“Will AI replace my job?” is the wrong question, and most of the public discourse about AI and work is wrong because it answers it anyway. Jobs are not the unit of analysis. Tasks are. And tasks don't all sit on the same spectrum of replaceability — they fall into one of four economically distinct classes. Once you see the classes, the question changes shape: not can AI do my role, but which fraction of my role lives in which class, and what that implies for cost, compensation, and where to spend the next year of your career.
This is the canonical Wagecore taxonomy. We assign every task in every Wagecard to exactly one of these four classes, and the headline substitution-exposure number is a weighted aggregate over them. The framework draws on a decade of automation economics literature (Autor, Frey/Osborne, Brynjolfsson, Acemoglu) plus the AI deployment post-mortems of the last three years. The contribution is committing to a small, mutually exclusive partition that the math can ride on.
Class 1: Replaceable
A task is replaceable when AI runs it end-to-end with minimal human oversight, audit rates under 10%, and an error cost low enough that the cases where it gets it wrong don't wipe out the savings on the cases where it gets it right. The work is bounded, repetitive, structured, and the consequences of an individual mistake are recoverable cheaply.
Concrete examples: triaging an inbound customer support ticket to the right queue, extracting structured data from invoices into an ERP, OCR plus first-pass classification on incoming forms, generating product descriptions from a SKU and an image, summarizing a long meeting transcript into action items, transcribing audio. These are tasks where the AI's output is checked at the moment it produces a downstream signal (the ticket went to the wrong queue → a human moves it), not weeks later in a courtroom.
Most roles have a non-zero fraction of replaceable work. Even surgeons have a sliver: dictating clinical notes is now replaceable in many practices. Even therapists have it: appointment scheduling, intake forms, insurance verification. The mistake is assuming the role's replaceable fraction is the whole role. Usually it's 15–35% of a knowledge worker's time.
Class 2: AI-augmented
A task is AI-augmented when AI produces the first version, the human owns the last 20–30%, and that last fraction is where the value comes from. The AI does the bulk of the keystroke work; the human supplies judgment, context, and accountability for what gets shipped.
Concrete examples: writing a marketing email (AI draft, human refines for voice and audience), drafting a legal demand letter (AI pulls precedent and structure, lawyer applies case-specific facts), generating code for a feature (AI writes the scaffold, developer integrates with the codebase and handles edge cases), creating slides for a client pitch (AI builds layout, salesperson edits for positioning), preparing a financial model (AI builds the template, analyst tunes assumptions).
This is the largest single class for most knowledge workers, typically 25–40% of time. It's also the class with the most growth potential as models get better and the human-review surface shrinks. But it has a ceiling: as long as the human is on the hook for what ships, they need to know the work well enough to catch the AI's mistakes — which means the human still has to be in the loop, still has to be paid, and still has to have the underlying skill. AI-augmented isn't a path to zero-headcount; it's a path to leverage.
Class 3: Human-led, AI-assisted
The inverse of AI-augmented. The human leads; the AI is a tool — fast lookup, summary, code completion, retrieval. The human is doing the thinking and making the decisions; the AI is shrinking the time between question and relevant information. If you removed the AI, the work would still get done, just slower.
Concrete examples: a doctor querying the literature for similar presentations before a diagnosis, a lawyer asking the AI to find the clause in a 300-page contract that contradicts a position, an engineer asking for the syntax of a library they last used three years ago, a teacher generating worksheet variations to differentiate a lesson, an architect using AI to render a façade option they've already designed.
Roles in regulated, high-stakes, or relationship-heavy work tend to cluster heavily here: 30–50% of time. The AI doesn't make the decisions, doesn't carry the accountability, and isn't allowed to — either by law (medical, legal advice) or by physics of the work (the therapy session, the customer relationship, the team conflict). What it does is make the human faster at the parts of their job that are information-bound rather than judgment-bound.
Class 4: Human-critical
A task is human-critical when AI delivers no net value, and often negative value, because the value of the task is in something the AI cannot produce: trust, accountability, ambiguity tolerance, relational judgment, persuasion under pressure, contextual reading of an unfamiliar room. These are not tasks that AI hasn't caught up to yet. They are tasks where the AI being plausible-sounding is itself the failure mode.
Concrete examples: a senior salesperson reading a stalled deal and deciding whether to escalate or back off, a manager delivering hard news to a team member who has been with the company for fifteen years, a therapist sitting in silence while a client gathers themselves, a board member calibrating a CEO's confidence claim against what they saw at lunch, an investigative journalist deciding which of two contradictory sources to believe, a teacher noticing that a usually-engaged student has gone quiet and choosing whether to address it now or in private later.
Human-critical work is what doesn't scale, and that's the point. It's also where pricing power lives. Roles that are 40%+ human-critical are the roles where AI deployment makes the work more valuable per hour, not less — because the augmentation strips the lower-leverage time and concentrates compensation against the irreducible core.
Most roles are a mix, not a single class
Here is the part that the public discourse keeps getting wrong: very few roles are 100% in any single class. A software engineer's actual weekly time might land at roughly 20% replaceable (boilerplate, ticket triage), 35% AI-augmented (feature implementation under review), 30% human-led, AI-assisted (debugging gnarly production issues, architecture decisions), 15% human-critical (negotiating scope with a PM, mentoring a junior, navigating a politically-loaded code review). A customer-support team lead might land at 30% replaceable (tier-1 ticket handling), 25% AI-augmented (drafting macros and policy docs), 30% human-led + AI-assisted (handling escalations the AI can't defuse), 15% human-critical (1:1s with team members, conflict mediation, performance conversations).
The distribution matters more than any single number. A role that is 80% replaceable will compress on price even if its average task is non-trivial, because the deployment economics are clear. A role that is 50% human-critical will retain pricing power even if the rest of it gets automated to zero — and the average compensation per remaining hour will go up.
Why the four-way partition (and not three, or six)
Earlier frameworks used two classes (replaced / not replaced) or three (replaced / augmented / unaffected). Two is too coarse — it collapses AI-augmented and human-critical into “not replaced,” which hides the central truth that augmentation can compound pricing power while replaceable work strips it. Three is closer but folds the most economically distinct cases — human-led, AI-assisted versus human-critical — into a single bucket. They behave differently. A diagnosis is human-led, AI-assisted (the AI helps with literature review). A patient telling their doctor they don't trust their spouse is human-critical (the AI is actively in the way).
Six or more classes is overfitting. The marginal granularity stops carrying economic content and starts being aesthetic. Four maps cleanly to the dimensions that actually move per-task cost: who does the work, who carries the accountability, how often is it audited, and what does the error cost.
Where this changes how you think
Three practical shifts come from holding the four classes in mind:
Career planning is about portfolio, not category. The question is not “is my role safe” (which assumes binary). It's “what does my class mix look like, and which classes do I want to grow into.” The reliable move is to bias time toward human-critical and human-led + AI-assisted, even within a role that started in the replaceable end of the spectrum.
Org design follows the distribution. A team that operates against a workload that's 60% replaceable will shrink in headcount but retain or grow in compensation per remaining seat. A team that operates against 60% human-critical work won't shrink at all and will become harder to staff, not easier. The org chart of 2028 looks different from 2024 not because total headcount halved but because the per-role mix shifted.
AI deployment ROI tracks the classes. Replaceable tasks generate fast, defensible ROI when automated. AI-augmented tasks generate productivity gains, not headcount savings — the ROI is real but it's a velocity story, not a cost story. Human-led + AI-assisted tasks generate small per-hour gains that don't justify a dedicated deployment project. Human-critical tasks have negative ROI on deployment — the AI inserts errors that the human now has to clean up. This is the discipline most failed AI rollouts skipped: they deployed against tasks that were not actually in class 1.
The full picture for your role
Wagecore computes the four-class distribution for any role you describe. The wizard takes about two minutes and the methodology is open at /methodology. You'll see exactly how your work splits across the classes, what operational cost AI would carry to do the replaceable portion, where your human advantage concentrates, and the headline substitution-exposure score derived from the mix. None of it is prediction. It's measurement against today's capability matrix, refreshed monthly.
If the framing here is useful, the related deeper read on the operational economics is Why operational AI cost is 3–10× what the demo shows — it picks up where this leaves off and walks through what it actually costs to deploy AI against a class-1 task in production.