New B2B buyers often sign up on a marketplace and never place a first order. Email and SMS nudges get ignored; human sales teams cannot call every signup at marketplace scale. From August 2026 I have been building an outbound multilingual AI voice sales agent that places personalised sales calls in 5 languages - Hindi, English, Hinglish, Marathi and Gujarati - grounded in a live offers knowledge base, with callback scheduling and WhatsApp follow-up after the call. This case study is the architecture and the decisions - not a vendor pitch.

Why does outbound voice fail after marketplace signup?

Most post-signup outreach fails because it is generic, delayed, or invents commercial details the buyer can immediately distrust. On a B2B marketplace the buyer already showed intent by registering. The job of the first call is narrow: confirm need, present a real offer that matches that need, and either start a first purchase or leave a clear next step.

A free-form chatty agent fails in three predictable ways. It pitches the wrong category. It invents a discount or MOQ that is not on the live catalogue. It drops the thread when the buyer says “call me later” or asks for the offer on WhatsApp. Fixing those is harness work, not a better system prompt alone - the same lesson as in agent harness engineering.

This build sits alongside the multi-client pattern in voice AI lead generation platform architecture and the inbound clinic pattern in voice AI appointment booking. Those posts cover intake, CRM and booking. This one covers outbound sales after marketplace signup.

How do you structure a multilingual sales conversation as a state machine?

Model the call as ordered phases with allowed tools and explicit exits, not as one long generative dialogue. Each phase has a goal. The agent may only call tools that belong to that phase. Leaving a phase requires a clear signal: identity confirmed, need captured, offer grounded, objection resolved, or next step booked.

PhaseGoalTypical tools
Open & identifyConfirm who answered and why the marketplace is callingBuyer profile lookup; language preference
Discover needLearn category interest, volume and urgency without pitching yetCRM notes; prior browse or signup context
Grounded offerPresent only live, retrieved offers that match the needOffers knowledge-base search; price and availability tools
Handle objectionsAnswer logistics, MOQ and payment questions from factsPolicy KB; FAQ retrieval; human escalate
Close or continueTake a first order intent, schedule a callback, or hand offCallback scheduler; WhatsApp send; CRM update

State machines keep interruptions recoverable. If a buyer jumps from discovery to price, the harness routes to the grounded-offer phase instead of improvising. If ASR drops a turn, the agent resumes from the last completed phase rather than restarting the pitch. That is the same control-flow instinct I use with LangGraph on non-voice agents, applied to a live call.

How do you stop the agent inventing offers or prices?

Every commercial figure the agent speaks must come from a tool result retrieved in that turn - never from model memory. The offers knowledge base is the system of record for what can be sold right now. Before the agent pitches, it searches that knowledge base with the buyer’s need. The retrieved rows - SKU, price, availability, terms - enter the turn context. The prompt instructs the model to cite only those rows.

Guardrails sit after generation. If the draft utterance contains a number or discount that is not present in the retrieved context, the harness blocks or rewrites it and forces another retrieval. Invented figures are a trust failure on a marketplace: one wrong price can burn the brand faster than a missed call. Treat invented offers like irreversible cloud actions - constrain them in the harness, as described in the production agent harness guide.

Tool calling also covers what the agent should not invent: shipping windows, payment terms and MOQ. Those live in policy or offer metadata. If retrieval returns nothing, the honest move is to say so and escalate or schedule a human callback - not to approximate.

How do you handle Hinglish and code-switching?

Design for five languages as first-class paths, and treat Hinglish as a mode of speech rather than a translation error. Buyers mix Hindi and English mid-sentence when talking about products, payments and logistics. Forcing a single language per call sounds unnatural and raises hang-ups. The agent mirrors the buyer’s language on each turn while keeping offer facts language-agnostic in the knowledge base.

LanguageHow it shows up on the call
HindiPrimary for many north and central India buyers; formal and informal registers both appear.
EnglishCommon for B2B terms, SKUs and payment language even when the rest of the call is Hindi.
HinglishCode-switching mid-sentence is normal; the agent must not force a single language per turn.
MarathiRegional preference for Maharashtra-facing buyers; same offer grounding rules apply.
GujaratiRegional preference for Gujarat-facing buyers; keep product facts language-agnostic in the KB.

Prompting alone is not enough. Evaluate transcripts for language fidelity and for whether offers stayed grounded. A call that sounds fluent in Hinglish but invents a price is still a failed call. A slightly stiff call that quotes a live offer correctly is a win you can improve with better phrasing later.

What happens after the call - callback and WhatsApp?

The hang-up is not the end of the sales loop. Callbacks and WhatsApp carry the same grounded offer forward. When a buyer asks to continue later, the agent schedules a callback through a tool instead of promising vaguely. When a buyer wants the offer in writing, the harness sends a WhatsApp follow-up with the same retrieved offer facts the agent just spoke - not a new generative summary that can drift.

CRM updates close the loop for human sellers: who was called, which phase completed, which offer was presented, and what next step was booked. That record is also how you tune prompts with evidence. The multi-channel pattern - voice plus messaging plus a system of record - is the same shape as the lead-generation platform, documented in voice AI lead generation architecture, applied here to first-purchase sales.

What is in the stack, and what stays out of the model?

The production stack is voice AI, tool calling, a live offers knowledge base and prompt engineering for sales phases - with CRM and messaging tools for callbacks and WhatsApp. Telephony and STT/TTS vendors can be swapped; the harness boundaries should not. Buyer PII and credentials never belong in the model context beyond what the current phase needs. Offer search returns commercial facts for the turn; it does not dump the whole catalogue into the prompt.

Prompt engineering here means phase-specific instructions: how to open, how to ask for need, how to present a retrieved offer, how to refuse unknown prices, and how to hand off. Those instructions are versioned with the harness, not edited live in a vendor dashboard without review. When a transcript shows a failure - wrong language mirror, skipped grounding, missed callback - the fix lands in the phase prompt or the tool schema, then gets re-tested on a small golden set of call scenarios.

What would you reuse from a voice sales agent build?

Reuse the state machine, offer grounding and post-call tools on any outbound sales voice agent. Those layers are product-agnostic. Redesign the phase goals and CRM fields for each marketplace or catalogue. Do not copy a clinic booking harness wholesale into sales - inbound appointment booking optimises for calendar slots; this agent optimises for a first order. The shared DNA is the harness: typed tools, grounded context, guardrails and tracing.

I would also reuse the evaluation split: language fidelity separate from commercial correctness. Teams that only score “did the call sound natural?” ship fluent agents that invent discounts. Teams that only score “was the price right?” ship stiff agents buyers hang up on. You need both scores, and you need them on multilingual transcripts - including Hinglish - not only on clean English test lines.

For the broader portfolio of systems behind this work, see the homepage project case studies. For how I think about tools, permissions and verification around any agent, start with agent harness engineering.

Frequently asked questions

How do you build a multilingual voice agent that handles Hinglish?

Assume code-switching is normal. Keep product facts and prices in a language-agnostic knowledge base, instruct the model to mirror the buyer's language turn by turn, and evaluate transcripts for language fidelity - not only for whether an offer was correct.

How do you stop a voice sales agent inventing offers or prices?

Never let the model speak a number that did not come from a tool. Retrieve live offers into the turn context, require tool-backed claims for prices and availability, and add a guardrail that blocks or rewrites utterances that introduce figures absent from retrieval.

Is an outbound voice sales agent the same as a lead-generation voice platform?

No. Lead-generation platforms often qualify and book meetings across many clients. This sales agent drives a first purchase for new B2B buyers on one marketplace, with offer grounding and post-call WhatsApp as first-class features. The harness ideas overlap; the business loop does not.

What stack do you use for production voice sales agents?

Voice AI with tool calling, a live offers knowledge base, prompt engineering for multilingual sales phases, and CRM or messaging tools for callbacks and WhatsApp. The exact telephony vendor can change; the state machine and grounding rules should not.

When should a human take over the call?

When the buyer asks for custom pricing outside the catalogue, disputes an order, or the agent cannot ground an answer after retrieval. Escalate with full transcript context rather than restarting the conversation cold.

References

Related: voice AI appointment booking harness, voice AI lead generation architecture, agent harness engineering.