Ganu.

The problem

Customer expectations have shifted with the Gen AI wave: customers now expect outcomes, not just a tool to do the job. But the AI customer support tools they encounter still get the most consequential moment wrong — the escalation from bot to human.

The friction stack: escalation timing is too early (wasting human support time on issues the bot could have handled) or too late (frustrating the customer who's been stuck in a loop for ten turns); no handoff summary for the support agent (just a raw transcript); missing context (the agent has to re-ask the customer for details already given); no state transfer to the case tool (the agent re-creates the case from scratch in Salesforce); and unclear escalation paths when complex issues need a specific team.

The solution

Ganu introduces two components alongside the standard support bot: a Silent Evaluator that runs every turn and decides when escalation is needed based on confidence, sentiment, and escalation criteria; and a Structured Handoff Card that displays a concise summary of everything the agent needs — Issue Summary, Steps Already Tried, Recommended First Action, Customer Sentiment, Routing Suggestion — in a side panel, instead of dumping a raw chat log.

The result: the human agent walks into the conversation already oriented, the customer doesn't have to re-explain, and escalation happens at the right moment rather than too early or too late.

How it works

Conversation flow: customer enters problem details → support bot asks clarifying questions, keeps track of resolution confidence + sentiment → the silent evaluator runs every turn, scoring against escalation thresholds → at the escalation point, the Handoff Generator builds the structured card → the case is routed to the appropriate team (Salesforce integration in next release).

Multi-agent model strategy matched to each role's requirements: Opus 4.6 for the Support Bot (high precision in customer-facing replies justifies the cost), Evaluator agent runs every turn (straightforward task, JSON output, candidate for a cost-efficient optimal model), Handoff Generator produces the card. All three share the conversation transcript as input; each has a distinct system prompt tuned to its function. Evaluation criteria: escalation precision, handoff card completeness and clarity, routing relevance, accuracy of "steps tried," sentiment accuracy, confidence calibration, and coverage of edge cases.

Who it's for

External users: the product company's customers' data engineers and engineering staff — they interact with the Support Bot 24/7 to resolve product issues, expect technical depth, and don't want to re-explain themselves on escalation.

Internal users: Tier 1 Customer Support Engineers at the product company — they receive escalations and benefit from the structured handoff card. Buyer: the product company itself (B2B), looking to reduce escalation volume, improve support agent productivity, and raise customer satisfaction on the issues that *do* escalate.

Why it matters

Hard competition is reshaping the customer support category — newer players with niche capabilities are wiping out legacy support tools that don't deliver outcomes. The Gen AI wave has raised the customer expectation bar from "respond fast" to "actually resolve" — and the products that can't transition lose share quickly.

Ganu's structural bet is that the bot-to-human handoff is the highest-leverage point to fix. Every other AI support tool focuses on the conversational interface or the knowledge base; almost none invest deeply in the handoff layer where the productivity gain (or loss) compounds across every support interaction. Structured handoff cards + silent evaluator + auto-routing is a defensible position because it requires upfront investment in evaluation infrastructure that AI-first support competitors haven't built.

At a glance

Project

Ganu

Built by

Nilesh Acharya

One-liner

AI customer support bot with a silent escalation evaluator and structured handoff cards — fixes the weakest link in support: the human handoff.