A component nobody gets to ignore.
Rope goes into cranes, mining, oil and offshore, ports, elevators, fishing and ropeways. In every one of those, the rope is the part holding the load. A question about it is never a low-stakes question.
Customer engagement · Usha Martin
A rope fails at 2am.
Somebody has to answer.
Status: live since 4 August 2026, about 25 calls a day. First results reported September 2026.
Usha Martin makes steel wire rope for cranes, mines, offshore rigs, ports and elevators. When one of those ropes is in question the call does not wait for Monday. Since 4 August 2026 that call has been routed to a voice agent that runs inside Usha Martin’s own cloud.
Industrial manufacturer · India · Inbound complaint handling, English first
Complaint lineInbound call · severity triage
SUNDAY · 02:14Site call, no desk open
ILLUSTRATIVE CALL · DESIGN INTENT, NOT A CUSTOMER RECORDING
01 Complaint intake, 24/7
02 Severity triage and callback
03 Salesforce case creation
04 RMA portal handoff
01 — The customer
Wire rope,
sold into the hard industries.
Usha Martin Limited describes itself as one of the world’s leading manufacturers of wire ropes, operating in over 75 countries, with plants in Ranchi and abroad.
Industrial accounts, not consumers.
The caller is a maintenance head, a site engineer, a procurement manager or a distributor. They know the vocabulary, they have the paperwork in front of them, and they want a name and a time, not sympathy.
Time zones that do not cooperate.
Selling into more than 75 countries means an inbound complaint can land at any hour against a support desk that keeps Indian office hours. The gap is structural, not a staffing mistake.
02 — The problem
The complaint is not the problem.
The silence after it is.
Failures do not keep office hours.
A rope is inspected when the shift allows and it is replaced when the crane stops. Neither event checks the calendar. Complaints arrive at night, on Sundays, and during Indian public holidays from customers who do not share them.
Inbound, unscheduledEvery caller says urgent.
A stopped crane and a query about a test certificate arrive on the same line in the same tone. Without a consistent first question, the order in which they get handled is decided by whoever picked up.
Severity set on the callA complaint taken by voice tends to stay in voice.
Notes get written later, or into an email, or not at all. The case in the CRM is created after the fact, with whatever the person still remembered.
Case created during the callSilence gets read as an answer.
On a safety-critical part, a customer who waits does not wait quietly. They stop the asset, they call the distributor, they open a claim, and they start pricing the competitor. None of that is recoverable by replying on Monday.
First response is the product03 — What we designed
One line, one script,
one place the case lands.
The agent is designed to do the part that is identical on every call: identify the account, classify the complaint, set severity, capture the evidence a quality engineer will ask for anyway, and put it in Salesforce before hanging up.
It is not designed to decide whether a rope is defective. That stays with the people qualified to say so.
- 01
Identify the account and the rope.
Order or invoice reference, plant, rope construction and diameter, application. The same opening questions every time, so two engineers reading two cases are reading the same fields.
- 02
Classify into ten categories.
The category list was drawn with Usha Martin’s quality team and covers what goes wrong in service and what goes wrong before the rope gets there: condition and wear in the field, specification and material mismatches, certification and documentation, and delivery and packaging.
- 03
Set severity, commit to a callback.
Severity is decided on the call against written criteria, and the caller is told the callback window before the line drops. Safety-critical cases page the on-call quality engineer immediately rather than waiting in a queue.
- 04
Create the case in Salesforce.
Account, category, severity, transcript and the captured fields are written to a Salesforce case during the conversation. The engineer opens a record, not a voicemail.
- 05
Hand material returns to the RMA portal.
Where the outcome is a return, the agent connects to Usha Martin’s existing RMA portal so the authorisation is raised in the system the warehouse already works from.
- 06
Escalate to a human, with the context attached.
Anything outside the ten categories, anything with an injury in it, and anything where the caller asks for a person goes to the QA on-call rota with the transcript and the draft case already written.
| Tier | What it covers | Callback target |
|---|---|---|
| Severity 1 | Safety exposure or a stopped asset | 2 hours |
| Severity 2 | Product performance in dispute, operation continuing | 8 hours |
| Severity 3 | Documentation, certification, packaging and administrative issues | 48 hours |
English at launch. Mandarin and Spanish are planned for the subsidiaries, on the same category list and the same severity rules.
04 — The constraints
It runs in their cloud,
because it had to.
Data residency was set before the first design session, and it decided the architecture rather than decorating it.
Indian data stays in India.
The agent runs in Usha Martin’s own AWS India account. Call audio, transcripts and case data are stored in region. Nothing about the deployment depends on a vendor tenant somewhere else.
Inside the VPC, not beside it.
The services sit in a private network inside the customer’s VPC. Salesforce and the RMA portal are reached over controlled egress, and the customer’s security team holds the keys to the account the whole thing runs in.
RDS Postgres as the record.
Case state, category mapping and the severity audit trail live in managed Postgres under the customer’s own backup and retention policy. There is no second copy of the complaint history anywhere else.
The constraint shaped the build.
Running in someone else’s account means every component has to be deployable, upgradable and auditable by their team. That ruled out anything we could not hand over, and it is why the design took the shape it did rather than the shape that would have been faster.
05 — The second track
What the field sales team
said in the meeting.
Scoped alongside the complaint line, on the same cloud footprint.
Usha Martin’s field sales engineers spend their week in plants and at ports, and the record of those conversations is whatever gets typed up afterwards. The second track records the call, summarises it, and posts the summary into MS Teams where the regional manager already reads everything else. Same principle as the complaint line: capture the thing while it is still accurate, and put it where somebody will actually see it.
If your own field team is the only place your customer intelligence lives, the pattern is the same one described in AI agents for manufacturers.
06 — What we measure
First numbers are in.
Measured against the same list.
| Measure | How it is read | Industry benchmark | Reported |
|---|---|---|---|
| Complaints captured outside office hours | Calls the agent took that previously had nowhere to land | 47% of US business calls land after hours or on weekends, across a general call-log sample, not complaint lines specifically. AnswerConnect, 2026 | 48% of calls |
| Severity agreement | How often the quality engineer keeps the severity the agent set | No published benchmark for AI/human severity agreement in an industrial complaint setting. | 88% kept |
| Category accuracy | How often the case stays in the category it was filed under | Up to 93% category-classification accuracy in an academic multilingual logistics-support benchmark, lower on noisy real-world queries. A 13-way taxonomy, not wire-rope complaints. arXiv, 2026 | 82% |
| Cases created without human retyping | Salesforce cases complete enough to work from as filed | No published benchmark for case-completeness at creation. | 95% of cases |
| Escalation rate | Share of calls handed to the QA on-call rota | 76% average resolution rate across 12,000+ customers on a chat AI agent, meaning roughly a quarter hand off to a human. SaaS chat support, not voice or B2B manufacturing. Fin AI (Intercom), 2026 | 18% of calls |
The engagement was approved in June 2026 and went live on 4 August 2026, handling about 25 calls a day. The Reported figures above are Usha Martin’s first results, as reported in September 2026. Benchmark figures are third-party industry data, linked at the source, not projections for this deployment.
07 — In their own words
Usha Martin,
on camera.
Your turn
Bring us your after-hours line.
Tell us what happens today when a customer calls your complaint line at midnight, and which system the case is supposed to land in. We will scope one workflow, price it, and say so if the maths does not work.
Book a discovery callA short qualification form first. A call if there is a fit.
Prefer email? hello@voxdonna.com