WhatsApp Sales Discovery Bot

The Job to Be Done

When a seller reaches out about using EshopBox, I want to understand their category, volume, and logistics setup immediately, so I can route them to the right sales rep without spending discovery call time on poor-fit leads. (User: EshopBox sales rep)

Problem

EshopBox’s sales team was spending significant time on discovery calls with sellers who were not a fit - wrong product category, wrong order volume, wrong geography. The qualification step was happening too late in the funnel, after human time was already committed.

The workaround was manual: a sales rep would take an initial call, run through a standard set of questions, and then either route or disqualify. This worked at low volume. At scale, it created a bottleneck where the best reps spent disproportionate time on unqualified leads.

Insight

The qualification criteria were deterministic. Category, monthly order volume, and current logistics setup could be captured via a structured conversation before any human involvement. The only reason a human was doing it was that no system existed to do it first.

Root cause (via 5 Whys): The qualification step had never been productized. It was a human process by default, not by necessity.

Key Decisions

Decision 1: WhatsApp over a web form

Sellers were already on WhatsApp. A web form would require them to find a URL, open a browser, and context-switch. WhatsApp meets them where they already are. Completion rate for conversational flows on WhatsApp significantly outperforms web forms for the seller demographic we were targeting.

Decision 2: Multi-LLM chain for intent parsing

A single model broke on ambiguous seller inputs - particularly for sellers who responded in fragmented messages or mixed languages. A multi-LLM chain with fallback routing improved accuracy on these edge cases.

AI Triangle trade-off: Quality + Cost, accepted higher Latency. In a WhatsApp conversation, users expect a few seconds between messages - the latency increase from the chain was within acceptable range. Misrouting a qualified lead costs more than an extra API call.

Decision 3: Qualification logic stays deterministic

The LLM was used only for intent parsing and message normalization - not for the scoring decision itself. Scoring logic was rules-based and auditable. This made it easier to debug, explain to stakeholders, and adjust thresholds without retraining.

What I Cut and Why

Multilingual support (full): Medium Impact / High Effort → deferred. Hindi intent parsing was included. Full regional language support was not justified until the core flow was validated.

CRM auto-update on disqualification: Low Impact / Medium Effort → deferred. Qualified leads got CRM entries with context. Disqualified leads were logged but not pushed to CRM - adding noise to the pipeline without clear value.

Sentiment analysis on conversations: Low Impact / High Effort → cut entirely. No downstream action was defined for sentiment data. Building instrumentation without a clear use case is waste.

Outcome

North Star: Qualified lead rate into the sales pipeline.

L1 results:

40% of L1 support and discovery queries handled without human involvement
Sales team received pre-qualified leads with context already attached
Reduction in unqualified discovery calls

What I’d do differently: Define a precision threshold before shipping. We didn’t set a formal accuracy target for disqualification before launch. In retrospect, agreeing on “we will not disqualify a lead unless confidence exceeds X%” would have given us a clearer validation criterion and made the post-launch audit more structured.