AI agents are rewriting how CDPs ingest, orchestrate, and activate customer data. Here's what Southeast Asian marketing teams need to rethink now.
The CDP you bought two years ago was designed for a world where humans orchestrated data pipelines. That world is gone.
The shift isn’t subtle. Towards Data Science contributor Eivind Kjosbakken recently documented running 100+ Claude-based AI agents in parallel — each handling discrete analytical tasks simultaneously, with outputs feeding back into a central orchestration layer. What took a team of analysts days now resolves in minutes. For CDP practitioners, this isn’t an interesting experiment. It’s a direct challenge to how we’ve architected customer data infrastructure.
The Agent Layer Is the New Middleware
Most CDPs were built around a familiar model: ingest data from sources, resolve identities, build segments, push to destinations. Human analysts sat between each stage, validating logic, catching anomalies, writing SQL. That human-in-the-middle assumption is what’s collapsing.
Agent orchestration frameworks — like the multi-agent approach Kjosbakken details — allow you to assign discrete tasks (anomaly detection, segment refresh, propensity scoring) to parallel agent threads. The orchestrator coordinates outputs without human intervention at each gate. For a regional e-commerce brand managing customer data across Shopee, Lazada, and a direct-to-consumer app simultaneously, this matters: you’re no longer choosing between speed and coverage. You can run identity resolution across all three channels at once, with agents flagging conflicts rather than humans hunting for them.
The practical implication is architectural. If you’re still treating your CDP as a destination — a place where data lands and waits — you’re leaving activation velocity on the table.
What Happens to the Analysts Who Ran the Old Stack
Rashi Desai’s piece in Towards Data Science is more strategically relevant than its career-advice framing suggests. Her argument: analysts who survive the AI transition aren’t the ones who mastered SQL or BI tooling — they’re the ones who understood why a metric mattered to a business decision, not just how to pull it.
For CDP teams, this reframes the talent question entirely. The profiles you need aren’t people who can write complex audience segment logic. You need people who can define what a useful unified customer profile actually looks like for your specific activation use cases — and then govern the agent layer that builds it.
In practice, this means your CDP strategy lead needs to be fluent in both data modelling and commercial outcomes. At a Southeast Asian financial services brand, for instance, that means understanding that a “high-value customer” segment means something structurally different in Indonesia (where cash-on-delivery still accounts for a meaningful share of transactions) versus Singapore (where digital wallet penetration changes the behavioural signal entirely).
Open Infrastructure Is Quietly Changing Your Vendor Calculus
The dbt team’s announcement that OSI — their open-source semantic layer — is now entering incubation at the Apache Software Foundation as Apache Ossie is easy to scroll past. It shouldn’t be.
A vendor-neutral semantic layer graduating to Apache governance means the definitional layer of your data stack — what “revenue” means, what “active user” means, what “churn” means — can now be standardised across tools without being locked to any single platform. For organisations running hybrid CDP environments (a common reality across Southeast Asia, where enterprise brands often straddle a regional CDP instance and local market tools), this reduces the taxonomy drift that silently corrupts unified customer profiles over time.
Concretely: if your Bangkok team and your Jakarta team are calculating “purchase frequency” differently in their respective BI tools, your CDP’s identity graph is building on sand. Apache Ossie-style semantic standardisation is the plumbing fix that makes the customer profile actually trustworthy.
The Activation Gap Is Where CDP ROI Goes to Die
Here’s the part no CDP vendor puts in their pitch deck: most platforms earn less than 40% of their theoretical activation value because the pipeline between unified profile and channel execution is still largely manual. Marketing ops teams pull segments, export CSVs, upload to ad platforms, wait for sync. By the time the audience reaches the channel, the behavioural signal that generated it is already stale.
Agent orchestration closes this gap directly. An agent layer that continuously monitors segment membership — and triggers downstream activation the moment a customer crosses a threshold — turns the CDP from a data warehouse with a nice UI into an actual real-time decisioning asset. For a Grab or a regional telco with millions of daily touchpoints, the difference between a 24-hour activation lag and a 90-second one isn’t a minor efficiency gain. It’s a fundamentally different customer experience.
Implementation note: before you bolt an agent orchestration layer onto your existing CDP, audit your identity resolution logic first. Agents amplify whatever is already in the pipeline — including errors. A poorly resolved customer graph running at agent speed generates bad personalisation at scale, which is worse than slow personalisation.
Key takeaways:
- Redesign your CDP architecture around an agent orchestration layer — not as a future consideration, but as the operational baseline for real-time activation in 2026.
- Prioritise analysts who can translate commercial context into data logic, not just analysts who can execute against existing definitions.
- Standardise your semantic layer across tools before scaling agent-driven activation — data consistency is what makes speed safe.
The deeper question here isn’t whether AI agents will change how CDPs work. They already are. The question is whether your organisation’s data culture — the way decisions get made about what to measure, what to activate, and what to trust — is evolving at the same pace as the tooling. Fast infrastructure on top of slow thinking is just a more expensive way to get the wrong answer.
At grzzly, we work with mid-to-large brands across Southeast Asia on exactly this: auditing CDP architectures, defining unified customer profile strategies, and building the activation pipelines that actually move the commercial needle. If your platform is underperforming its licence fee, that’s usually a solvable problem — and it rarely starts with the 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.