Why AI Context Is the New Customer Data Advantage

Every marketing leader in the room is running some form of AI right now. The honest ones will admit it isn’t working as well as the vendor demo suggested. The reason is almost never the model — it’s the data being fed into it.

Tealium’s Robbie Rizman puts it plainly: the AI advantage won’t belong to those with the most sophisticated models, but to those who can give AI the richest context. That distinction matters enormously if you’re architecting customer data infrastructure in 2026.

Context Is a Data Architecture Problem, Not an AI Problem

The brands seeing real AI lift — measurable improvements in conversion, retention, or lifetime value — share one thing: a unified customer profile that stitches behavioural signals, transactional history, and declared preferences into a single, accessible view. Without that foundation, your AI is essentially guessing in a crowded room.

In Southeast Asia, this challenge is structurally harder than in Western markets. Customers move fluidly across Shopee, Lazada, LINE, and brand-owned channels, often within a single purchase journey. A customer who browses on the Shopee app, clicks a LINE OA message, and converts on a brand’s mobile site generates three sets of signals that rarely talk to each other by default. The brands closing that loop — through proper identity resolution and event streaming — are the ones whose AI-driven recommendations feel eerily accurate rather than obviously algorithmic.

The infrastructure investment here is real. But so is the cost of not making it: Tealium’s research suggests AI systems operating on incomplete or siloed profiles consistently underperform against personalisation benchmarks, while generating the kind of irrelevant outreach that erodes customer trust faster than no personalisation at all.

The Transformation Layer Is Getting an Upgrade — and It Matters

dbt Labs and Fivetran’s announced direction for dbt Core v2.0 signals something important for data teams: the transformation layer — long the unglamorous backbone of any CDP — is being rebuilt with AI-native workflows in mind. This isn’t cosmetic. The ability to define, version, and govern data models with AI assistance changes how quickly teams can iterate on the customer profile itself.

For marketing data teams at mid-to-large brands, the practical implication is this: the gap between raw event data and AI-ready customer context used to be measured in months of engineering time. With the direction dbt and Fivetran are heading, that gap compresses. Semantic layers become more accessible. Non-engineering stakeholders — your CRM manager, your growth lead — get closer to the data without needing a data engineer as an interpreter.

The caveat worth holding: tooling improvements don’t substitute for data strategy. A faster transformation pipeline built on poorly governed source data still produces context that misleads rather than informs. The architectural discipline comes first.

---

What Eight Years of ML Failures Actually Teach Us

Pascal Janetzky’s retrospective on eight-plus years in machine learning lands a lesson that’s uncomfortable but necessary: most ML projects fail not because of model choice or compute resources, but because of misaligned expectations, poor problem framing, and insufficient patience with iteration cycles. The same pattern plays out in CDP implementations.

Brands routinely underestimate how long it takes to get from CDP deployment to a genuinely useful unified profile. The vendor promises six weeks to value. The reality, with identity resolution tuning, data quality remediation, and cross-channel event schema alignment, is closer to six months — and that’s with a disciplined team. The brands that get stuck are usually those that skipped the unglamorous middle: defining what “unified” actually means for their specific customer journey, stress-testing their ID graph against real edge cases, and establishing governance before scale.

Janetzky’s framing of discipline as a prerequisite to insight applies directly here. A CDP that earns its licence fee isn’t a plug-and-play platform — it’s a programme. The distinction changes how you resource it, how you measure it, and how you explain it to a CFO who is eyeing the contract renewal.

Turning the Unified Profile Into Activation That Moves Revenue

Contextualised AI is only valuable at the point of activation. The unified profile needs to flow — in real time or near-real time — into the tools your teams actually use to engage customers: your marketing automation platform, your push notification stack, your on-site personalisation engine, your call centre CRM.

The architectural pattern that’s gaining traction is the reverse ETL approach: warehouse-native customer profiles pushed downstream to activation tools on a continuous basis, rather than batch exports that are stale before they’re used. For brands running high-frequency engagement channels — Grab promotions, Shopee Flash Sales, LINE broadcast campaigns — stale profiles mean your AI is optimising against yesterday’s customer, not today’s.

The measurement framework matters as much as the data flow. Teams that build holdout groups into their activation logic — a discipline borrowed from rigorous ML practice — can actually quantify what the contextualised AI is contributing versus a baseline. That number, when it’s real, is what secures next year’s data infrastructure budget.

---

Key Takeaways

• Unify before you personalise: AI running on fragmented customer profiles doesn’t underperform quietly — it actively damages trust through irrelevant outreach.
• Treat your CDP as a programme, not a platform — budget for the six months of iteration between deployment and genuine insight, not just the licence fee.
• Build reverse ETL and holdout measurement into your activation architecture from day one; retrofitting them later is expensive and politically awkward.

---

The real question heading into 2027 isn’t whether your brand has AI — everyone does. It’s whether your AI knows your customer well enough to earn the next interaction. The brands investing now in the unsexy infrastructure work — identity resolution, semantic data layers, governed transformation pipelines — are building a structural advantage that model improvements alone won’t close. What does your customer profile actually know today that it didn’t know six months ago?

---

At grzzly, we spend a lot of time in exactly this space — helping brands across Southeast Asia figure out what their customer data architecture needs to look like before they invest in the AI layer on top of it. Whether you’re evaluating CDPs, stress-testing your current data stack, or trying to translate infrastructure investment into a CFO-friendly business case, we’ve probably seen the version of your problem before. Let’s talk