The Timing Problem
Agentic commerce systems — AI agents that make recommendations, allocate budgets, and execute purchases on behalf of users — operate on a fundamentally different time scale than traditional consumer research. An agent browsing product options, comparing alternatives, and making a purchase decision does so in seconds or milliseconds. The preference data it needs to make those decisions well must be available at the same speed.
This creates what we call the preference timing constraint. Agents encounter thousands of scenario variations per user: different product categories, price points, brand alternatives, feature configurations, and contextual factors. For each variation, the agent needs to know what a consumer with this profile would prefer. It cannot pause to commission a survey. It cannot wait days for a panel to field. It needs an answer in milliseconds.
Two Non-Negotiable Requirements
Any system that provides preference data to agentic commerce must satisfy two requirements simultaneously. First, speed: the system must operate at agent time scale, returning results in milliseconds, not hours or weeks. Second, accuracy: the results must be accurate enough to drive real purchasing decisions with real money.
Traditional market research satisfies the accuracy requirement — well-designed surveys produce reliable preference data — but fails on speed. A consumer survey takes days to weeks to field, analyze, and deliver. Conversely, simple heuristics and rule-based systems satisfy the speed requirement but fail on accuracy. Hardcoded preference rules cannot capture the nuance of real consumer decision-making across thousands of scenario variations.
The gap: Traditional research delivers accurate preferences too slowly. Heuristics deliver fast preferences too crudely. Agentic commerce needs both speed and accuracy simultaneously — and until now, nothing has provided both.
Synthetic Respondents as the Preference Layer
Synthetic respondent systems satisfy both requirements. They operate at agent speed — returning preference estimates in milliseconds via API — and they are validated against live survey data to ensure accuracy. This combination makes them uniquely suited to serve as the preference layer for agentic commerce infrastructure.
The architecture is fundamentally different from traditional research. Instead of discrete research projects that produce static reports, synthetic respondents function as continuous preference subsystems: always available, scenario-specific, on-demand, and embedded directly in the agent's decision loop. An agent does not request a research project and wait for results. It queries a preference API and receives a distributional estimate of consumer preference for a specific scenario in real time.
From Research Projects to Preference Subsystems
This shift — from research projects to preference subsystems — changes the economics and operational model of consumer insight. Traditional research operates on a project basis: scope, field, analyze, report, repeat. Each project costs tens of thousands of dollars and takes weeks. The output is a static snapshot that begins decaying the moment it is delivered.
A preference subsystem operates continuously. It provides unlimited scenario variants at a cost of pennies per query, with latency measured in milliseconds. There is no fielding period, no analysis phase, no report generation. The agent queries, receives a preference distribution, and acts. The cost structure shifts from large fixed investments per project to marginal per-query costs that scale with usage.
For agentic commerce platforms, this means preference data becomes infrastructure rather than a discretionary research expense. It is embedded in the system architecture the same way a pricing API or inventory feed is embedded — as a continuous, real-time data source that the agent consumes programmatically.
Validation in an Agentic Context
The validation approach for preference infrastructure must be rigorous precisely because the outputs drive automated decisions. Every synthetic respondent model used in agentic commerce is validated against published live survey benchmarks using KL divergence for single-select items and Rank-Biased Overlap for multi-select items. These validation metrics are computed at the question level, not just in aggregate, so that downstream systems can assess confidence for specific preference dimensions.
This is not a theoretical exercise. If an agent is allocating a consumer's grocery budget across brands, the preference estimates for each category must be accurate enough that the consumer recognizes the output as reflecting their actual preferences. Validation against live surveys is the mechanism that provides that assurance.
The Opportunity
Agentic commerce is still in its early stages, but the infrastructure requirements are becoming clear. Agents need real-time access to consumer preference data that is accurate, scenario-specific, and available at machine speed. Synthetic respondent systems are the only current technology that satisfies all three requirements simultaneously.
The full technical architecture, validation methodology, and implementation framework are detailed in our CTO Brief white paper (PDF). For more on the Oracle preference API, visit the Oracle product page.