Why AI Agents Need Structured Data, Not Raw Prompts

There’s a quiet assumption running through most AI marketing conversations right now: that better models solve messier problems. Feed the chaos in, get clean insight out. It’s a seductive idea — and it’s wrong.

The brands that will actually extract value from AI agents aren’t the ones with the most sophisticated prompts. They’re the ones that built structured, trustworthy first-party data pipelines before the AI hype cycle hit.

The LLM-as-Oracle Problem Is a Data Architecture Problem

Clara Chong’s analysis in Towards Data Science is a useful corrective to the current orthodoxy. Working with 100 unstructured PDFs, she found that treating an LLM as a general-purpose problem solver — dump documents in, expect insights out — produced inconsistent, brittle results. The fix wasn’t a smarter model. It was a deterministic loop: structured extraction first, validation second, synthesis only after the data had a reliable shape.

For marketing teams in Southeast Asia, this maps directly onto a familiar pain point. Customer data across Shopee, LINE, and proprietary CRM systems rarely arrives clean. Behavioural signals from mobile-first journeys are fragmented across sessions, devices, and platforms. If you’re routing that raw material into an AI agent and expecting coherent customer intelligence, you’re not building a data strategy — you’re hoping.

The discipline required is less about AI and more about data design: defining what a trustworthy record looks like before any model touches it.

Agent Experience Is Raising the Stakes for Data Quality

Montecarlo’s Lior Gavish makes a pointed observation: for thirty years, “user” meant a human being. That assumption is dissolving. AI agents — systems that query data, make decisions, and trigger actions without a human in the loop — are now operational infrastructure, not experiments.

This shift has direct consequences for how marketing and data teams should think about consent and data provenance. When a human analyst pulls a customer segment, they apply judgment. When an agent does it, it applies only what the data structure tells it to. A mislabelled consent flag, an ambiguous opt-in timestamp, or a siloed identity graph doesn’t just produce a bad report — it can trigger a non-compliant activation at scale, automatically, before anyone notices.

Brands running loyalty programmes on Grab or executing CRM campaigns through regional telco partnerships need to treat data schema design as a compliance instrument, not just a technical nicety. The agent doesn’t read the privacy policy footnote. The pipeline architecture has to encode that intent.

---

First-Party Data Programmes That Actually Feed AI

Building a first-party data programme that’s useful to AI agents means thinking about structure from the consent layer up — not retrofitting structure after collection.

Practically, this means three things. First, consent metadata needs to be machine-readable and attached at the record level. Not a checkbox in a CRM field, but a structured attribute that travels with the data wherever it moves — into a data warehouse, a CDP, or an agent’s context window. Brands like Lazada have invested in this kind of consent infrastructure precisely because their data surfaces across so many downstream systems.

Second, data taxonomy matters more than data volume. A deterministic extraction loop — the approach Chong describes — only works if the categories it’s extracting into are clearly defined and consistently applied. For multilingual Southeast Asian audiences, this includes language-normalised fields and region-aware segmentation that doesn’t collapse Thai, Bahasa, and Filipino consumer behaviour into a single generic bucket.

Third, build validation checkpoints before the synthesis layer. AI agents work best when they receive pre-validated, structured inputs. Routing unvalidated data directly into agent workflows is the equivalent of asking a new analyst to present findings from a spreadsheet they’ve never audited. The output might look confident. It probably isn’t.

The Competitive Moat Is the Infrastructure Nobody Sees

The brands that will differentiate on AI over the next three years won’t be the ones that adopted the latest model first. They’ll be the ones that made their first-party data genuinely reliable — structured, consented, and connected — before the agents arrived.

This is, in some ways, the unsexy version of the AI story. It doesn’t have the drama of a GPT-5 demo. But in markets where consumer trust is hard-won and regulatory environments across PDPA in Thailand, PDPB in the Philippines, and Singapore’s PDPA are tightening, a first-party data programme that’s compliant by design isn’t just good practice. It’s the only foundation an AI stack can safely stand on.

The question worth sitting with: if you ran an audit of your current data pipeline today and mapped every record to its consent basis and structural completeness, how much of your customer intelligence would you actually trust enough to hand to an agent?

---

At grzzly, we help brands across Southeast Asia build first-party data programmes that are designed from the ground up to feed modern activation — including AI-driven workflows — without cutting corners on consent or data quality. If your team is trying to figure out where the architecture gaps are before your next AI investment, we’d like that conversation. Let’s talk