By Sagar Shankaran, Founder of CallSphere
What a device maker loses to unanswered support calls, and what an instant-answer voice line changes about complaint records, RMAs and service parts revenue.
Key takeaways
You tried this in 2024. The answering service picked up after four rings, read a script, spelled the hospital's name wrong, and emailed a form at 9 a.m. your complaint specialist had to phone back on anyway. It made more work than it saved and you cancelled it in five months. Fair. Here is what is actually different now, and why it matters more to a device maker than to almost any other shop.
Your plant is in Anaheim. A clinical engineer at a Boston hospital starts at 6:10 a.m. Eastern and finds one of your reusable instrument washers throwing an error on the pre-shift check, first case at 7:30. He calls your support line. It is 3:10 a.m. in Anaheim. He gets a greeting and a beep.
He does not leave a message. He calls the regional service company, gets a technician on-site by nine, and pays for a part your service contract would have covered. You never hear about it.
That is the sales loss, and it is the smaller half. A complaint, in a device plant, is not a bad review — it is any communication alleging deficiencies in the identity, quality, durability, reliability, usability, safety, effectiveness or performance of a device after it is released for distribution, and the obligation to record it starts the moment someone at your company hears it. A voicemail nobody returns is not a complaint you avoided. It is a complaint that reached your company and did not get a record.
Pull your phone system's report for last quarter. The pattern is predictable: a meaningful slice of support calls land outside staffed hours, and most of those callers leave no message. Hospital sterile processing runs overnight. Biomed departments start at six. European distributors call during their afternoon, which is your night.
Counted in money, the loss is parts orders, service bookings and return authorisations that went to a third-party refurbisher. Counted in regulatory terms, it is complaint records never opened — and complaint handling and CAPA are exactly where a QSIT inspection starts. An investigator who finds a voicemail box of unreturned messages beside a complaint log with no matching entries has a Form 483 observation without trying.
Add the seasonal shape. Elective surgical volume climbs hard in the fourth quarter as patients hit their deductibles, sterile processing runs longer, and calls about reprocessing wear and cycle faults come with it. Your quietest staffing quarter and your busiest call quarter are the same quarter.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for healthcare in your browser — 60 seconds, no signup.
Here is the change. Speech-to-speech systems now answer in roughly two-tenths of a second — faster than an average person's reaction time — in one pass, without the old chain of transcribe, think, then read aloud. And because looking something up or booking is built into the same conversation, the agent can check a serial number or put a service visit on the calendar mid-sentence, with no gap the caller can hear.
Google's Gemini 3.1 Flash Live arrived in March 2026 with audio-to-audio across more than ninety languages and strong performance on that kind of lookup, and OpenAI's realtime line pushed the same capability into ordinary business phone systems over standard telephone lines. That last part is the boring bit that made it usable: it attaches to the number printed on your instructions for use, not a web widget.
The 2024 version failed at the two things that matter to you: it mangled spelled-out alphanumeric lot numbers, and it could not check anything while the caller was still talking. Both are now routine.
flowchart TD
A["Biomed calls 6:10 a.m. Eastern, washer error code"] --> B["Agent answers on the first ring, states it is an AI assistant"]
B --> C["Captures facility, model, serial, lot, date and description"]
C --> D{"Patient involved or procedure cancelled?"}
D -->|Yes| E["Pages the on-call RA lead, stays on the line"]
D -->|No| F{"Unit under service contract?"}
F -->|Yes| G["Books the field service visit, issues an RMA number"]
F -->|No| H["Quotes the service call, emails the parts list"]
E --> I["Complaint record opened in the eQMS before 7 a.m."]
G --> I
H --> I
Your complaint form already exists, and it is the specification for what the agent must collect: reporter name, facility, callback number and email; device trade name and catalogue or model number; serial or lot number and the device identifier portion of the UDI; date the event occurred and date the reporter became aware; the event in the reporter's own words, not summarised; whether a patient was involved; whether clinical intervention was required; whether the device is being returned; and any existing case number.
A voice agent can reliably get all of that, and one of them better than a tired person at 3 a.m.: the description in the reporter's own words, recorded and transcribed rather than paraphrased into a form field. That matters later, when the reportability decision is made from the file.
What it must never do is characterise the event. "That sounds like the seal failure we have seen before" is an admission on a recorded line, and it will be read back to you. The agent takes the report. It does not classify it.
Assumptions, illustrative, for a mid-size maker of reprocessing and surgical instrument equipment: 240 inbound calls a month to the technical support and complaint line, staffed 8 a.m. to 5 p.m. Pacific.
| Line | Value |
|---|---|
| Inbound support calls per month | 240 |
| Arriving outside staffed hours (27%) | 65 |
| Of those, abandoned without voicemail (58%) | 38 |
| Captured by an instant-answer line (90% of 38) | 34 per month |
| Of the 34: service parts orders | 8, averaging $410 = $3,280 |
| Of the 34: complaint reports that now get a record | 11 per month, 132 per year |
| Of the 34: return authorisations and service bookings | 6 |
| Parts revenue recovered per year | $39,360 |
| Cost of the answering line at $600/month | -$7,200 |
| Net, before counting a single regulatory benefit | $32,160 |
The 132 complaint records a year are not in that total, because there is no honest way to price them until the year one is missing. Then the price is a Form 483 observation in the complaint handling subsystem, a commitment letter, a remediation project and a follow-up inspection — which every quality manager can cost out from experience, and which dwarfs the parts revenue.
One more line worth measuring separately: daytime abandonment. One support person cannot hold three lines at 10:30 on a Monday, and the calls that drop on hold are the same calls, arriving during business hours.
Write these into the escalation rules before it goes live, and test each by calling in yourself.
Still reading? Stop comparing — try CallSphere live.
See the healthcare AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
Anything involving patient harm. If the caller says a patient was injured, a procedure was cancelled, or the device failed during use, the agent takes the identifying details, says a person will call within a stated time, and pages the on-call regulatory affairs lead. Not a queue. A page.
The reportability decision. Whether an event is a 30-day report, a five-day report or not reportable is a trained judgement made against your written decision tree by a named person, and it stays that way. The agent's contribution is that the clock starts against a properly dated record rather than a voicemail nobody played until Thursday.
Any instruction to stop using a device. Telling a hospital to quarantine an instrument is a field action decision. It comes from your quality organisation, in writing.
The angry caller. A surgical services director whose case was cancelled does not want an efficient intake. Route on tone, get a person on within minutes, and accept that as a cost of doing this well.
Yes, and that is the point. Awareness sits with the company, not an individual employee, so a call that reached your line and was never returned already counted against you — you simply had no record of it. An agent that opens a dated complaint record at 6:12 a.m. puts you in a better position than a voicemail box, not a worse one. Make sure the record's creation time is the call time and that your procedure says so.
Disclose it in the first sentence, every call, and keep the recording. Several states now have AI statutes on the books — Texas TRAIGA and California SB 53 both took effect on 1 January 2026, with Colorado, New York, Utah, Nevada, Maine and Illinois running their own, and federal preemption is still unsettled. If you take calls from EU distributors, the EU AI Act's transparency obligations carry a 2 August 2026 compliance date and can reach a US company whose systems affect users there. Disclosure is cheap and removes the argument.
This is the right thing to be sceptical about, and it is testable in an afternoon. Call your own line twenty times with real lot and serial formats from your labels, including the ones with a leading zero and the ones where a letter O sits next to a zero. Require the agent to read the value back and get confirmation before it writes anything. If it cannot pass your own twenty-call test, do not go live — and never accept a demonstration on someone else's part numbers.
Speech-to-speech across ninety-plus languages works well enough for intake now. The judgement is what you do with the record: a complaint from an EU customer can pull in your notified body and vigilance reporting obligations, a different route than a domestic call. Have the agent tag country of origin and route non-US complaints to whoever owns your EU technical documentation.
Do not start by replacing your daytime support line. Point the after-hours overflow at it for one quarter, leave the daytime staff alone, and compare two numbers against the same quarter last year: complaint records opened, and abandoned calls. Those two make the case or kill it.
CallSphere builds the voice and chat agents that do this — answering around the clock, capturing facility, model, serial and lot accurately, booking the service visit, and handing anything alleging a device problem to your on-call quality contact with a dated record behind it. What it does not do, and should not, is decide whether an event is reportable. That has your quality manager's name on it.

Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Sixty calls a day, one person at the counter. See what unanswered banner and yard-sign calls cost a print shop, and what an instant-answer voice agent changes.
Luxury buyers in Scottsdale call nights and weekends, and the first team to answer usually wins. How AI voice and chat agents keep Arizona brokerages first.
Late-night lockouts, dead AC, and towed cars define after-hours calls at Key West, FL rentals. An AI night desk fixes most 2 a.m. calls without waking you.
AI now drafts the CAPA investigation. What a medical device plant should teach a new quality engineer in week one, and what stopped being a job requirement.
Community bank calls that hit voicemail at noon cost real deposits. A 2026 voice agent answers in 200ms, reads holds and payoffs, and hands the rest to a banker.
Discharge planners call three suppliers until one answers. What DME loses to abandoned calls, and what an instant-answer line changes about that number.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI