Agent Z
Scaling enterprise support with an intelligent AI front door
Problem
Zscaler’s support front door had become fragmented and difficult to navigate. Customers could search help documentation, send email, call support, use multiple Customer Support Portal login experiences, or submit a case through an unauthenticated public form that anyone could access.
The result was more than a poor experience. Support received approximately 325,000 tickets annually across more than 10,000 customers, with 35% of that volume coming through the unauthenticated form alone. Ticket demand was growing faster than the organization could add capacity, while inconsistent entry points made cases harder to frame, route, and resolve.
Executive leadership wanted 10% AI deflection, but the organization first needed an intelligent front door. You cannot deflect demand that you cannot authenticate, understand, or route.
Context
Agent Z began as a product strategy for redesigning support entry, not simply adding a chatbot to the existing portal. The opportunity was to create one intelligent front door that could answer questions, guide customers toward the right resolution, collect better context when a case was necessary, and reduce avoidable demand before it reached Support.
Zscaler had no existing AI deflection capability at this scale. The product needed to work across more than 10,000 customers, thousands of technical documents, and a wide range of cybersecurity products and configurations. Accuracy mattered because an incorrect answer could delay resolution or provide harmful technical guidance.
The strategy eventually expanded into a multi-agent support stack. Agent Z became the customer-facing intelligent front door and deflection product. ResolvR became its feedback loop, closing knowledge gaps by identifying unanswered questions and automatically creating new knowledge-base and help documentation. Teammate became the internal sister product that helped support engineers investigate and resolve the complex cases that still reached them.
Zscaler had no existing AI deflection capability at this scale. The product needed to work across more than 10,000 customers, thousands of technical documents, and a wide range of cybersecurity products and configurations. Accuracy mattered because an incorrect answer could delay resolution or provide harmful technical guidance.
The strategy eventually expanded into a multi-agent support stack. Agent Z became the customer-facing intelligent front door and deflection product. ResolvR became its feedback loop, closing knowledge gaps by identifying unanswered questions and automatically creating new knowledge-base and help documentation. Teammate became the internal sister product that helped support engineers investigate and resolve the complex cases that still reached them.
My Role
As Director of Business Systems Product Management, I owned the end-to-end product vision and strategy for Agent Z, including the support front-door redesign, build-versus-buy decision, knowledge strategy, customer experience, staged launch, and go-to-market decision that made Agent Z the primary path to case creation.
Strategy & Process
- 1Fix the Front Door FirstConsolidated multiple support entry experiences into one authenticated path and deprecated the unauthenticated public form, reducing low-quality and unnecessary demand before AI deflection began.
- 2Prove Build Versus BuyRan a competitive bake-off between Salesforce Agentforce and an internally developed Agent Z prototype. Evaluated both options against cost, content ingestion, accuracy, scalability, and the support experience Zscaler needed.
- 3Treat Knowledge as Product InfrastructureDirected an audit of approximately 6,000 help documents and knowledge articles. The team removed duplicates, stale content, and conflicting answers, producing a curated corpus of approximately 5,000 authoritative documents.
- 3Measure Real Task SuccessTested whether Agent Z could answer real technical questions accurately, not simply retrieve relevant documents. Subject-matter experts evaluated the correctness of answers and the quality of customer guidance.
- 3Launch in Controlled StagesBegan with a five-customer beta to validate accuracy and the end-to-end experience, expanded to approximately 100 customers to test scalability and deflection, and then released Agent Z to more than 10,000 customers.
- 3Separate Early Signals from Scaled ResultsTreated the 70% deflection observed during limited availability as a promising small-cohort signal, not a production claim. After general availability initially produced roughly 30% deflection, used scaled behavior to identify remaining knowledge gaps.
- 3Build a Continuous Feedback LoopCreated ResolvR to analyze the questions Agent Z could not answer, identify missing knowledge, and automatically generate new knowledge-base and help content. This allowed the system to improve coverage instead of repeatedly failing on the same questions.
- 4Extend the Architecture InternallyApplied the same AI and knowledge architecture to Teammate, an internal sister product that helps support engineers investigate and resolve the complex cases that still reach them.
Key Decisions
- 1Build Instead of BuyChose the internally developed Agent Z after it outperformed Agentforce in the bake-off. Agentforce’s cost and content-ingestion limitations made it a poor fit for the required scale and knowledge architecture.
- 1Redesign Before DeflectingRefused to treat AI as a layer on top of a broken support journey. Simplified authentication, entry, routing, and case creation before measuring Agent Z’s deflection performance.
- 1Prioritize Accuracy Over SpeedAccepted a slower first phase while the team curated the knowledge corpus and raised task success above the 90% launch threshold. A fast release with unreliable technical guidance would have damaged customer trust.
- 1Create One Intelligent Front DoorSecured C-suite support to make Agent Z the unified path to support case creation, removing the fragmented alternatives that prevented consistent guidance, context gathering, and measurement.
- 1Scale in StagesUsed beta and limited availability as distinct learning environments rather than moving directly from prototype to general availability. Each stage tested a different risk: experience, accuracy, scalability, and customer behavior.
- 2Do Not Overclaim the PilotDeclined to present 70% limited-availability deflection as the expected scaled result. Used the lower initial general-availability rate to expose where the product and knowledge corpus still needed improvement.
- 2Build ResolvR as a Feedback LoopChose to address knowledge gaps systematically instead of relying on recurring manual cleanup. ResolvR automatically creates new support content from unanswered questions, allowing Agent Z’s coverage and deflection to improve over time.
- 3Design a Multi-Agent StackPositioned Agent Z, ResolvR, and Teammate as connected products serving different parts of the support journey: customer deflection, knowledge creation, and engineer-assisted resolution.
Impact
- 1Front-Door Demand ReducedDecreased annual ticket volume 26%, from approximately 325,000 to 240,000, by closing the unauthenticated public form and consolidating entry paths before AI deflection began.
- 2Deflection ScaledDelivered 37% deflection across the first full year after general availability, 3.7 times the C-suite’s original 10% target.
- 2Accuracy Reached Launch QualityIncreased answer accuracy from approximately 50% to more than 90% through corpus curation and human validation, meeting the product’s launch threshold.
- 2Resolution AcceleratedReduced mean time to resolution 40% across all support cases through the combined effect of Agent Z, ResolvR, and Teammate.
- 2Customer Growth AbsorbedKept support headcount flat while customer and case demand grew, allowing the organization to absorb year-over-year growth without additional support hires.
- 2Customer Friction Stayed LowMaintained abandonment below 3%, measured as customers leaving before receiving an answer. Abandoned sessions were excluded from the deflection calculation.
- 2Knowledge Became Self-ImprovingTurned unanswered questions into inputs for new knowledge-base and help documentation, continually closing gaps in Agent Z’s corpus.
- 3A Reusable AI Platform EmergedExtended the architecture beyond customer deflection into internal support-engineer assistance, creating a connected multi-agent support platform rather than a standalone chatbot.
Lessons Learned
The knowledge corpus is the real ceiling on AI accuracy. Model capability and prompt design matter, but neither can overcome stale, duplicative, or contradictory source material. Treating content quality as product infrastructure was essential to getting Agent Z above the 90% accuracy threshold.
The staged rollout also reinforced the difference between an encouraging pilot and a defensible scaled outcome. The 70% limited-availability deflection rate showed potential, but the lower initial general-availability result exposed real coverage gaps. Building ResolvR turned that gap into a product opportunity and created the feedback loop Agent Z needed to improve over time.
Finally, the bake-off changed my broader product strategy. Building internally was not automatically the right answer, but the comparison proved that a focused internal product could outperform a major platform vendor when the organization’s scale, data, workflow, and experience requirements were sufficiently differentiated.
The staged rollout also reinforced the difference between an encouraging pilot and a defensible scaled outcome. The 70% limited-availability deflection rate showed potential, but the lower initial general-availability result exposed real coverage gaps. Building ResolvR turned that gap into a product opportunity and created the feedback loop Agent Z needed to improve over time.
Finally, the bake-off changed my broader product strategy. Building internally was not automatically the right answer, but the comparison proved that a focused internal product could outperform a major platform vendor when the organization’s scale, data, workflow, and experience requirements were sufficiently differentiated.
