The AI Jobs Apocalypse Already Started. You Just Didn't Get the Memo.

Everyone's waiting for the AI jobs apocalypse. The data says it already started, and it looks nothing like a headline.

The European Central Bank just published research tracking AI's effect on US jobs and wages between 2019 and 2025. The aggregate number? Muted. No mass unemployment, no crater in the labor market. So the story dies in most feeds right there, filed under "AI disruption was overblown."

Read one layer down and it's a completely different story. Jobs at high risk of AI substitution grew about 15 percentage points slower than low-risk jobs. Employment in those high-risk roles didn't just slow. It fell more than 4%.

That's not nothing. That's the shape disruption actually takes before it has a name.

We Keep Expecting Extinction. What Shows Up First Is Reallocation.

I've watched this movie four times now. The spreadsheet didn't fire the finance department in 1985. It quietly redrew what an analyst did all day, and the people who only did the part the spreadsheet now handled got squeezed out of the growth curve. Same with the factory robot in the '90s. Same with offshore teams in the 2000s. Same with cloud automation in the 2010s.

The first casualty is never the job. It's the raise, the new hire, the next opening that never gets posted.

Nobody wakes up to a pink slip that says "eliminated by AI." What happens is subtler and more dangerous: the most dangerous disruption is the one that arrives slowly enough to look harmless. Your department goes from replacing three people who retire to replacing one. Then to replacing none. Then leadership starts asking why the team is sized the way it is.

By the time it feels like a problem, you're three years behind.

The Slope You're Standing On

Here's what 4% annual employment decline actually looks like on the ground.

A mid-sized accounting firm has twelve tax preparers. One retires. They don't backfill — the AI-assisted review tools are handling more of the routine compliance work, and the remaining eleven can absorb it. A year later, another leaves for a competitor. They don't replace that one either. No dramatic announcement. No all-hands about restructuring. Just... a slope.

The people still there don't feel it as disruption. They feel it as "we're stretched thin" or "leadership is cheap." The person who left and can't find a comparable role? They feel it as "the market is tough right now."

This is how reallocation works. It's not a headline. It's a thousand small decisions that compound.

For anyone whose day is routine cognitive work — and I'm talking to CPAs running standard compilations, auditors executing testing procedures, analysts building the same monthly variance reports — the market is already repricing what you do. Not with a layoff. With a slope. And a slope is easy to miss while you're standing on it.

Execution vs. Judgment: The New Dividing Line

I was talking to a client last month — partner at a regional firm — and she said something that stuck with me: "I can get a first draft of any technical memo in four minutes now. What I can't get is someone who knows which footnote actually matters to the CFO."

The work that holds value is shifting from execution to judgment.

The AI can run the analysis. It can draft the memo. It can pull the comparables and format the workpapers. What it can't do — yet, and maybe ever — is sit across from a nervous audit committee and read the room. It can't prioritize which of seventeen findings actually threatens the deal. It can't translate "material weakness in IT general controls" into language that makes a board take action.

Execution is the commodity. Taste and translation are the moat.

This isn't a prediction. This is already the operating reality in every firm where people are actually using this technology, not just talking about it. The analysts who only knew how to execute the testing procedure are getting managed out or plateaued. The ones who can frame what the finding means and what to do about it are getting pulled into more complex engagements.

If you're betting your career progression on being really good at the part the AI is really good at, you're making a bad bet.

The Uncomfortable Questions

So here's what I want you to sit with, because I don't have a clean answer for you and neither does anyone else:

If your role were repricing 4% a year, would you feel it, or would you only notice once the raise stopped coming?

What percentage of your day is judgment versus execution? And if you're honest about it, is that percentage moving in the right direction?

When you think about the skills you're building this year — the training you're seeking out, the projects you're volunteering for — are you doubling down on execution mastery, or are you building the translation and taste that machines can't replicate?

These aren't rhetorical. I'm asking you to actually answer them, because the firms and the careers that come out ahead in the next three years will be the ones who got honest about this in 2025, not 2027.

What to Do Monday Morning

This is a planning question, not a panic one. But it is a now question.

Here's what I'd do if I were managing my own career risk — and what I'm advising clients to do with their teams:

Audit your skill mix. Track your time for two weeks. Be ruthlessly honest about what percentage is execution (running the process, filling out the template, performing the procedure) versus judgment (deciding what matters, translating for the client, navigating the edge case). If it's more than 60% execution, you're in the danger zone.

Volunteer for the messy work. The AI handles the clean, repeatable processes. It struggles with ambiguity, conflict, and context. So the person who takes the difficult client conversation, the contentious audit finding, the implementation with seventeen exceptions — that person is building AI-resistant skills.

Build translation as a competency. Your value isn't knowing the technical answer. It's making the technical answer useful to someone who doesn't live in your world. Practice explaining your work to people who aren't accountants. Get good at it. That's the moat.

The aggregate numbers will keep looking fine for a while. The headlines will keep missing the story. But the reallocation is already here. The question is whether you're going to wait for it to have a name, or whether you're going to treat it like the planning problem it actually is.

The slope doesn't announce itself. But what do I know — I've only watched this movie four times.

What's one thing you do in your role today that you couldn't train an AI to do in six months? If you can't answer that quickly, it's time to start building something new.