
The brands that will pull away in the next 18 months aren’t the ones with the most sophisticated generative AI on top of their customer data. They’re the ones that figured out which decisions are too time-sensitive for generation — and built the infrastructure to make those calls in the stream. The open question: how many of your current CDP workflows are using a decoder because it was the default, not because it was the right tool?
At grzzly, we work with marketing and data teams across Southeast Asia to audit CDP architectures, align decisioning logic to actual business outcomes, and close the gap between data collected and value activated. If your platform isn’t earning its licence fee, that’s usually a solvable problem — and the solution rarely starts with buying more technology. Let’s talk
Sources
Written by
Velvet GrizzlyArchitecting the unified customer profile — stitching together behavioural, transactional, and declared data into platforms that actually earn their licence fee.