Designing AI triage and RAG self-service for enterprise support

Enterprise support creates two different information problems. The first is routing: understanding a request well enough to get it to the right team. The second is resolution: finding a trusted answer quickly enough to help the user or support professional act.

At PepsiCo GCC, I led Kognito-based AI triage for misrouted tickets and conceived Oscar, a user-facing RAG-enabled ITSM chatbot for guided self-service and standardized responses. I later expanded the AI-led operations roadmap with an SRE agent, email-to-ticket automation, RAG analytics, and agentic integration patterns.

The leadership lesson was not that AI can replace a service organization. It was that AI can make the service system easier to navigate - if the operating model around the model is designed deliberately.

Choose problems with a visible cost of friction

Misrouted tickets create avoidable queue time, repeated reassignment, unclear ownership, and user frustration. Knowledge search creates a different form of delay: support professionals and users may find multiple documents, outdated guidance, or no clear answer.

Both are good AI candidates because the current friction is observable and a human fallback already exists. That gives the team a baseline, a safe escalation path, and meaningful measures beyond a model demo.

Separate classification from authority

An AI classifier can recommend a category or assignment group. That does not mean it should always make the final decision. I treat confidence as an operating input. High-confidence, low-risk patterns may be automated. Lower-confidence or high-impact cases should be routed to human review with the recommendation and evidence visible.

This separation protects service quality and creates a learning loop. Overrides are not failures to hide; they are labeled feedback that can improve data, prompts, rules, knowledge, or the model.

Ground answers in governed knowledge

RAG improves usefulness by retrieving relevant enterprise content before generating a response. But retrieval only helps if the source knowledge is trustworthy. Oscar therefore represents a service-management challenge as much as an AI challenge.

• Curate the source setIndex approved, relevant content rather than everything that can be reached.
• Assign knowledge ownershipEvery important article needs a steward, review cycle, version, and retirement path.
• Expose groundingUsers and support teams should be able to understand which source informed the answer.
• Design an escalation pathThe experience must say when it is uncertain and make hand-off to the right team straightforward.
• Protect access boundariesRetrieval must respect the permissions, sensitivity, and service context of the person asking.

Measure the service outcome

Model accuracy is important but insufficient. For triage, I look at reassignment patterns, time to correct ownership, confidence distribution, overrides, and the downstream effect on MTTR and SLA performance. For RAG self-service, useful measures include grounded-answer quality, escalation rate, unresolved questions, knowledge gaps, repeat usage, and the consistency of support responses.

These measures keep AI connected to the operating objective. They also help distinguish a problem in the model from a problem in the underlying process or knowledge base.

Create ownership across the lifecycle

AI-enabled support crosses product, service management, engineering, security, data, knowledge, and operations. I make responsibilities explicit: who owns the use case, the source content, the model behavior, the integration, the controls, the performance review, and the decision to expand or pause automation.

Without that ownership, a pilot can appear successful while the knowledge grows stale or exception handling remains manual. With it, the capability can be operated and improved like any other enterprise service.

Responsible AI in operations is not a disclaimer attached to a model. It is a set of decision rights, controls, evidence, and human hand-offs built into the service.

Closing perspective

Kognito-based triage and Oscar demonstrated two practical roles for AI: reducing the distance to the right owner and reducing the distance to a trusted answer. Their value depended on grounded knowledge, confidence-aware automation, visible fallback, and service measures.

That remains my approach to AI-led operations. Start with a real point of friction. Define the safe decision boundary. Keep ownership visible. Learn from overrides. Expand only when the service outcome supports it.

Professional note: This article reflects my leadership experience and personal perspective. Examples are intentionally generalized and do not disclose confidential company information.