Self-ServeAISupport

The default ask was a generalist AI assistant spanning the whole platform: ten modules, every PM and engineer aligned before a single pixel shipped. That is years of committee gridlock, and the broken self-serve experience was bleeding cost now.

I scoped down to a specialist tool covering only the support-intake flow. Narrow enough to ship, valuable enough to prove the concept. I gave up breadth on purpose: a specialist routes into a broader assistant later, so this was an additive first step, not a competing one.

My first assumption was that we needed a better chat interface. Research killed it: support engineers cited lack of customer context as their number one bottleneck. The interface was not the problem, the inputs were. No matter how capable the LLM, vague inputs produce vague answers.

So I scrapped freeform chat. Open chat puts the burden of context on the user, and admins under pressure give almost none. I replaced it with a structured checklist capped at the top three suggestions by model confidence. Users lost the ability to phrase things their own way; the LLM gained context it could actually use. Each checked item silently enriched the prompt, so every Regenerate produced a sharper answer at no extra effort.

Our users were senior network engineers who reverse-engineer failures for a living. A confident, wrong suggestion is the fastest way to lose them, and once trust is gone it does not come back. Cisco's internal model, "Sherlock," was capable but locked behind engineers who approved every output, because a hallucination on a production network is a real liability.

Making it safe for customers was an interaction problem. When inputs are thin, the tool asks probing questions instead of guessing. When answers keep failing and model confidence drops below a threshold, it escalates to a human at TAC. And I constrained retrieval to Cisco's verified internal documentation, surfaced as popover footnotes. The constraint was the product: a RAG pipeline that can only cite first-party Cisco docs cannot wander the open internet, and visible provenance let a skeptical engineer verify nothing was invented.

The cheapest support case is the one never opened

The tool intercepts support cases at the source, before a ticket is ever opened. Business analysts projected $15M in annual operational savings from the deflection.