Agentic Investigation Platform for Integrity Teams

Context

I owned the product strategy and execution for Meta's agentic investigation platform for integrity teams.

The product gave reviewers, investigators, policy teams, enforcement teams, and cross-functional partners a single place to investigate complex cases. Instead of forcing users to jump across case tools, signal dashboards, runbooks, source-of-truth systems, and specialized agents, the platform let them start with a natural-language question and receive grounded reasoning, evidence, summaries, and recommended next steps.

The problem

Integrity investigations were becoming too fragmented to scale. Investigators had to inspect cases and objects, extract IDs, look up enforcement history, interpret signals, follow workflow-specific runbooks, query specialized tools, draft notes, write responses, and maintain evidence trails. Too much of this work depended on tool knowledge, tribal knowledge, and manual synthesis.

At the same time, teams across integrity were building specialized AI agents for narrow domains such as account compromise, URL analysis, enforcement history, phishing, paid compromise, and policy interpretation. That created a new platform problem: even with more agents, users still had to know which one to use, what input it required, whether its output was authoritative, and how to combine results across tools.

The insight

The opportunity was not to build a chatbot. It was to build the orchestration layer for agentic investigation. Investigators should investigate problems, not tools. The platform needed to understand the user's intent, resolve the right entities, retrieve the right evidence, invoke the right source system or specialized agent, and synthesize the result into an investigation workflow the human could trust.

The key product decision was to separate investigation modes. Open-ended investigations should stay human-led. Strict workflows should become procedural runbook execution. Low-risk, high-confidence steps could move toward automation. Each mode needed a different user experience, launch bar, evaluation method, and governance model.

What I did

I built the core investigation loop: understand the question, resolve the relevant entities, retrieve case and object context, invoke the right capability, synthesize evidence, and preserve follow-up state so the investigation could continue naturally.

I defined the platform as an orchestration layer, not a replacement for domain agents. Specialized teams continued to own their expertise; the platform owned routing, user experience, evidence synthesis, governance, and the end-to-end investigation experience.

I shaped the agent ecosystem model: registration, input and output contracts, capability routing, multi-agent planning, quality tiers, ownership, monitoring, and removal criteria. That allowed the platform to grow without becoming an uncurated list of unsupported tools.

For the first hero workflow, I helped reframe account-compromise investigation from open-ended AI chat into procedural investigation support. The system needed to follow runbooks, resolve entities, detect duplicates, classify issues, check compromise signals, distinguish recommendations from actions, ask for clarification when identifiers conflicted, and stop safely when evidence was incomplete.

I also changed how we measured success. Usage alone was not enough. The product needed to show whether investigations were shifting from human-only work to human-plus-agent work, whether some workflows could become autonomous, whether accuracy improved, and whether investigators completed work faster with stronger evidence.

Impact

• Reached 94% investigation accuracy, above human-expert performance, with 1,000+ weekly active users and 7,000+ AI-led investigations a day.
• Reduced investigation friction by giving teams one place to ask questions, inspect evidence, understand signals, invoke specialized capabilities, and draft next steps.
• Created the platform model for trusted agentic investigation: intent understanding, entity resolution, capability routing, source-of-truth integration, evidence synthesis, human confirmation, evaluation, monitoring, and governance.
• Moved the product from a single-use-case assistant toward an investigation operating system, where specialized agents could improve independently while users experienced one coherent investigation surface.

What I learned

In high-trust investigation workflows, fluency is not enough. The product has to know what workflow it is in, what evidence is required, which source system is authoritative, what it can recommend, when it needs human confirmation, and when it must stop.

Agentic products need clear product contracts. A copilot, a runbook executor, and an automation engine can all use AI, but they should not be designed, measured, or launched the same way.

Platform governance is product quality. If an agent is unsupported, poorly evaluated, or unclear about its limitations, users blame the platform. For agentic systems to scale, quality has to be earned through evidence, ownership, monitoring, and clear boundaries.