Alert fatigue, bloated token costs, and public AI distrust share one root cause. Here's how smart data modeling fixes all three.
Most AI activation problems aren’t model problems. They’re plumbing problems dressed up in a lab coat.
Three separate developments this week — a study on public AI skepticism, a triage agent built to manage alert overload, and a data modeling story that cut token costs by 20x — point to the same underlying dysfunction. Teams are trying to activate on AI before they’ve cleaned up what they’re feeding it. The result is noise at scale: systems that fire too often, cost too much, and earn too little trust.
Alert Fatigue Is a Data Architecture Problem in Disguise
Monte Carlo’s Virna Sekuj describes a situation every data team will recognise: one channel — usually Slack — receives every system alert indiscriminately. A freshness break on a critical revenue table sits alongside a volume dip on a staging table nobody queries. Both ping at 2am. Both get ignored.
The solution Monte Carlo built — a triage agent that classifies alert severity before it reaches a human — is clever. But the more important insight is what this problem reveals about data infrastructure at large. When systems can’t distinguish signal from noise, humans stop trusting them entirely. That same dynamic plays out in customer engagement: if your CEP fires a re-engagement message every time a user misses one session, the user’s inbox becomes that Slack channel. They stop opening. You’ve trained them to ignore you.
The fix isn’t a smarter alert. It’s tiered monitoring logic baked into the data model — defining upfront which signals actually warrant action, and at what threshold. In a Southeast Asian mobile context, where users receive an average of 46 push notifications per day across super-apps, the cost of undifferentiated alerting is brand trust, not just operational noise.
Token Cost Is a Modeling Problem, Not a Vendor Problem
Getdbt’s Britton Stamper published a remarkably practical post about Gong call transcript data. The team was passing raw, unstructured transcripts directly to AI models — burning through API tokens at a rate that made the use case economically unviable. By modeling the transcripts in the data warehouse first with dbt, structuring them into clean, queryable entities rather than walls of text, they reduced token consumption by a factor of 20.
That’s not a marginal efficiency gain. That’s the difference between a use case existing and not existing.
The same principle applies directly to real-time personalisation pipelines. Teams building AI-driven customer journeys often make the same mistake: they pass raw event streams or unprocessed CRM exports into a model and wonder why results are inconsistent and costs are ballooning. The model is doing work that the data warehouse should have done upstream. When you model for the AI consumer — not for the analyst — you dramatically reduce the surface area of what the model has to interpret, and dramatically improve what it returns.
For teams running on Shopee or Lazada seller data, or ingesting LINE OA interaction logs, this is especially relevant. Platform-native data is messy, inconsistent in schema across markets, and often duplicated. Cleaning it at the warehouse layer before it touches any AI pipeline isn’t optional housekeeping — it’s the prerequisite for any reliable activation.
Public AI Skepticism Is a Trust Calibration Signal
Stephanie Kirmer’s analysis in Towards Data Science surfaces something the martech industry tends to brush past: people’s willingness to accept AI-driven experiences is conditional on perceived value. When they see a clear benefit — faster resolution, a more relevant recommendation, a better price — they tolerate the tradeoff. When they don’t, skepticism hardens fast and is difficult to reverse.
This is directly relevant to how brands in Southeast Asia deploy AI in customer journeys. The region has high smartphone penetration and strong platform loyalty, but also acute sensitivity to feeling manipulated or surveilled. A personalisation engine that surfaces a relevant recommendation during a browse session feels helpful. The same engine sending a WhatsApp message referencing a product someone viewed once, two weeks ago, at 8pm, feels like surveillance.
The distinction isn’t the data. It’s the context model. AI-driven engagement that earns trust is built on frameworks that ask: what signal, at what moment, through which channel, justifies this intervention? That requires investment in journey architecture before a single model is deployed — defining the value exchange explicitly, not discovering its absence through churn.
Kirmer’s underlying point is a useful forcing function for any growth team: if users can’t articulate the benefit of an AI-powered experience you’ve built, you’ve probably optimised for your metrics, not their moment.
The Common Thread: Upstream Decisions Determine Downstream Quality
What connects alert triage logic, token-efficient data modeling, and public AI trust? All three fail or succeed based on decisions made before the AI does anything. The model is the last mile. The architecture is the race.
For customer engagement teams, this means the real leverage isn’t in selecting a better AI vendor or adding another personalisation layer. It’s in:
- Defining signal hierarchy early — which user behaviours actually predict intent, and which are just activity
- Modeling data for the AI consumer — structuring inputs at the warehouse layer so models receive clean, contextualised entities rather than raw event dumps
- Building the value exchange into the journey design — so the AI’s output is one the user can feel, not just one the dashboard can count
The teams that get this right aren’t necessarily the ones with the most data. They’re the ones who’ve been ruthless about what the data is actually for.
The question worth sitting with: if your AI activation stack disappeared tomorrow, would your customers notice something better — or just something quieter?
At grzzly, we work with growth and data teams across Southeast Asia on exactly this kind of structural problem — helping brands build CEP frameworks where the architecture and the activation strategy are designed together, not bolted together after the fact. If your AI outputs feel noisy or expensive relative to what they’re returning, the answer is usually upstream. 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.