When two teams report the same metric with different numbers, your AI doesn't split the difference — it picks a side. Here's how to fix semantic debt before it scales.
Two teams. Same metric. Different numbers. That scenario — the one every data lead has lived through at least once — is about to get dramatically worse. Not because your data quality is deteriorating, but because you’re about to hand that disagreement to an AI agent and ask it to act on it.
The Problem Isn’t Your Data. It’s Your Definitions.
Semantic debt — the accumulated gap between what a metric is called and what it actually measures — has always existed in enterprise marketing stacks. “Active users” means something different to your CRM team than it does to your product analytics team. “Conversion” in your Shopee campaign dashboard doesn’t map cleanly to “conversion” in your attribution model. These discrepancies used to surface in quarterly reviews and get resolved over a heated slide deck.
As dbt’s Dustin Dorsey argues, AI makes this impossible to ignore — and not in a good way. When a human analyst hits a conflicting number, they pause and investigate. When an AI agent hits a conflicting number, it doesn’t pause. It picks a definition, runs with it, and generates recommendations at scale based on that choice. By the time the error surfaces downstream, your personalisation engine has already sent 200,000 messages based on a flawed cohort definition.
For Southeast Asian marketing teams running cross-platform campaigns across LINE, Shopee, Grab, and owned channels simultaneously, the surface area for definitional conflict is enormous. Each platform exports metrics with its own naming conventions, attribution windows, and session logic.
AI Agents Don’t Fail Loudly — They Fail Silently
Monte Carlo’s release of Agent Lineage this week is a signal worth reading carefully. The data observability platform is now offering lineage tracking specifically for AI agent runs — because, as their team notes, when agents misbehave or produce degraded outputs, teams typically have no idea whether the fault lies in the model, the retrieval context, the upstream data, or the tool calls the agent made along the way.
This is the black box problem applied to marketing operations. A customer engagement platform running real-time decisioning might touch twelve data sources in a single journey trigger — session data, purchase history, segment membership, suppression lists, channel preference, recency scores. If any one of those upstream definitions is inconsistent, the agent’s decision logic is compromised. And unlike a batch campaign where you can QA before send, real-time agents act before you can check.
The implication for teams building CEP frameworks right now: observability isn’t a post-deployment concern. Lineage tracking, metric documentation, and definition governance need to be baked into your data architecture before you wire agents into your activation layer — not retrofitted after your first incident.
Structure Is the Prerequisite for Intelligence
There’s a quieter but instructive parallel in the world of enterprise document intelligence. A piece published this week on Towards Data Science tackled a surprisingly common problem in RAG systems: PDFs that visually display a table of contents but ship no structural outline in their metadata. The fix requires reconstructing section boundaries from the document itself — essentially reverse-engineering the structure that should have been there from the start.
It’s a small technical problem. But it’s a precise metaphor for what’s happening in marketing data stacks. The structure that makes information retrievable, scopeable, and trustworthy has to be built deliberately. It doesn’t emerge automatically from the data itself. A well-labelled column in a database is the equivalent of a properly tagged PDF section — it tells downstream systems not just what the data is, but where it belongs and how it relates to everything else.
For teams investing in AI-driven personalisation across multilingual, multi-platform Southeast Asian markets, this is more than an analogy. Localised content variations, regional pricing logic, platform-specific creative specs, and market-level suppression rules all need to be structurally encoded in your data layer — not left as implicit knowledge sitting in someone’s head or a shared Google Doc.
What Fixing Semantic Debt Actually Looks Like
The path forward isn’t glamorous, but it’s concrete. Start with a metric alignment audit: pull your ten most-used marketing KPIs and document exactly how each team, platform, and system defines them. You will find disagreements. That’s the point.
From there, a semantic layer — whether via dbt metrics, a headless BI tool, or a centralised metric store — gives you a single definition that all downstream consumers, including AI agents, reference. This is not a data engineering luxury; it’s a prerequisite for any AI activation work that needs to scale reliably.
Pair that with agent lineage tooling. If you’re running AI decisioning in production — even in a limited pilot — you need visibility into which data sources and definitions each agent decision touched. When a journey trigger fires unexpectedly or a personalisation recommendation looks wrong, you need to trace the reasoning, not just the output.
Finally, treat definition governance as a recurring process, not a one-time cleanup. Southeast Asian platforms update their data exports and attribution models with frustrating regularity. What your Lazada seller dashboard called “click-through” six months ago may not be what it calls it today.
The brands that will get real value from AI-driven engagement aren’t the ones who deployed agents fastest. They’re the ones who did the unglamorous work of making sure the agents had something trustworthy to reason with.
Is your organisation’s AI activation roadmap built on a semantic layer that’s actually maintained — or on a quiet assumption that everyone means the same thing when they say the same word?
At grzzly, we work with marketing and data teams across Southeast Asia to design CEP frameworks where the activation layer and the data architecture are built in parallel — because one without the other is just expensive guesswork. If your team is navigating the gap between AI ambition and data reality, Let’s talk
Sources
Written by
Brooding GrizzlyDesigning CEP frameworks that move beyond batch-and-blast into real-time, context-aware engagement — across channels, devices, and the messiness of actual human behaviour.