Augmenting Intelligence with AI-Powered Trend Discovery for a PE-Backed Intelligence Platform

Fractional AI built a research agent system that autonomously surfaces and validates emerging market trends. Using structured debate between reasoning agents, it delivers data-backed, decision-ready reports in under 35 minutes for about $3 – laying the groundwork to multiply Analyst throughput and speed up proactive trend discovery for a PE-backed intelligence platform.

Opportunity

Generating high-quality insights from this market intelligence platform's data has traditionally required Senior Analysts with significant experience. Even then, Analysts were limited to primarily responding to client-initiated questions. Meanwhile, important signals – internet trends, consumer sentiment, emerging behaviors – often precede measurable sales shifts, but there was no scalable way to detect or validate them.

The company wanted to flip the model: proactively identify trends their clients hadn't yet considered, then validate those signals using proprietary point-of-sale data.

Solution

Fractional AI built a modular research agent system designed to mimic the workflow of an experienced Analyst. The system operates in two stages: topic generation and structured evaluation.

In the first stage, agents generate candidate "areas of concern" by scanning public web data for emerging trends tied to a specific brand or category (e.g., Emerging Threat of GLP-1 Therapies to Calorie-Sweetened Soda Demand). These agents surface both direct and adjacent trends that could impact a client’s performance. These topics are selected for their relevance, novelty, and potential impact, covering both direct threats and adjacent opportunities.

In the second stage, each topic is analyzed through a multi-agent debate: two AI agents (affirmative and negative) argue for and against the trend’s importance using a blend of sales data and external web evidence. This reasoning-focused model allows each side to make informed arguments based on data.

The output is then evaluated by a third agent simulating a brand executive, who reviews the arguments and renders a final decision on the topic’s relevance within a 12–24 month business horizon. Each finalized report summarizes the decision and rationale in clear language, suitable for delivery to clients.

Impact

This trend report agent sets the foundation to shift from reactive analytics to forward-looking and insight-driven consulting.

Key outcomes:

• Faster time-to-insight: Reports are generated in under 35 minutes, enabling rapid delivery of insights to clients.
• Cost-efficiency: Each report costs approximately $3 to produce.
• Proactive detection: Identify early trends backed by internal data.
• Scalability: The modular framework sets the stage for future productization.

Looking Under the Hood

Mitigating Model Bias through System Design 

Analysts do two jobs – find evidence and make an informed decision. LLMs often blur these tasks, steering research toward their own preferred outcome. The architecture for this project separates the roles and uses intentional bias – pitting pro and con researchers against each other – so a neutral judge decides based on briefs, not self-justified searches.

The critical design choice was not just modularity, but limiting each agent’s scope. When one agent both gathers evidence and renders judgment, its research naturally drifts toward supporting its own view. This system inverts that tendency: one agent is tasked to argue “this matters,” another to argue “it doesn’t,” and a third to decide. The judge reads the briefs without running its own tools, avoiding feedback loops and producing more stable outcomes.

The system is a set of composable agent loops – topic generation, proposition evaluation, debate orchestration, stakeholder simulation – built with OpenAI’s Agent SDK. Tool access is role-scoped: research agents can search the web and query sales data; the judge cannot. This keeps research directional but judgment conservative and tool-independent. Modularity also allows updates to any loop without disturbing the rest.

‍

Tooling and Infrastructure

• OpenAI Agent SDK for managed agentic workflows, tool chaining, and context management.
• Streamlit for displaying intermediate builds and iteration; FastAPI for production-grade integration.
• Braintrust for evaluating topic quality, proposition relevance, and decision consistency across runs.

Evaluations

Because the core question, “does this matter to a brand executive in the next 12–24 months?”, is inherently subjective, we focused evaluation on reliability and quality rather than a single accuracy score. Despite the ambiguity, we used structured checks at each stage to improve the system:

• Topic generation was scored for diversity, actionability, insightfulness, and coverage.
• Proposition generation used a lightweight “LLM-as-judge” approach to assess whether each proposition was important and relevant to the topic.
• Stakeholder decisions were evaluated for consistency (across multiple runs) and “correctness” based on human-labeled ground truth. While correctness was subjective for some topics, consistency across runs became a key quality signal.

Balancing Speed, Quality, and Cost in Production

Performance tuning focused on cost, quality, and latency. In production, a full report took 25–35 minutes with o4-mini at about $3 in compute. Quality was strong: ~80% of report pairs reached the same verdict on re-run. The adversarial research + independent adjudication pattern drove that consistency by keeping the judge out of the evidence-gathering process.

This behind-the-scenes architecture created a repeatable, extensible system for trend discovery and validation – paving the way for broader productization.