In many Indian businesses a buyer's day includes the same call made again and again: "Has the material for this order been dispatched? When will it reach? How many?" Email goes unanswered and a portal goes unused, but a supplier almost always picks up the phone. A voice agent that makes those calls in Hindi, records the answers against the purchase order and hands anything unusual to the buyer is one of the applied AI services we are building.
It is in build. It does not run at a client, and we have no results to report. This piece sets out what we are designing for, because the decisions are the same for anyone considering a voice agent, whether they build it with us or not.
Our starting position is that bounded scripts and an exception queue beat a clever assistant. A supplier follow-up call has a narrow job. Treating it as an open conversation adds risk without adding value.
What the call is for
The first version of the call has three questions, tied to one open purchase order:
- Has the material been dispatched, or when will it be?
- What quantity, and is it the full order?
- If dispatched, what are the vehicle or transporter details?
Everything else the supplier might raise (a price change, a quality complaint, a request for an advance) is out of scope for the agent. It says a person will call back and records the request. Price is never discussed by the model.
Narrowing the call this way makes it testable. Each call has a known set of answers to capture, so each call can be marked as complete, partial or failed, and that is the metric we would agree with a client before building: calls confirmed per day, measured against the baseline of a buyer making them by hand.
Hindi, and how suppliers actually speak
"Hindi" undersells the speech problem. A supplier in Ludhiana or Surat will mix Hindi with English and often with a regional language in the same sentence. Order numbers, quantities and units are frequently said in English. Dates come as "parson", "Monday tak" or "next week". Item names are whatever the two companies have called them for years.
That shapes the speech recognition (ASR) choice:
- Code-mixed speech must be the test, not an edge case. We evaluate recognition on recordings of real supplier calls, gathered with consent, not on clean read-aloud Hindi.
- Telephone audio is narrow and noisy. Calls come from shop floors, trucks and roadsides on mobile networks. Models are tested on that audio.
- Numbers and names need help. The agent knows the order it is calling about, so it can check a heard quantity or item against the order lines and confirm, rather than trusting a raw transcript.
- Read back before writing. "So, 400 kilos, dispatched on Thursday, vehicle number ending 2231. Is that right?" A confirmation turn costs a few seconds and catches most recognition errors.
Speech synthesis (TTS) matters as much. A voice that sounds robotic or speaks formal, textbook Hindi gets hung up on. The agent should use the register a buyer would: polite, plain, conversational Hindi, with English words where people use English words. It also needs to handle interruptions, because people talk over a caller, and to keep pauses short, because long silences on a phone line read as a dropped call.
Consent and the rules around automated calls
Three commitments are in the design from the start.
- The agent says what it is. Every call opens by naming the buying company and stating that this is an automated call about a specific order. It does not pretend to be a person.
- Recording is disclosed. If calls are recorded and transcribed (and for review they need to be), the supplier is told at the start, and the recording and transcript are handled as personal data under the Digital Personal Data Protection Act, with a stated retention period and access limited to the people who need it.
- Calls only to known contacts, at agreed times. The agent calls the contact recorded against the supplier in the business system, at times the supplier has agreed, and stops calling anyone who asks it to. Telecom rules on automated and commercial calling apply and are checked with the client's counsel for each deployment.
The data stays in the client's tenancy. Calls to hosted speech or language models go through our gateway to a provider the client approves, under terms that exclude training on their data.
Writing the answer back to the purchase order
A call is only useful if its result lands where the buyer works. In ERPNext the natural home is the purchase order itself.
- Each call becomes a call record linked to the purchase order, with the time, the outcome, the answers captured, a confidence for each answer, and the transcript and recording.
- Confirmed answers update dedicated fields on the order or its lines, such as a supplier-confirmed dispatch date and quantity. The agent does not overwrite the required-by dates the buyer set; the difference between the two is the information the buyer needs.
- When the confirmed date or quantity differs from the order, the record goes to the buyer as an exception rather than being applied quietly.
- Dispatch details, once confirmed, are ready for the goods receipt, so the store knows what is coming.
Nothing financial happens because of a call. The agent does not amend an order, cancel a line or change a price. Those remain decisions for a person, recorded against the person's name.
The exception queue is the product
Most of the design effort goes into what happens when a call does not go to plan, because that is where a buyer's time is saved or wasted. The calls that should reach a person include:
| Situation | What the agent does |
|---|---|
| No answer, or switched off | Retries at the agreed interval, then flags the order for the buyer |
| Wrong person answers | Asks for the right contact, records it as a suggestion, never updates the master on its own |
| Dispatch slipped or quantity short | Captures the new date or quantity, reads it back, sends it as an exception |
| Supplier raises anything out of scope | Says a person will call back, records the request |
| Low confidence on any answer | Marks the answer unconfirmed and queues the call for review |
| Supplier is unhappy or asks for a person | Ends politely and hands over at once |
A buyer starts the day with a list: orders confirmed, orders with exceptions, orders not reached. They call only the second and third groups. That is the time saving we would set out to measure.
How we would test it before it calls anyone
The governance we apply to all applied AI work holds here:
- An eval set of real calls. Recorded supplier calls, transcribed and labelled with the correct answers, that the agent must handle before release and after every change to a model or a script.
- A cost budget per call. Agreed in advance and monitored; if a call would exceed it, the order goes to the manual queue instead.
- A named approver. One role owns the exception queue and the decision to widen the agent's scope.
- Honest labels. Until a client runs it in production and we have measured the result, we describe the expected effect and the method, not a number.
If this is your situation
If your buyers spend their mornings on follow-up calls, or you are weighing a voice agent for supplier confirmations, payment reminders or delivery scheduling, read how AI fits inside ERPNext and business workflows and the applied AI services we are building, then tell us about the calls your team makes.