When the AI Goes Down, Do You Still Know What's Right?

A colleague told me this morning: "I can't get my work done. The AI's down."

I nodded. Then spent the rest of the day unsettled, because I knew exactly what he meant—and what it revealed about where we are in this transition.

We've crossed a threshold most of us didn't notice crossing. Somewhere between "this tool is helpful" and "I can't work without it," AI stopped being optional assistance and became critical infrastructure. Not for everyone, not in every role—but for enough people, in enough workflows, that an outage now feels like the power going out.

The question isn't whether that's happening. It demonstrably is. The question is whether we're losing something more fundamental than productivity in the process.

The Calculator Precedent (And Why This Is Different)

Try doing real math without a calculator. Long division, by hand, on paper. It's miserable, and most of us can barely do it anymore.

The calculator made an entire generation worse at arithmetic. We know this. We don't care. Nobody actually needs to do long division. We quietly agreed the machine could have that one, and the world kept turning. Dependence on tools isn't new—it's the entire history of human progress. We're terrible at remembering phone numbers now. We've forgotten how to navigate by landmarks. Each technology erodes some prior skill, and we accept the trade because what we gain matters more than what we lose.

So when my colleague said "the AI's down," my first instinct was: this is fine. Tools break. We depend on tools. This is what technology adoption looks like.

But here's what's been nagging at me all day, and it's the part that makes this different from calculators.

The Part of the Job That Moved

A calculator does the labor. You still decide what to calculate. You set up the problem. You apply the formula. You interpret the result. Punch in a number wrong and get 40,000 where you expected 400, and something in your gut says "that's not right." The judgment stayed with you. Only the grunt work left.

That's the critical boundary. The tool handled execution. You retained verification.

I've watched this movie before. When Excel macros arrived, accountants worried they'd become obsolete. What actually happened? The tedious reconciliation work disappeared, and accountants spent more time on analysis and client advisory. The judgment work expanded to fill the space. Same pattern with tax software, audit automation, every wave of professional services technology for the past thirty years.

AI isn't following that script.

It's not taking the grunt work. It's taking the judgment. The analysis. The first draft of the thinking. The part that used to be the actual job. My colleague wasn't stuck on data entry when the AI went down—he was stuck on the analytical work he'd delegated without fully realizing it.

The Smell Test Is Degrading

Here's what I'm watching for with my own team, and it's what keeps me up at night about this transition:

Can you still tell when the output is wrong?

Not obviously wrong—nonsense, hallucinations, factual errors. Those are table stakes, and honestly, most professionals I work with catch those pretty quickly right now. I'm asking about the subtler failure mode: plausible but incorrect. Confident but flawed. Well-formatted but fundamentally misaligned with what the situation actually requires.

The calculator never degraded your ability to smell a wrong answer. If anything, it sharpened it—you did more problems, built more intuition, caught errors faster because you weren't exhausted by arithmetic.

I'm not sure we can say the same about AI-assisted analysis.

I was reviewing a risk assessment last month that was beautifully written, properly structured, hit every compliance checkbox. It was also investigating the wrong risk entirely—focused on technical implementation when the real exposure was operational process. The analyst who'd drafted it (with significant AI assistance) didn't catch it. Not because they weren't smart, but because they'd stopped doing the repetitions that build pattern recognition.

They'd outsourced the thinking before they'd fully developed the judgment.

The Outage as Stress Test

When my colleague said "the AI's down," my second thought—after the initial unease—was: this might be the most valuable thing that's happened to their team this month.

The outage isn't really a productivity problem. It's a competence test.

Not a punitive one. An honest diagnostic. When the AI comes back online, the easy measure of success is "we're fast again." The real measure is whether you can still verify the output. Whether you've maintained the skills that let you distinguish great analysis from plausible nonsense delivered in a confident paragraph.

I'm not anti-AI. I use it constantly. It's genuinely transformative for certain workflows, particularly the ones where I'm synthesizing across domains or need to explore multiple analytical angles quickly. But I've also started deliberately doing some work the slow way—not as my primary workflow, but as maintenance. The equivalent of a pilot practicing manual landings even though autopilot handles most flights.

Because here's the thing about erosion: you don't notice it's happening until you need the thing that's gone.

What Skills Are You Not Practicing?

This is the uncomfortable question I've been sitting with since this morning's conversation:

What's the one skill you've quietly stopped practicing because the AI does it now?

For me, it's initial research synthesis. I used to read five sources, make connections, spot the gaps, build the framework myself. Now? I'll often feed sources to AI and ask it to synthesize. It's faster, it's usually pretty good, and I tell myself I'm still verifying the output.

But am I actually verifying—or just checking whether it sounds right? There's a difference, and I'm not always sure which one I'm doing anymore.

I'm asking this of my team now, too. Not to police AI usage—that's a losing battle and the wrong fight anyway. But to identify which skills need deliberate practice because they're no longer getting incidental practice through daily work.

Could you still catch it when it lies to you, confidently, in a clean paragraph?

That's the capability that matters. Not "can you work faster with AI" but "can you still work correctly when AI gives you something that looks right but isn't."

The Railroads Arrived Quietly, Too

Nobody gets fired the day the railroad arrives. The town just slowly empties out.

I've watched enough technology transitions to know the pattern. The disruption doesn't announce itself. It compounds quietly until one day you look up and realize the critical skills in your organization are concentrated in people who learned their craft before the tool arrived—and nobody's developing those skills anymore because the tool handles it.

What happens when those people retire? When the institutional knowledge walks out the door?

This isn't hypothetical. I'm seeing it in audit teams where junior staff have never manually traced a transaction because the software does it. In tax practices where associates can't explain the logic behind a position because they've always started from AI-generated drafts. The tool isn't making them faster at a skill they possess—it's preventing them from developing the skill in the first place.

Maybe that's fine. Maybe those skills don't matter anymore, the same way long division doesn't matter.

But I'm not convinced yet. And the difference between calculator-level automation and AI-level automation is that we knew what we were giving up with calculators. We understood the boundary. We made the trade deliberately.

I don't think we understand the boundary yet with AI. We're making the trade before we fully understand what we're trading away.

What to Do Monday Morning

Here's what I'm asking my team to do—and what I'd encourage you to consider for yours:

Pick one skill that AI has been handling for you. Do it manually, once a week, for the next month. Not as your primary workflow—I'm not suggesting we abandon productivity gains. But as deliberate practice. As calibration.

For analysts: draft one assessment without AI assistance, then compare it to what you would've produced with AI. What did you catch in the manual version? What did you miss?

For managers: review one AI-assisted work product by recreating the analysis independently first. Did you catch different issues? Would you have approved the AI version without that check?

The goal isn't to prove you don't need AI. The goal is to maintain the capability to verify AI. To preserve the smell test.

Because depending on a tool is fine. We depend on tools constantly, and it's made us more capable across nearly every domain.

Losing the ability to know when it's wrong is not.

The AI will come back online. It always does. The question is whether we'll still be able to tell the difference between good output and confident nonsense when it does.

What skill have you stopped practicing? And more importantly—could you still verify the AI's work in that domain, or have you outsourced the judgment along with the labor? I'm genuinely curious where others are drawing this line. Hit reply or find me on LinkedIn.