AI helps doctors find more precancerous growths during a colonoscopy. But in one study, doctors who had been using AI for only about three months found fewer of those growths on their own than they had before the AI arrived.
I think quality work is about to face the same trade. AI can make us better at work that needs breadth and depth, like reading every record in an audit or pulling together the evidence for an investigation. It can also make us worse at the parts that need our own judgment, like asking what the evidence doesn't show or questioning what the AI hands back to us.
In my view, some deskilling comes with any AI good enough to be worth using, so we can't prevent it, but we can control it. We can hand over the work we are willing to get worse at, and keep practicing by hand the skills we can't afford to lose. Deskilling that we choose is a planned trade-off, and deskilling that just happens to us is an unmanaged risk.
Here's what's inside:
- The pull to trust AI. Why it's tempting to let the AI vouch for everything, and why the market will push most of us toward a hybrid anyway.
- Where AI enhances. When the AI does the reading and we keep checking its work, we see far more than we could alone.
- Where AI deskills. When a well-supported AI answer becomes the reason we stop asking questions, we start down a dangerous path of deskilling.
- Making the choice deliberately. Results will show us where the line belongs, but we still have to decide which skills to hand over and which to keep.
The pull to trust AI
I feel this pull myself. I can point an AI at every SOP and every record a client has, ask it what's missing, and in a real sense audit all of those documents in five minutes. The temptation is to look only at what the AI flags and let it vouch for the rest.
The usual answer to that temptation is to create friction by still doing the work yourself so you do not lose the skills. For example, you still check a sample of everything yourself, just as you would in a normal audit, even if you are using AI. But that raises a nagging question: If I don't really trust the AI, and if I still have to do the same amount of work, why am I using it at all? My answer is that we still use AI because it makes the results better, even if it doesn't save us much work.
This is where the market comes in, because it always pushes toward the best result for the least cost, and that will almost certainly mean a hybrid where the AI does much of the work. Say I spend a day and a half on an audit, and a competitor uses AI to do the same audit in a fraction of the time for much less money. If the AI finds 97 or 98 percent of what I would have found, most clients will take that deal over paying me much more for 99.
And let's be honest, the read-every-record audit was disappearing long before AI showed up. In my view, audits have been getting thinner for years, with fewer days on site and fewer records pulled, and the certificate still gets issued at the end. So I don't think most of us will get to choose whether we deskill, only which skills we let go, and if we don't make that choice ourselves, price and convenience will make it for us. The real advantage goes to whoever finds the right balance first.
Where AI enhances
AI makes an audit better when it does the reading and the auditor keeps checking its work. Here's a fictional example. An auditor has the AI review every supplier lot that a manufacturer received last year, across the receiving records, lab results and release decisions.
One lot's release packet looks complete, with a passing test result and a release signature. This is what the auditor might have found on his own. But the AI can easily see the whole integrated universe of records, including the lab system, and it finds an earlier test on that same lot that failed. The AI would flag that discrepancy, and further investigation would show that nobody investigated the failure, the lot was retested, and only the passing result went into the packet.
That doesn't mean the auditor stops sampling. He still pulls a sample and reviews it himself, the way he always has, but the sample now has a different job. Before AI, the sample was the audit, because nobody could read every record. Now the AI reads every record, and his sample is how he checks the AI, including the records it said were fine.
He also checks every finding against the source records and talks to the people involved before he writes it up. What he lets fade is his speed at paging through records, and I think that's a fair trade for an audit that covers every lot.
So maybe the real question isn't how good the AI is, but how good the person driving it is, and whether that person keeps their own skills sharp.
Where AI deskills
AI starts to cost us skill when a well-supported answer becomes the reason we stop asking questions. CAPA investigations are where that worries me most.
Here's a second fictional example. A cleaning check keeps failing on the same piece of equipment. The AI reviews the records and finds that the same operator skipped a required cleaning step each time, even though he was trained on it. It concludes that he didn't follow the procedure and recommends retraining, and every record it points to is real.
What the AI couldn't see is why he skipped the step. A replacement part had been installed that jams when the equipment is hot, and when he reported it, his supervisor told him to clean around it so production wouldn't be delayed. None of that was written down, so none of it was in the records.
Quality people know the five whys. The AI got the first four right, but the fifth, why a trained operator would skip a required step, could only be answered by going out to the floor and doing a proper investigation.
This is the trade I don't think we can afford. The AI's answer looks like a finished investigation, so it would be easy for a quality manager to sign it without doing one of his own. Retraining won't fix the part, and the blame lands on the one person who reported the problem. What we lose is the habit that has always been at the heart of CAPA work: asking what isn't in the records and going out to look.
Making the choice deliberately
In both cases the AI did good work with the records it had. The difference was what the people around it kept doing. The auditor chose which skills to let fade and kept practicing the ones he needs to check the AI, while in the CAPA nobody made that choice, so the AI's thoroughness made it for them.
Where the line falls will be different for every kind of work, and I think results will settle most of it. The checks that keep catching what the AI misses are worth what they cost, and so is practicing the skills behind them.
This is the work we do at Certify Consulting Group. We help regulated manufacturers decide deliberately which skills an AI-assisted quality system can let fade, and design their audit and CAPA processes so their people can still challenge what the AI tells them. If your team already uses AI in audits or CAPAs, our AI in the Quality System Review is a good place to start: a fixed-fee, two-week look at where AI is already writing, checking or deciding inside your quality system. You can book a free 30-minute call from that page, or send a note through the contact page.
None of the doctors in that colonoscopy study chose to get worse without AI. Most of us in quality will let some skills go over the next few years, and I think that's fine, but I would rather we make that trade on purpose.
On your next AI-assisted audit or CAPA, which skill will you be practicing less, and will you still need it to catch the AI when it's wrong?
First published in my LinkedIn newsletter, AI in Regulated Industries.
Jared Clark
Principal Consultant, Certify Consulting
Jared Clark is the founder of Certify Consulting, helping organizations achieve and maintain compliance with international standards and regulatory requirements.