From Dashboards to Decisions: The Data Stack Is Finally Growing Up

Three things happened this week that, taken separately, look like product news. Taken together, they describe a structural shift in what a data stack is actually for.

The Fivetran-dbt merger closed. Netcore Unbxd shipped a conversational AI layer on top of ecommerce search analytics. And buried in the technical press, a paper on Proxy-Pointer RAG quietly proposed a smarter way to extract meaning from enterprise knowledge graphs. None of these are coincidences. They’re symptoms of the same underlying pressure: the gap between having data and acting on data has become the most expensive gap in marketing.

The Pipeline Is No Longer the Point

For years, the dominant narrative in data infrastructure was consolidation — get everything into one warehouse, clean it up, model it consistently. Fivetran moved data in. dbt transformed it. They were complementary tools that lived in the same workflow but different companies. As Tristan Handy wrote in the official merger announcement, the combined entity now exists to deliver “data infrastructure for agents you trust” — a phrase that would have meant nothing three years ago and now means everything.

The signal here isn’t operational efficiency. It’s architectural intent. When your ingestion layer and your transformation layer share a product roadmap, the next logical step is that the activation layer — the CEP, the personalisation engine, the recommendation system — gets built closer to the data itself, not bolted on downstream. For brands running complex customer journeys across Shopee, Line OA, and their own apps simultaneously, that proximity matters enormously. Latency kills relevance.

Conversational Analytics Is a UX Problem Disguised as a Tool Problem

Netcore Unbxd’s new Insights Agent is being positioned as a replacement for complex dashboards — a conversational layer that lets merchandising teams ask questions in plain language and get answers fast enough to act on. The framing is right, even if the execution will vary.

The real problem it’s solving isn’t technical. It’s organisational. Most ecommerce teams in Southeast Asia are running lean: one or two analysts supporting a merchandising team of ten, with reporting cycles that lag commercial decisions by days. By the time a category manager knows that search-to-purchase conversion on “wireless earphones” dropped 18% last Tuesday, the promotional window has closed. A conversational query layer that surfaces that signal in real time — and connects it to inventory, pricing, and campaign levers — changes the economics of merchandising decisions fundamentally.

The implementation risk is the same as any AI interface: garbage framing, garbage output. If the underlying data model is messy or the entity definitions are inconsistent, natural language queries will return confident-sounding nonsense. That’s not a tool failure; it’s a data governance failure that the tool just made more visible.

---

Knowledge Graphs and the Cost of Extraction

The Proxy-Pointer RAG paper from Towards Data Science addresses something less glamorous but equally important: how enterprise AI systems waste compute — and introduce error — by repeatedly extracting entities and relationships from documents they’ve already processed. The proposed approach uses structural pointers to reuse prior extraction work rather than redoing it from scratch on every query cycle.

For a customer engagement context, this matters in a specific way. If your personalisation engine is querying a knowledge graph to understand product relationships, customer segment affinities, or campaign context, the efficiency of that extraction loop directly affects how quickly you can resolve a decision — and how much it costs to do so at scale. A brand running 50 million monthly active users across mobile in Indonesia or Vietnam isn’t doing one query. They’re doing millions. The compounding cost of wasteful extraction is real, and architectures that ignore it will hit ceiling economics faster than they expect.

This isn’t academic. It’s a concrete consideration for any team evaluating GraphRAG systems as part of their next-generation CEP or recommendation layer.

What Activation-Ready Architecture Actually Looks Like

Pull these threads together and a picture emerges: the data stack is being re-architected around the moment of decision, not the moment of reporting. That requires three things most brands haven’t fully resolved.

First, a unified data model that doesn’t require translation between systems — where what dbt knows about a customer segment is the same thing the CEP platform acts on, without a synchronisation lag. Second, an inference layer that’s fast and cheap enough to operate in real time — which is where RAG architecture efficiency stops being a nerd concern and starts being a P&L concern. Third, interfaces that put signal in front of the person who needs to act on it, not just the person who can interpret it.

Southeast Asian brands have a structural advantage here that often goes unrecognised: the platform ecosystems they operate in — Grab, LINE, Shopee, Lazada — are already generating rich behavioural signal at enormous volume. The problem isn’t data scarcity. It’s decision latency. The teams that close that gap first — architecturally, not just tooling — are the ones who’ll compound their engagement advantage into revenue outcomes that are genuinely hard to replicate.

The question worth sitting with: is your current data architecture designed to answer questions you’ve already thought of, or to surface ones you haven’t yet asked?

---

At grzzly, this is exactly the territory we work in with growth teams across Southeast Asia — designing CEP and data frameworks that connect infrastructure decisions to activation outcomes, not just reporting outputs. If your stack is producing insight but not action, that’s a solvable problem. Let’s talk