Executive Summary
Your employer owns your job. That has always been true. What's changing — fast — is that AI is now doing significant parts of it.
Sixty-one percent of employees say their job description no longer matches what they actually do, according to SHRM's 2025 State of the Workforce report — the highest figure on record. Job titles were never built to survive this kind of change. The workers navigating it well aren't the ones who learned the most AI tools. They're the ones who figured out what they actually produce — the Work Unit — and built their identity around that instead of the role title on their contract.
What It Is
A Work Unit is a modular, outcome-oriented package of work that can be assigned, executed, measured, and improved as part of a larger value flow. Unlike a job description, which stabilizes responsibility over time, a Work Unit stabilizes the outcome and the execution logic while letting the mix of contributors evolve — human, AI, system, or some hybrid of the three.
It is not your job title, and it is not your list of responsibilities. For a financial analyst, the Work Unit isn't "financial reporting" — that's a task. It's investment-grade interpretation of incomplete data under time pressure: the judgment to know what's missing and what it means anyway. That judgment is what a job description was always a rough proxy for, and it's the part AI cannot yet do on its own.
Why It Matters
McKinsey estimates that 75% of knowledge work activities are now technically automatable. That doesn't mean 75% of jobs disappear — it means most of the tasks inside those jobs are now candidates for automation, while the roles that contained them stay nominally intact. Static job descriptions can't show anyone which 25% is left, or why it's the part that matters.
Microsoft's 2025 Work Trend Index found that 82% of business leaders believe AI will fundamentally change how work is done, but only 23% of employees feel prepared for that change. That gap isn't primarily a skills gap. It's a clarity gap. Workers who haven't named their Work Unit don't know what to protect, what to develop, or what to hand off — so they either resist AI adoption wholesale or hand over parts of the work they shouldn't. Microsoft WorkLab also found that organizations where employees design their own AI use cases see 3.4 times the productivity gains of top-down AI rollouts, because the people closest to the work unit are the ones best placed to see where automation helps and where it doesn't.
Core Components
A Work Unit is defined by four elements, and AI is not neutral across them — it is steadily absorbing two while leaving two stubbornly human.
Outcome definition — what result the unit is accountable for. AI can now draft plausible outcome statements from context, and increasingly proposes the target itself. This is the element most exposed to automation.
Execution logic — the steps, rules, and exceptions required to produce that outcome. This is exactly what generative AI is good at compressing: the repeatable, pattern-based playbook. Most AI adoption inside knowledge work happens here first.
Execution structure — who or what performs each part of the unit. This is where human judgment reasserts itself. Deciding which parts of a workflow AI should own, which need a human check, and where accountability sits if the output is wrong is itself a judgment call, and it's one that carries consequences for the person who makes it.
Learning loop — using performance signals to improve the unit over time. AI can surface the signals faster than any human review cycle. But deciding which signal actually matters, when a pattern is a fluke versus a real shift, and what the organization should change in response is context-dependent, relationship-laden, and carries accountability in a way a model's output does not.
Execution structure and the learning loop are where your work unit lives now. They are the layers that don't fully delegate, because delegating them means delegating the accountability that comes with them — and accountability doesn't transfer to a system.
How to Read the Framework
Read the four elements as a gradient, not a list — from most automatable to least. Outcome definition and execution logic sit closest to what generative AI already does well: drafting targets, compressing playbooks, producing a first pass. Execution structure and the learning loop sit at the other end, because they require someone to own the call when it matters — who does what, and what the organization should actually learn from what just happened.
The mistake is treating all four elements as equally exposed, or equally protected. A worker who assumes AI will eventually absorb everything overinvests in defending tasks that were never the real value. A worker who assumes none of it will touch their role underinvests in the parts of the job that are, in fact, disappearing fastest. The useful question isn't "will AI change my job" — it's "which of these four elements is mine to own, and which am I still doing by habit."
Practical Implications
For individual workers, the shift from task ownership to output ownership is the defining move. Workers who make it stop describing their days in terms of activities and start describing them in terms of what they're accountable for producing. That reframing changes how they use AI: not as a replacement for judgment, but as a way to clear the execution-logic layer so more time goes into execution structure and the learning loop — the parts of the work that were always theirs.
For managers and organizations, the same logic applies at scale. A team redesigned around Work Units — rather than static role descriptions — can reassign the automatable layers to AI without leaving anyone's accountability unclear, because the outcome and the judgment layer stay explicitly owned even as the execution mix changes. Teams that skip this step tend to get partial AI adoption with no coherent accountability model behind it: work gets faster, but nobody can say who owns the result when it's wrong.
Simple Application Prompt
Run these against your own role:
- What is the one output you produce that would be missing, or worse, without you specifically?
- Of the four Work Unit elements — outcome definition, execution logic, execution structure, learning loop — which have you already handed to AI, deliberately or by default?
- If something in your work goes wrong, whose judgment is actually being relied on to catch it?
- What would you take with you — not your access, not your title — if you left tomorrow?
