Jobs Are Not the Unit of Work (And AI Vendors Know It)

Clu Labs
Jun 1
3 min read

Every AI displacement study you've read is probably measuring the wrong thing.

When a research paper says '80% of legal work is exposed to AI,' what it means, if it's being precise, is that a subset of tasks currently performed by people with the job title 'lawyer' or 'paralegal' are automatable. Document review. Precedent search. First-draft contract generation. Those tasks are real, and AI is genuinely reshaping them.

But the prediction consistently fails to account for the context in which those tasks are embedded.

The problem with jobs as a unit of measurement

Jobs are administrative containers. They're the payroll and legal fiction that bundles together dozens of distinct tasks, skills, and judgments. They're the org chart's best approximation of what a person does, and as anyone who has run a serious capability audit knows, that approximation is frequently outdated, incomplete, and politically shaped before AI enters the picture.

Skills are not isolated. They cluster. They depend on each other. A paralegal conducting document review is simultaneously reading for context, noticing anomalies that don't fit the expected pattern, making quiet judgment calls about what to escalate and what to let go, and building the institutional knowledge that makes the next case faster and better.

The document review task is visible and measurable. The surrounding skills - observation, contextual reasoning, risk sensing, escalation judgement - are far less so.

What the research actually shows

MIT research has found that AI can perform around 14% of occupational tasks across the US economy; a figure less than half of Goldman Sachs' more optimistic estimates, and a fraction of the headlines those estimates generated.

The gap between 'theoretically automatable' and 'actually automated at scale' is not a technology problem. It's a structural problem. Organisations don't know, at the task level, what their people actually do, which means they can't identify which tasks are genuine automation candidates, which skills sit underneath those tasks, and what the risk exposure is when those skills change.

You can automate the task. You cannot yet automate the judgment that determines whether the automated output is trustworthy, complete, or safe to act on. And the more you automate, the more that judgment work expands.

What a task-first approach changes

At Clu, we build workforce intelligence from the task layer up, not the job layer down. Our task fingerprint captures six dimensions of every unit of work: the action being performed, the object being acted upon, the outcome produced, the domain it sits in, the work type (cognitive, collaborative, operational), and the reach of the decision.

This matters for AI workforce planning because augmentability, the degree to which a task can be meaningfully supported or replaced by AI, is a property of the task, not the job. Two people with the same job title may carry radically different task profiles. One role is 60% automatable. The other is 20%. A job-level analysis tells you neither.

When our clients run a diagnostic, the finding that most consistently surprises their ExCo isn't 'AI is coming for our people.' It's 'we didn't know what our people were actually doing.' That structural blind spot is the real risk, not the model on the other side of the API.

The practical implications for leaders using AI for workforce planning

Before your organisation commits to an AI transformation programme, ask a harder question than 'which roles are exposed?' Ask: 'Do we know, at the task level, what those roles actually contain?' If the answer is no, and for most organisations, it is, then any AI impact assessment you commission is an estimate built on administrative fiction.

The organisations making defensible workforce decisions right now are the ones who've done the structural work first.

They know what their people do. They know which tasks are candidates for automation or augmentation. They know where the cognitive and relational work lives that AI cannot yet touch. And they know, when something changes, what the second and third-order effects are.

That's not a technology advantage. It's a structural one. And it's available before you spend a pound on AI.