Agentic Workflow Automation for Integrity Escalations

Context

I owned the product strategy and execution for workflow automation across Meta's global integrity escalation workflows.

The work started with the central triage workflow, the front door for many high-priority escalations, and expanded from triage into investigation, enforcement, response, and case closure. Over time, it became a broader automation platform where operations teams could build, manage, monitor, and improve workflows through a governed lifecycle.

The problem

Integrity escalation volume was growing faster than operational capacity. Analysts were manually reading incoming cases, filling structured fields, determining priority, routing cases to downstream teams, investigating across fragmented tools, coordinating handoffs, taking enforcement action, updating cases, sending responses, and closing the loop.

This work was repetitive, time-sensitive, and difficult to scale. For front-door triage, the pressure was especially high because the workflow required fast response, around-the-clock coverage, and consistent intake quality for high-priority escalations.

The insight

The unlock was portfolio thinking. The global integrity operations org had thousands of escalation workflows, but not every workflow was ready for automation. Some were high-volume and rule-bound. Some were low-volume but high-risk. Some had mature runbooks. Others depended on expert judgment. Some had clean systems of record. Others were stitched together through manual workarounds.

Before we could scale automation, we had to turn the workflow universe into a prioritized product portfolio: which workflows were valuable, automatable, safe, repeatable, and ready to move through scoping, building, testing, shadow mode, human oversight, and operational monitoring.

What I did

I worked with global integrity operations to prioritize the highest-value workflows based on volume, business criticality, risk, and feasibility.

I first focused on central triage, building automation for repeatable intake work: custom field filling, routing, duplicate detection, subcase updates, case summaries, and reviewer context. The system documented what it did and why directly in the case, creating an auditable trail for reviewers, operations teams, and quality measurement.

I then expanded the product from triage to end-to-end workflow automation across investigation, enforcement, response generation, and closure. Low-risk, high-precision work could be automated; judgment-heavy work stayed with humans.

To scale beyond bespoke builds, I drove the workflow inventory and prioritization model across the broader workflow surface, then ran an eight-week co-location sprint with operations owners and forward-deployed engineers to convert priority workflows into reusable automation patterns.

I also shaped the self-serve automation hub: a central place to register workflows, track lifecycle phase, test cases safely, run shadow mode, manage launch approval, monitor human oversight, graduate workflows to live operation, and track portfolio-level health, failures, alerts, and automation opportunities.

Impact

• Launched triage automation for central intake workflows, cutting repetitive manual intake work and creating the first reusable pattern for escalation automation.
• Expanded from triage to end-to-end automation across investigation, enforcement, response, and closure. One sprint launched automation for 20 high-volume workflows, projected to save 1.4M content reviews, 140K cases, and 95K hours a year.
• Helped shift integrity operations from fragmented manual execution to governed automation, where workflows could be built, tested, launched, monitored, and improved through a repeatable lifecycle.

What I learned

Workflow automation is not just about replacing manual steps. It is about understanding where judgment, risk, policy, and operational context sit inside a process.

The hardest part is often not the agent or the builder. It is turning undocumented human behavior into structured, measurable, auditable systems that teams can trust.

A durable automation platform needs more than one successful launch. It needs prioritization, reusable patterns, launch gates, quality guardrails, monitoring, ownership, and a way for workflow owners to keep improving without depending on custom engineering every time.