If you're here from the other post, you might be confused. There I argued — with data, citing companies, with actual receipts — that the macroeconomic story for AI and white-collar work is probably fine. We're going to be fine.

This post is the companion piece, and the argument is narrower. AI is going to eat your job. Or at least, the parts of it that look like a job.

I want to be specific. Most of what people in knowledge work spend their day doing is task-shaped. Producing a deliverable. Drafting a memo. Filling in a form. Generating a report. Running an analysis. Triaging a queue. The cognitive content varies, but the structure of the work is consistent: input arrives, work happens, output is produced. That entire structure is exactly what current AI is good at, and it's getting better at it on a curve that doesn't appear to be slowing down.

If your job is mostly task-shaped, large parts of it are going to be eaten. Maybe not all, maybe not next year, but the trajectory is not subtle and pretending otherwise is its own form of denial. Aggregate gains, individual costs — and you might be the individual.

So the question is: what do you invest in if you want to still have a job in five years?

My answer, drawn from years of doing knowledge work and watching technology eat tasks at every layer of it, is that there are four things that don't get eaten. Not because of the limits of current models, those move every six months, but for structural reasons that I think will hold even when the models are better than I am at most of what I currently do. They're worth listing carefully, because they should also be where you're spending your professional development time.

Judgment under uncertainty

The most valuable thing I do at work — the actual reason I'm paid like a CIO and not a sysadmin — is making calls when there's no rulebook, no clean data, and the situation is genuinely novel. Two vendors, neither of whom has solved this exact problem before, and we have to pick one by Thursday. A security incident in progress where the playbook covers half of it. A staffing decision where the right answer depends on factors I don't fully know yet and waiting until I do would mean missing the window.

This is the work I find hardest, and it is also not coincidentally, the work that hasn't gotten any easier over time. There's no framework for it. The senior people you respect are the ones who can sit with the discomfort of an incomplete picture, weigh the cost of waiting against the cost of being wrong, and call the call.

AI is excellent at applying patterns from training data. It is much weaker at situations where the right move requires committing to action with insufficient information, knowing you might be wrong, and being prepared to live with the consequences. That gap doesn't get smaller as the surrounding tasks get cheaper. If anything, it gets more valuable, because more decisions hit the judgment-required stage faster when the analysis-required stage is automated away.

If you spend your career getting good at making calls under uncertainty (and being right often enough that your colleagues trust the next call), that's a skill that compounds. Compounding skills are how you stay employed.

Translating between domains.

A CIO's job is half technical, half political, and the entire job is translation. I take a security risk and explain it to a CFO in cost terms. I take a budget cut and explain it to my team in operational terms. I take a vendor's marketing pitch and translate it into "what does this actually do, and what does it cost us to operate over five years." I take an incident in progress and translate it into a board-ready update for an audience that doesn't know what an ASN is and shouldn't have to.

This is the work that is most clearly bilingual: speaking to two audiences at once and serving as the bridge between them. AI can summarize. AI can translate languages. AI cannot yet, and I suspect not for the stakes-bearing decisions, read the room. AI does not know what your CFO actually cares about, understand what your engineering manager is really worried about underneath the polite question they asked, or know that the right move in this meeting is to say less rather than more.

Translation between domains requires actual standing in both domains. You cannot credibly translate between communities you are not part of. The accumulation of that standing (your relationships, your read on individuals, your sense of what the unspoken concern in the room is) is irreducibly human and built over years.

If you can hold a real conversation with the engineers about implementation reality and with the executives about business outcomes, and translate one into the other in something close to real time, you are nearly impossible to automate out of a job. There aren't enough of those people to begin with, and the work needs more of them, not fewer.

Owning outcomes

I want to be careful with this one because it sounds preachy, and preachy is not the point.

The point is structural: accountability cannot be delegated to AI. Not legally, not socially, not organizationally. When something goes wrong, someone has to take the L; face the board, talk to the customer, write the apology, decide what changes. That responsibility falls to a human because no organization has yet figured out how to fire a model. And when they do figure it out, it'll be the human who deployed the model who gets fired.

This is the part of knowledge work that most people actually under-invest in, partly because it's uncomfortable and partly because the org chart doesn't always make it clear who owns what. The bias in most knowledge work is to be busy, to produce deliverables, to keep the queue moving. None of that is the same thing as owning a result. You can be very busy for years without ever being on the hook for anything that matters.

The people who consistently take ownership, those who say "this is mine, the outcome is on me, here's what I'm going to do about it", accumulate trust at an exponential rate, because the supply of people willing to do this is much smaller than the demand. AI has done nothing to increase that supply. It's possibly decreased it, because the easy availability of plausible-sounding work output has made it easier than ever to look productive without ever being responsible for an outcome.

Find the things in your work that you actually own end to end. Take more of them. The willingness to be on the hook is the most career-defining trait I've watched play out across twenty-something years and a lot of careers, mine and other people's.

Knowing what not to do

The fourth one is the hardest to describe and possibly the most important. It is the wisdom (taste, restraint, judgment, whatever you want to call it) to know what to ignore.

Every knowledge worker is now bombarded with possibility. AI makes it possible to do more things, faster, than ever. The implication a lot of people seem to be drawing from this is that they should be doing more things, faster, than ever. This is wrong. The implication is that you should be doing the right things, and the marginal cost of identifying the wrong things has gone up sharply, because the wrong things are now also free and fast.

The most valuable work I've done in IT has often been the work I decided not to do. The migration we didn't run because the system was being replaced anyway. The vendor we didn't sign because their roadmap didn't actually align with ours. The integration we didn't build because the underlying business problem turned out to be a process problem dressed up as a technical one.

AI is exceptionally good at giving you more options. It is conspicuously bad at telling you which options are worth pursuing in your specific situation, because that requires understanding your organization's history, its actual goals (as opposed to its stated goals), and the political topology of who's pushing for what and why. None of that is in the training data. And even if it were, there is no model that can actually be on the hook for what gets ignored.

The signal-from-noise filter is a human function. It has gotten more valuable, not less, because the noise floor went up.

So if I were a knowledge worker thinking about my career under the assumption that AI is going to eat the task-shaped parts of my job (which seems like a reasonable assumption) I would not be panicking. I would be doing four things, in roughly this order.

I would be deliberately taking on situations that require judgment, and seeking honest feedback on whether my calls are actually any good. I would be building genuine relationships in adjacent domains so I can credibly translate between them. I would be looking for things to actually own end to end and saying yes to that ownership when offered. And I would be practicing saying "no, this isn't worth doing", and being right about it often enough that people start trusting the no.

The aggregate story is fine. The individual story is up to you. None of the four things I just listed get eaten by AI. All of them are skills that compound. All of them get more valuable as the surrounding work gets cheaper.

This is, I think, an OK time to be a knowledge worker who's serious about the work. It is a much less OK time to be a knowledge worker who's mostly producing task-shaped output and hoping nobody notices.

Pick a side.

— Chris