A sales call that ends inside the CRM.
An electric two-wheeler maker sells through a dealer network, and most of its enquiries arrive by phone. Settl's voice agent now takes those calls in English, Hindi and Hinglish, asks what a good salesperson asks, and hands the company's own Salesforce a scored lead, a dealer-routed opportunity and a booked test ride — with the rep's coaching already on the record. This is what was built, what we found on the way in, and what we measured.
Enquiries by phone, and a CRM the team had stopped trusting.
The company makes electric scooters and sells them through showrooms across Indian cities. Enquiries come in by phone: a rider comparing three models, a daily run of twenty-five kilometres, a plug in the building's parking or not, a question about EMI. Every call was a rep's time, and what the rep learnt reached the CRM only if they typed it in afterwards — usually a line or two, often nothing.
The company already had a Salesforce org built for exactly this funnel: lead scoring rules, a qualification threshold, dealer assignment by pincode, city, state and territory, and a nine-stage pipeline from enquiry to delivery. On paper, a lead that scored high enough would convert and route itself to the nearest dealer with room. In practice the sales team had stopped believing the routing, and once we were inside the org we found out why.
The agent asks the discovery questions, then speaks CRM.
Settl's voice agent answers the sales line. It asks how the scooter will be used, how far, where it will charge, what it is being compared against and when the buyer plans to decide, and it books a doorstep test ride before the call ends. It answers in whichever language the caller uses: a question in Hindi gets a Hinglish answer, not a transfer.
What the caller said is translated into the CRM's own vocabulary before anything is written. "One lakh twenty" becomes a budget of 120,000. "Before the end of this month" becomes a purchase timeline of Within 30 Days. "Can I do EMI?" sets financing interest. The lead is written once per conversation, so a second sync enriches it rather than duplicating it, and a lead that has already converted enriches the opportunity instead. Only values the call actually produced are sent, so nothing a rep typed by hand is ever blanked.
The org's own Apex decides the score, the threshold and the dealer.
Settl did not reimplement the company's sales rules. Lead scoring, the qualification threshold, conversion and dealer routing already lived in the org as Apex, exposed as invocable actions, so the agent calls them. A lead created by the AI and one a salesperson keys in by hand go down the same code path, and sales ops can move the threshold in custom metadata without a Settl deploy.
Routing is evaluated most-specific first: a pincode rule, then city, then state, then territory. A dealer at capacity is skipped and the lead falls through to the next rule, with capacity counted live at assignment time. The connection is an OAuth JWT bearer against a Connected App: no password, no security token, and no AI key stored anywhere in the org.
Seven defects in the org, fixed on the way in.
Wiring a live pipeline into an org is the most thorough test it ever gets. The rules that were supposed to route leads had never matched a record, and the capacity that was supposed to protect dealers was never enforced. We brought the org's metadata into version control — until then it lived only in the org, one accidental overwrite from gone — and fixed what we found:
- Territory, state and city rules could never match, so every lead fell to the default dealer.
- Dealer capacity was queried but never enforced; a full showroom kept receiving leads.
- The open-opportunities counter on each dealer was never written to.
- 33 fields across four objects were invisible to every profile — an audit trail nobody could read.
- A stale picklist value on purchase timeline, and a duplicate Pune dealer splitting that city's leads.
The rep opens the record and the first call is already prepared.
After every conversation the model writes a summary and the discussion points, the single most useful next step, a talk track with an answer for each objection the caller raised, a follow-up message ready to send on WhatsApp, a win likelihood with its rationale, and the risks that could lose the deal. These are stored as ordinary fields on the lead and copied to the opportunity at conversion, so they survive a change of model and appear in the same cards a rep already uses.
The rule score and the model's likelihood are allowed to disagree, and the card says so when they diverge. One prospect scored 45 on the rules, below the threshold, because no configured signal could see that their apartment-charging objection had been resolved mid-call. The model read the call and put the chance at 65%. Both numbers stayed on the record, and the rep rang.
The same funnel on Postgres, for a dealer network that won't pay per seat.
Salesforce is licensed per seat, and a dealer network with thirty showroom staff pays that before a single lead is worked. So the integration was built as a port with two adapters. Settl's own CRM implements the same objects, scoring rules, routing and nine-stage funnel on the deployment's own Postgres, and switching is one environment variable. Same conversation, same extraction, same coaching: the same lead, dealer and test ride land in either.
It also carries what the org never had — field history on every object, an error log that is actually written, calls and tasks on a timeline — and what a showroom sells on: a per-unit deal price with a floor no rep can go under, and the registration paperwork the funnel really ends at, from VIN and chassis to RTO status, number plate, insurance and invoice.
Speech-to-speech on the sales line
About 0.9 to 1.2 seconds from the caller stopping to the agent speaking, measured with a live lookup in the middle of the call. A human can join the same call when the caller asks for one.
The org's rules, called by the agent
JWT bearer auth, one write per conversation, Apex scoring and routing untouched. No Agentforce licence: the AI runs on Settl's side and the org stores the result.
The same funnel without the licence
Fifteen tables on Postgres, scoring and routing rules edited in Setup, list views, record pages with a stage Path, reports and field history.
Frequently asked
Who is the customer?
An Indian electric two-wheeler maker with a dealer network, working with Settl as a design partner. They have not asked to be named, so this page describes the deployment without naming them. Ask on a demo call and we will say what we can.
Are the numbers real?
The measurements are. The latency was measured on the deployed voice worker, the seven defects are the ones fixed in the org, and the 45-versus-65 example is a real record. We do not publish the company's sales figures, so there is no conversion-rate claim on this page.
Did it need Salesforce Agentforce?
No. Agentforce is not enabled in the org and nothing waits on it. The model that ran the call already holds the transcript and the grounding; it writes its coaching into ordinary fields the org stores and shows. If the company switches Agentforce on later, it can read the same fields.
Would this work with our CRM?
The agent talks to a port in the sales team's vocabulary — interest, timeline, test ride — and an adapter translates. Salesforce and Settl's own CRM exist today; another CRM is an adapter, not a rewrite.
What happens to a call the agent can't handle?
A caller who asks for a person gets one: an available agent joins the same call and the AI steps aside. If nobody is free, a ticket with the transcript opens in the inbox, and the lead still syncs with everything learnt so far.