CoCo

1,000 new hires. 12 weeks. 20+ fragmented systems.

Pfizer split its marketing business between IPG and Publicis right before peak vaccine season. Publicis inherited Paxlovid, Abrysvo, and Comirnaty—and needed to scale from zero to nearly 1,000 staff in weeks.

I embedded with Pfizer, Publicis, and IPG for two weeks before writing any code. Shadowed content producers. Sat with MLR reviewers. Mapped the actual workflow.

What I found: staff spent 25+ minutes per asset just searching for approved templates and prior claims. 40% of MLR submissions were rejected—not for content issues, but for using wrong templates or outdated claims.

CoCo: a hybrid-cloud, multi-tenant governed AI platform for enterprise knowledge.

CoCo—"Company Companion"—is embedded in Microsoft Teams. Instead of searching 20 systems, employees ask CoCo. It retrieves approved content, answers policy questions, and guides new hires through processes—with citations back to source documents.

The operating environment was Azure-native for collaboration, identity, and enterprise content access: Azure ML + AKS for the pipeline, Cognitive Search for hybrid retrieval, Graph-RAG for relationship mapping across brands and regions. The platform architecture followed portable control-plane patterns aligned with Amazon Web Services (AWS) for agent orchestration, validation, and observability.

Every model output was treated as an untrusted artifact. Before any response surfaced to users or triggered downstream workflows, it passed through a validation layer: schema enforcement, citation verification against approved sources, policy checks aligned to MLR rules, and confidence thresholds. Outputs that failed validation were blocked, routed for clarification, or escalated—ensuring the system behaved predictably and prevented non-compliant content from entering production workflows.

Multi-tenancy was enforced across brand, region, and partner boundaries—ensuring each tenant had isolated retrieval scope, policy rules, and audit trails while sharing a common platform control plane.

Sources: Veeva Vault · SharePoint · Workfront · CLM · Regional Drives

Production RAG with governed retrieval and compliance gates.

CoCo doesn't replace Veeva, SharePoint, or Workfront. It sits on top. The systems of record remain authoritative. CoCo indexes their content, understands relationships, and retrieves the right information—with every answer traceable back to its source.

Built with clear SLOs: sub-second retrieval, 99.9% uptime, full audit trail. Every response includes citations. Every query is logged. MLR gateway validates compliance before content surfaces to users. Every agent decision carried a trace ID across retrieval, generation, validation, and tool execution—allowing full reconstruction of system behavior for debugging, audit, and evaluation.

Agentic onboarding: learn by doing, not by reading manuals.

Traditional onboarding fails at scale. You can't have 200 trainers for 1,000 new hires. CoCo turns onboarding into an agentic experience: new employees ask questions as they arise, get immediate answers with context, and become productive in days instead of months.

Trained 200+ users across content production, MLR review, and brand management. Created playbooks for common workflows. Established feedback loops to continuously improve retrieval quality.

The Result

Publicis onboarded nearly 1,000 staff in time for peak vaccine season. MLR cycles dropped from 42 days to 14 days. Content throughput tripled. New hires were productive in days because CoCo was always there to answer questions, find content, and guide them through unfamiliar processes.

CoCo became the default way to work—not because it was mandated, but because it was faster. Today it serves as Pfizer's vaccines content backbone, with expansion planned for additional therapeutic areas.