The Human Preference Layer for AI Agents

AI agents are everywhere. They’re running ad campaigns, buying products on behalf of consumers, managing customer lifecycles, handling multi-step business workflows end-to-end. Things that used to take weeks now happen in minutes, no humans required. That’s great for speed. But something important got left behind.

The old systems had people in them. Marketers who actually understood their audience. Researchers who tested messaging before anything went live. Strategists who knew—from experience, from data, from gut—that 35-year-old mothers in the Midwest respond to safety messaging differently than 25-year-old urban professionals respond to performance messaging. That understanding was baked into the process. It was slow, but it was real.

Now the process runs at machine speed and that layer is just… gone. Agents are fast, they’re efficient, but they have no idea what people actually think. They optimize for clicks and conversions and historical patterns, but they can’t tell you how a specific demographic feels about a product, a message, or a decision. To get that, you’d have to ask real people—which takes exactly the kind of time agents were built to eliminate.

That’s the gap we built the Oracle to fill. We think of it as the human preference layer for AI agents.

What the Oracle Actually Does

The Simsurveys Oracle is trained on tens of millions of survey questions and hundreds of millions of demographic and psychographic data points. It understands human preference across any subgroup you can define—age, gender, income, location, values, lifestyle, attitudes—and it returns answers in seconds, not weeks.

That changes things. Instead of an agent guessing what people want, it can ask. An agent building an ad campaign can check: how does this audience actually feel about this value prop? An agent selecting products for a consumer can ask: what do people like this prefer, and why? An agent making a strategic recommendation can ask: what does this group care about most?

The Oracle gives agents something they were never designed to have on their own—an understanding of people.

A Concrete Example

Think about how programmatic advertising used to work. A media buying team would build a campaign strategy grounded in real research—focus groups, surveys, brand studies—to figure out which messages landed with which audiences. They’d test creative, refine targeting, and launch knowing the campaign was rooted in what people actually responded to.

Now agents handle most of that end-to-end. They generate creative variants, pick audiences, set budgets, optimize bids, iterate—all on their own, all in minutes. But that research phase? The part where someone figured out what the audience actually cares about? That took weeks. It doesn’t fit into a workflow that runs in real time.

So the agent optimizes on click-through rates and conversion data. It knows what people did, but not what they think or feel or want. It’s reactive, not informed.

With the Oracle plugged in, the agent gets the same kind of audience understanding a research team would have delivered—but at the speed it needs. The agent asks, the Oracle answers, and the campaign launches grounded in actual human preference instead of pattern-matching on past behavior. Better targeting, more relevant creative, higher conversion—because the decisions are based on what people want, not what the agent assumes they want.

The Piece in Every Agentic Loop

We see the Oracle as the human preference layer for any agent-driven system. Marketing agents, commerce agents, product agents, strategy agents—any workflow where an AI is making decisions that affect people, and where knowing what those people actually want would lead to better outcomes.

The output isn’t static. It’s tailored to what the agent needs in the moment—the right subpopulation, the right question framing, the right level of detail. It fits into MCP-based architectures, API calls within agent chains, or whatever orchestration framework you’re using. Wherever an agent needs to pause and ask “what would humans prefer here?”—that’s where the Oracle goes.

Agents are good at doing things. We’re making them good at understanding people.