Abhilash Asokan is an AI engineering director at ZS whose work applies agentic systems to pharmaceutical commercial analytics. He designs systems that connect prescription anomalies to their underlying causes and recommend actions grounded in the realities of drug coverage, geography, and brand performance.
At AI Engineer World’s Fair 2026, Asokan and fellow ZS engineering leader Subbiah Sethuraman described why their initial multi-agent analytics pipeline produced incoherent recommendations. Separate agents handled signal detection, localization, causal analysis, and synthesis, losing essential business context at each handoff. One system correctly attributed falling prescriptions to reduced insurance coverage but recommended deploying more sales representatives—an intervention unrelated to the actual problem.
Asokan rebuilt the architecture around three principles:
- Deterministic signal detection: Statistical methods, thresholds, guardrails, and prioritization identify meaningful prescription anomalies before an agent begins investigating.
- Centralized agent judgment: One primary agent owns end-to-end reasoning and delegates narrowly scoped investigative tasks without surrendering responsibility for connecting evidence, causes, and recommendations.
- Knowledge graphs as control planes: Relationships among geographic markets, insurance payers, brands, and performance indicators constrain which hypotheses the agent can pursue. The agent tests each plausible graph relationship against underlying data, narrowing national prescription declines to specific regions, accounts, payers, or contributing metrics.
His redesign drew directly on observing how Claude Code investigates problems and dynamically launches focused subagents. The resulting pharmaceutical analytics architecture assigns calculations to conventional software, investigation to an agent, and business context to an explicit domain model.
