Paper didn't fail us. The gaps between document control, training, and CAPA did.

Last year a batch of finished components shipped against a work instruction that had been revised three weeks earlier. The current revision sat in the document register. The acknowledgement that people had read it sat in a paper training folder. Nothing connected the two, so at audit time I had a current document, a training record, and no evidence that they belonged to the same revision.

That was the moment I stopped treating "digitise the QMS" as a filing project.

Some context, since it shapes everything below. I lead QA at a contract manufacturing organisation in Bangalore. We build components for three global medtech OEMs, we carry more than forty suppliers on our approved list, and the quality records I inherited in 2014 lived in three formats: a printed document register, an Excel CAPA log, and a training matrix nobody trusted. Each of those three was internally defensible. The joins between them were missing — and the joins are where an audit actually happens.

Paper was never the real problem

The usual complaint about paper is legibility, and it is the least interesting problem. Paper's real failure mode is that relationships live in a person's memory. "The technician was trained on revision 7" is, on paper, only reconstructible if someone remembers.

Each of our three systems held a true fact that the other two needed. The document register knew a revision existed. The training matrix knew who had attended a session on a date. The CAPA log knew a non-conformance had been investigated and closed. An OEM auditor, a notified body assessor, or my own reviewer asking "show me how these three connect" would get a guess and a lot of goodwill.

We came to call that nadu — that's how it is — which is a very Tamil way of describing a systemic weakness.

We did most of the redesign on a Sunday morning with filter coffee and no notifications, which is the closest thing I have to a reliable design session. An RCB match in the background is surprisingly bad for this kind of work, although the crowd's instinct to coordinate individual effort is both the problem and, in fairness, the answer.

What the standard actually asks for

Before changing anything, I read the clause rather than the summary.

• ISO 13485:2016, 4.2.4 (control of documents). Documents shall be reviewed and approved for adequacy, revision status identified, changes and their impact assessed, made available at the point of use, and obsolete versions prevented from unintended use. Note what is absent: it says neither "paper" nor "electronic". It also does not require you to prove which revision a specific operator used on a specific day — but that is precisely what an auditor will ask, and a binder at the line cannot answer it.
• ISO 13485:2016, 6.2 (human resources). Competence must be determined, training provided, the effectiveness of training actions evaluated, and records retained. An attendance sheet with a date proves an event occurred. It does not prove the person can now do the job.
• ISO 13485:2016, 8.5.2 and 8.5.3 (corrective and preventive action). Determine causes, evaluate whether the action eliminates them, implement, record results, and review effectiveness. The clause's own logic pushes you downstream: if the corrective action changes how work is performed, then a revised document and a competent person are part of the evidence, not follow-up chores.
• ISO 14971:2019 (risk management). Production and post-production information feeds risk evaluation. A repeated complaint is an input to the risk file, so a CAPA that changes a process should update risk. A CAPA record that sits next to the risk file rather than inside it breaks that chain.
• 21 CFR 820.40 and 820.70(d) (document controls, training) add the US expectations: training adequate and appropriate for the intended use of the device, effectiveness documented, records retained. Electronic signatures for approvals or training acknowledgements bring 21 CFR Part 11 with them.
• EU MDR 2017/745, Annex IX expects the QMS to close loops through audit, complaint handling and corrective action, with a notified body testing the system rather than the paperwork.

One more that gets skipped: ISO/TR 80002-2:2017, the guidance on validating software used in a medical device quality system. If any software sits inside your quality processes, that is the document your reviewer will eventually ask about.

Three moves that changed our day

1. Track revision state, not revision files. "Current" became a state attached to the work instruction and carried by anything issued from it, including the controlled print at the line. Paper at the point of use is defensible. A printed copy nobody can tie back to a revision is not.
2. Trigger training from the quality event, not the calendar. A CAPA closure that changes a process creates a training requirement. A new revision of a procedure creates a training item attached to that revision. Calendar-driven training, attended because a date arrived, is a filing exercise with a signature on top.
3. Make CAPA closure depend on downstream evidence. Root cause, corrective action, revised document, updated risk file, effectiveness check. If one is missing, the CAPA stays open. That is what a connected workflow means for us in practice: not one screen, but the ability to see what is missing.

The workflow behind this is built and maintained in-house. The tooling matters far less than the joins between records, and the joins are where the design decisions live.

Where assistance fits, and where it doesn't

We use AI-assisted drafting in two narrow places: candidate root-cause hypotheses from a non-conformance description, and draft training questions generated from a revised procedure. In both, the output is a suggestion that a named owner reviews, edits, and signs. Nothing closes a CAPA, and no training record is created, without a human decision recorded against it. If your procedure does not say who reviews the output, you do not have controlled assistance; you have an unattributed suggestion in a chat window.

Where paper is still the right answer

I would push back on the assumption that everything should move.

• If the process is unclear, digitising it produces a clearer-looking version of the same confusion. We had a CAPA investigation form that asked people to "be more careful"; in electronic form it remained an ineffective investigation.
• Some records are genuinely low-volume and high-consequence — certain batch release records for one OEM — where a controlled paper original plus a scan is a reasonable, defensible choice.
• Digitise the flow before the archive. We migrated six years of historical CAPAs early because it was visible. It changed nothing. The value came from CAPAs raised after the change.

What I would do differently

• Start with one CAPA type and run it end to end for a quarter — supplier non-conformance, in our case — including training and document revision, before touching complaints or audits.
• Define "connected" in testable terms before you build or buy anything. Ours is simple: one query returning which revision was in use, who was trained on it, and which CAPA changed it.
• Settle retention, electronic signatures and access control before