By Sagar Shankaran, Founder of CallSphere
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.
Key takeaways
At 7:05 on a Monday the new quality engineer is in the gowning room with a shift lead watching her put the bouffant on before the mask, and the mask before the coverall, and who will make her repeat the sequence twice more this week before she is signed off to enter the Class 7 room alone. That part of onboarding has not changed since 2019. What has changed is what she will be doing on Thursday afternoon.
Until about eighteen months ago, a new quality engineer at a 40-person device plant spent her first quarter learning where things live. Which folder in MasterControl holds the current revision of WI-0412 and which holds the superseded set nobody deleted. Which certificates of analysis are filed under the part number and which under the purchase order. That knowledge was the job, and when the person holding it left, it left with them.
A quality engineer's job in a 2026 device plant is no longer to write the first draft of a corrective and preventive action investigation. It is to decide whether the draft in front of them is true, and to put their name on that decision.
This lands harder in device manufacturing than most trades, because the Quality Management System Regulation took effect on 2 February 2026, folding ISO 13485:2016 into what used to be 21 CFR Part 820. Every shop that spent 2025 rewriting procedures against the new clause numbering now has a document set where half the cross-references were touched this year. A new hire cannot absorb that map by osmosis — the map itself is eighteen months old.
Take a contract manufacturer of injection-moulded and machined single-use components: 40 people, one full-time quality engineer, a quality manager half in the department and half in front of customers, a document control specialist, two receiving inspectors. In a normal week that department disposes of 22 nonconformance reports, carries nine to twelve open CAPAs, pushes four or five engineering change orders through Arena, and reviews device history records for every lot that ships.
An honest audit of those hours says most of it was assembly, not judgement. Pulling the last six NCRs on the same part number to see whether the trend was real. Comparing revision C of a work instruction against revision D by eye, because the redline skips table changes. Typing the same containment paragraph into the CAPA form for the eleventh time. Chasing training records for the operators who ran the affected lot. Nobody calls that skilled work, but you had to hire a skilled person to do it — it required knowing which cabinet to open.
Here is the change. US small-business adoption of these tools reached 66% this year, up from 55% a year earlier. In the same breath, about 70% of owners say their people need more training to use them effectively. The shortage is no longer software: every eQMS vendor a device shop already pays — MasterControl, Greenlight Guru, ETQ Reliance, Veeva — put drafting and summarising into the product you already license. The shortage is people who know when to reject what comes back.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for healthcare in your browser — 60 seconds, no signup.
That is a hiring and training problem, not a purchasing problem — and it is why the old job posting is now wrong.
flowchart TD
A["New QE starts Monday, gowning qualification"] --> B["Reads current rev of WI-0412 in the eQMS"]
B --> C["Assistant drafts CAPA investigation from NCR history"]
C --> D{"Does the draft match the DHR on the floor?"}
D -->|No| E["QE rejects draft, pulls the lot travelers"]
D -->|Yes| F["QE edits, cites evidence, signs under Part 11"]
E --> F
F --> G["Quality manager approves, training record updated"]
Day one, before anything else: the signature rule. An electronic signature under 21 CFR Part 11 is a legal act attributable to one human being. Logins are never shared, and nothing drafted by software gets signed until the signer has personally looked at the underlying record. Say it aloud and put it in the training record.
Day one, second hour: where the truth lives. The device master record is the truth. The traveler is the evidence. The drafted paragraph is a claim. Walk them to the floor with a completed device history record and have them find three fields a drafting tool could not have known.
Day two: how to reject. Give them a deliberately wrong draft — a containment statement naming the wrong lot range, an investigation citing a work instruction revision not effective on the build date. Their job is to find it. Do it three times in the first fortnight. A quality engineer who has never rejected a draft will not reject one in month six.
Day three: good documentation practices, unchanged. Single line through the error, initial, date, no correction fluid, no back-dating. Then show them the validation record for the assistant itself — under ISO 13485 clause 4.1.6, software used in the quality system is validated for its intended use, so the tool sits inside the system.
The document control specialist stops retyping approved change orders into the eQMS and re-linking affected work instructions — six to ten hours a week. The role moves to owning effectivity dates, the training assignments that fire on a revision, and the periodic review calendar, where a missed date is an audit finding.
The quality engineer stops writing first drafts and does more investigations — not the same number faster, more of them, because the ones that used to close as "operator error, retrained" now get a real look at the fixture, the mould tool or the incoming material.
The receiving inspector stops keying certificate-of-analysis values into a spreadsheet and spends that time on dimensional checks against the print's critical-to-quality features.
What stopped being a hiring requirement: eight years of writing CAPA narratives. What became one: the willingness to tell a confident-sounding system it is wrong. Hire for that and you can recruit from adjacent regulated manufacturing — food, aerospace, pharma packaging — instead of waiting nine months for a device candidate in the Twin Cities or Warsaw, Indiana, while the large OEMs outbid you.
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.
Assumptions, illustrative, for a 40-person shop. The department is 1.5 quality engineering headcount. Investigation writing, evidence pulls and revision comparisons take 44 hours a week combined. Roughly 60% of that is assembly rather than judgement, and the assistant handles 80% of it with a human reviewing everything.
| Line | Value |
|---|---|
| Combined QE hours per week on writing and evidence assembly | 44 |
| Portion that is assembly, not judgement (60%) | 26.4 hrs |
| Handled by the assistant, human-reviewed (80% of that) | 21.1 hrs |
| Working weeks per year | 46 |
| Hours returned per year | 971 |
| Loaded QE cost ($118,000 over 1,840 productive hours) | $64/hr |
| Value of returned hours | $62,100 |
| Assistant seats, 4 at $60/month | -$2,880 |
| Validation and procedure update, 60 hours at $64 | -$3,840 |
| First-year net | $55,380 |
That is roughly half the second quality engineer you were going to post in Q3. It does not cancel the hire if you are growing; it moves the date, and changes who you can afford, because you are hiring for judgement rather than typing speed.
The reportability decision under 21 CFR Part 803 — 30-day report, five-day report, or neither — is a trained judgement made against your written decision tree by a named person. Do not let a summary make it and do not let a summary phrase it.
Root cause is not a writing problem. A draft can tell you four of the last six failures came off the same cavity. Only a person standing at the press with the tool-room lead can tell you the cavity has a worn vent. If investigations start reading fluently and closing fast without anyone going to the floor, your quality system got worse, and it will show at the next ISO 13485 surveillance audit.
Supplier decisions — disqualifying a plater, accepting a deviation on a raw material lot — stay in the material review board. And training effectiveness is not a read-and-understand click. Watch the operator build the part.
Nothing in the Quality Management System Regulation cares what produced the text. An investigator cares whether the conclusion is supported by objective evidence, whether the signer was qualified, and whether the electronic signature meets Part 11. The real exposure is different: a fluent, well-organised, wrong investigation is easier to sign and harder to spot.
Yes — treat it like any other quality system tool. If a quality engineer uses it to prepare records, that is a competency with a training requirement, a record and a refresh. Auditors have started asking. Write a one-page work instruction covering what the tool may draft, what it may never do, and how human review is evidenced.
More easily than in 2024, but not blindly. The document-navigation and drafting load has dropped, which was the slowest part to teach. Device-specific judgement — complaint versus service request, when a nonconformance becomes a CAPA, what a notified body wants on a technical file — still takes months. Pair the new hire with your quality manager on reportability decisions for two quarters.
Tell them early and give them the better half of the job. Effectivity dates, training assignment triggers, the periodic review calendar and the internal audit schedule are things a shop your size does badly and an experienced document control person can own outright. That is a promotion — say so before the rumour version reaches the floor.
One last thing while you are rebuilding the department: look at the phone. In most small device plants the support line lands on whoever is nearest, which on a Tuesday is often the person you just hired. CallSphere builds voice and chat agents that answer the line around the clock, capture the caller's facility, device model and lot number properly, and book the callback with the right person — so a hospital biomed at 6 a.m. gets a real answer and your new quality engineer keeps her afternoon. The regulatory judgement stays where it belongs: with the person who signs.

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.
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.
Adoption is 66% and 70% of owners say staff need training. Which four district-office seats change shape, week-one onboarding, and what leaves the job posting.
Splitting an extent-of-condition review across several AI agents turns three weeks of record reading into three days, and bounds the field action before day 22.
Which property management roles change shape when AI drafts the abstract and the recovery schedule, what to teach in week one, and what leaves the job posting.
ITAR prints, customer mold designs and 21 CFR Part 820 records cannot leave your plant. What a local AI box in the tool crib actually does for second shift.
Release cycle time is the number that settles the AI argument in a device plant. The baseline to capture, the arithmetic, and where the reviewer still signs.
© 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