Why ZS ditched its multi-agent AI pipeline for pharma analytics

This article summarizes an episode of AI Engineer’s video series featuring Subbiah Sethuraman, head of AI engineering at ZS.

Image credit: Timmy Loen

In business analytics, chasing the wrong problem, even based on accurate data, is a costly mistake.

Subbiah Sethuraman, head of AI engineering at the healthcare consulting firm ZS, explains that this risk increases when AI agents are allowed to operate without strict guardrails.

To generate reliable recommendations in industries like pharmaceuticals, ZS found that standard statistical checks must trigger the investigation, and a single AI agent must own the entire decision process.

Verified task lists improve agent performance

Standard statistical methods should be used to detect issues before the AI ever gets involved. At ZS, the AI only wakes up when a verified signal is added to its queue.

This workflow keeps the AI focused on high-priority issues:

• Automate discovery: run scheduled statistical scans to identify anomalies, keeping this separate from user prompts.
• Filter the noise: set thresholds to separate true business issues from normal market fluctuations.
• Prioritize impact: direct the AI model to focus only on anomalies that directly affect revenue.
• Provide clear context: feed the AI a task list that includes the specific issue and a normal baseline for comparison.
• Define the role: restrict the AI strictly to explaining causes and suggesting solutions, rather than finding the initial problem.

By reserving AI for cases that have already passed math checks, businesses save computing time, prevent unnecessary scrambles by sales teams, and produce defensible results.

Multiple agents can weaken responsibility

Fixing the starting trigger solved one problem, but ZS discovered a flaw in execution. Passing tasks across multiple AI agents weakened the final recommendation.

Each agent produced a factually correct step, but the system as a whole often lost sight of the business goal. It might identify a patient affordability issue, for instance, without also targeting the insurance companies responsible for the solution.

To fix this, ZS redesigned the workflow to rely on one central AI agent supported by narrower, task-specific sub-agents:

• Centralize reasoning: one main agent controls the investigation from the initial clue to the final recommended action.
• Delegate tactically: the main agent decides when to assign specific research tasks to other tools.
• Restrict sub-agents: smaller agents only gather evidence; they are never allowed to make business recommendations.
• Consolidate findings: all evidence flows back to the main agent to maintain retains authority over the conclusion.
• Verify the outcome: the final recommendation is cross-checked against the original root cause before generating a market report.

Adding more agents doesn’t necessarily create a smarter system. ZS found that centralizing logic prevents fragmented reasoning and produces more reliable market strategies.

Knowledge graphs guide exploration