The Signal #122 — Dakota’s read on the AI news that actually matters to people running a business.

There is a version of AI adoption that looks like a win on every dashboard and quietly creates a serious problem underneath.

Automation handles the routine work. Response times drop. Escalations fall. Costs go down. Everyone celebrates. Then something genuinely strange happens, the kind of failure nobody has seen before, and the people who are supposed to handle it are rustier than anyone realized.

That is not a hypothetical. It is a pattern with a forty-year research history behind it. And it is worth understanding before you automate your way into it.

What happened

A newsletter from the observability team at SigNoz published a detailed piece on what AI is actually doing to site reliability engineers, the professionals responsible for keeping software systems running. The article is not a doom piece about job replacement. The concern is more specific and more interesting than that.

The argument starts with a 1983 paper by cognitive psychologist Lisanne Bainbridge called “Ironies of Automation.” It became one of the most cited papers in human factors research. Bainbridge studied factory automation and found something paradoxical. The more you automate a process, the more critical the human operator becomes during the rare moments automation fails, and the less practice they get, and the worse they become at exactly those interventions. Workers ended up with long stretches of passive monitoring punctuated by rare, high-stakes crises they were increasingly unprepared for.

The SigNoz piece draws the parallel directly. Industrial automation took over routine physical tasks. AI is taking over routine cognitive tasks: pattern matching, triage, known-issue resolution. What is left for humans is the rare, complex, ambiguous problem. The exact kind that requires deep expertise. The exact expertise that atrophies when you stop practicing.

The research backing this up is not soft. A study cited in the piece found that endoscopists who used AI assistance for polyp detection saw their unassisted detection rates drop from 28 percent to 22 percent after a period of AI use. They got worse at the core skill, not because they forgot the theory, but because they stopped exercising it. In aviation, research showed that long-haul pilots relying heavily on autopilot experienced measurable degradation in situational awareness and manual flying ability. The FAA now mandates more manual flying time specifically to counteract the effect.

In the SRE context, the article describes a tiered model that platforms like PagerDuty are moving toward. Tier 1 incidents, the known issues with known fixes, get fully automated. Tier 2 incidents, the partially understood problems, get AI recommendations with human validation. Tier 3 incidents, the novel, cascading, never-before-seen failures, stay human-led. The question the piece asks is a sharp one. If human practitioners only engage with Tier 3, where do they build the intuition to handle it?

Why it matters

This is not just an SRE story. The same dynamic applies anywhere an organization automates the routine work of a skilled role.

A real estate brokerage that uses AI to handle standard contract reviews still needs someone who can read a genuinely unusual deal structure without a template to lean on. An e-commerce operation that automates customer escalation routing still needs a support lead who can handle the situation that does not fit any existing category. A marketing agency that runs AI-generated campaign analysis still needs a strategist who can recognize when the model is confidently wrong.

The routine cases are exactly where humans build the intuition they need for the hard cases. Automate the routine entirely, and you are not just changing the workflow. You are changing what your team actually knows how to do.

What most people get wrong

The mistake is treating deskilling as a future risk rather than a current one.

Deskilling does not announce itself. Detection rates drop from 28 percent to 22 percent quietly. Pilots lose manual flying ability gradually. SREs lose the instinct for system behavior over months of watching dashboards instead of debugging. The metrics that an organization typically watches, resolution time, escalation rate, ticket volume, will look fine or better right up until the moment they do not.

The other mistake is assuming that because a team still technically has the knowledge, the skill is intact. Knowing how to do something and being practiced at doing it under pressure are different things. Bainbridge’s insight was that automation leaves humans with the worst possible job: monitoring a system they are progressively less equipped to take over when it breaks.

The lesson

Automation decisions are also training decisions. Whatever your AI systems handle routinely, your people stop practicing routinely. That trade can absolutely be worth making. The key is making it consciously.

If you are rolling out AI automation on anything your team used to do manually, it is worth asking which skills you are trading away, whether those skills matter when things go wrong, and what the plan is to keep them sharp. Scheduled drills, manual rotations, deliberate practice on the cases the AI handles, these are not inefficiencies. They are the insurance policy that makes the automation safe to run.

The organizations that get this right will have faster operations and capable people. The ones that do not will find out when something Tier 3 happens.

If you want to think through how AI fits into your operations without creating hidden fragility, xovionlabs.com is a good place to start.