![\"Developer](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/ai-skill-erosion-the-expertis-inline-1.jpg\")
<\/figure>\nWhat ‘forever junior’ looks like on a plant floor, not a codebase<\/h2>\n
A quality engineer signs off on an AI-drafted deviation report. The logic reads clean. Then an auditor asks why that root cause was selected over two plausible alternatives, and the engineer can only point back at the output. Approving is not the same as defending.<\/p>\n
Same pattern in planning. An ops planner accepts a scheduling recommendation that looks sensible, but cannot argue why it beats running the changeover in the opposite order. Baturone’s list of what makes a Senior on her team includes the ability to “defend code they’ve shipped, argue a design choice instead of just having one.” Swap code for a CAPA or a production plan and nothing about that standard changes.<\/p>\n
The reps that built your current experts no longer exist in the workflow<\/h3>\n
Your best process engineer earned that judgment by doing the tedious version of the job for years. Chasing down a contamination source through batch records. Being wrong twice before being right. Arguing with a supplier about a spec that turned out to be fine.<\/p>\n
Those reps were never on a training plan. They were a byproduct of the work being slow and manual. Automate the manual part and the byproduct disappears with it, silently, because nothing in your metrics tracks it.<\/p>\n
Why this shows up first in audits, escalations, and edge cases<\/h3>\n
Routine work hides the gap well. Ninety percent of deviations fit a known pattern, and the agent handles the pattern competently. You will not notice anything is missing while the distribution behaves.<\/p>\n
Then a regulator probes the reasoning behind a decision, or a failure mode appears that matches nothing in the training data, and you need someone who can tell when the output is heading the wrong way. That is the specific capability at risk. AI skill erosion does not announce itself in throughput numbers. It announces itself the first time an escalation lands on a desk and the person sitting there has only ever reviewed, never reasoned.<\/p>\n
What seniority actually means, defined by behaviour not years<\/h2>\n
Baturone did something most organisations never bother to do. She watched the Seniors on her team at Criteo and wrote down what they actually did differently. Five behaviours. None of them involve years of service, and none of them can be held by an agent on your behalf.<\/p>\n
The five behaviours, translated to quality and operations roles<\/h3>\n
Her list maps cleanly onto a plant. Translate it and you get a usable standard for what a senior quality or operations person looks like in an agentic workflow.<\/p>\n
- \n
- Defend what you shipped<\/strong>: explain a released batch, a validation decision, or a process change to a regulator six months later without reopening the tool that produced it.<\/li>\n
- Argue a choice, not just hold one<\/strong>: state why this control limit, this sampling plan, this supplier qualification route, and name the option you rejected.<\/li>\n
- Spot the wrong direction early<\/strong>: notice when a model’s suggested root cause contradicts what the equipment history says, before the CAPA is signed.<\/li>\n
- Know when to delegate and when to do it yourself<\/strong>: automate the recurring deviation write-up, personally walk the line when the failure mode is new.<\/li>\n
- Teach the rest<\/strong>: turn one investigation into a pattern three other engineers can reuse.<\/li>\n<\/ul>\n
Read that list against your own team. Most leaders find they have people strong on the first two and almost nobody strong on the third.<\/p>\n
A 20-minute exercise to score your team against them<\/h3>\n
Take your last five AI-assisted outputs: a deviation report, a schedule, a CAPA, a trend analysis, a supplier assessment. For each one, ask the person who approved it a single question. Why this conclusion over the next most plausible one? Write down whether they answer from reasoning or from the document.<\/p>\n
Score each behaviour red, amber, green per person. Twenty minutes gets you a map of where judgment skills exist and where they have quietly stopped forming. That map is more useful than any adoption metric on your dashboard, because it tells you which approvals are real and which are rubber stamps.<\/p>\n
![\"Senior](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/ai-skill-erosion-the-expertis-inline-2.jpg\")
<\/figure>\nDesigning AI workflows that build judgment instead of replacing it<\/h2>\n
Baturone’s first move was writing a
skill.md<\/code>, a set of instructions fed to the model so it behaved less like a vending machine and more like a mentor. She told the agent to hit her with Socratic questions after each decision, a pop-quiz on the plan before she accepted it. The mechanic is simple and it transfers directly to quality and operations work.<\/p>\nThen she noticed the harder thing. The real gap was not in her agent’s behaviour, it was in hers. You can configure a model to interrogate people all day; if nobody is held accountable for the answers, the ritual dies in a week.<\/p>\n
Three workflow changes that force judgment back into the loop<\/h3>\n
Start with the order of operations. Make the human state the expected outcome before the output appears on screen. Before the agent returns a root cause ranking, the engineer writes down which failure mode they expect and why. Mismatches become the most valuable training signal you have.<\/p>\n
- \n
- Ask before answer<\/strong>: configure the agent to surface its assumptions and two rejected alternatives, then ask the reviewer which one it ruled out and on what evidence.<\/li>\n
- Prediction before output<\/strong>: a one-line written call from the human, logged, before the recommendation is revealed.<\/li>\n
- Defend as a standing step<\/strong>: sign-off requires a short verbal or written defence to a peer, not a checkbox. If the only justification is “the system said so”, it goes back.<\/li>\n<\/ul>\n
Track the mismatch rate per person over time. That number tells you who is building judgment and who is rubber-stamping. It is the cheapest early warning on AI skill erosion you will ever install.<\/p>\n
Where to accept pure automation and stop worrying about it<\/h3>\n
None of this applies everywhere. Formatting, transcription, data extraction, translating a batch record into a summary table, pulling historical deviations matching a pattern. Automate those completely and never look back. Nobody’s expertise depends on retyping a lot number.<\/p>\n
The line is whether the task requires a defensible choice between plausible alternatives. Judgment tasks get friction by design. Clerical tasks get none.<\/p>\n
- Prediction before output<\/strong>: a one-line written call from the human, logged, before the recommendation is revealed.<\/li>\n
- Argue a choice, not just hold one<\/strong>: state why this control limit, this sampling plan, this supplier qualification route, and name the option you rejected.<\/li>\n