Measure the Gap, Not the Job
The OECD AI Capability Gap Index is the subject here: a way to compare what an occupation requires with what current AI systems can do across several kinds of human ability. This chapter explains the comparison itself—its nine domains, its weighting logic, and its composite score—so that a reader can interpret an exposure number as a distance to a capability frontier rather than as a label attached to an entire job.
The basic idea is straightforward. Every occupation depends on a profile of capabilities, and AI systems have their own uneven profile. If an occupation requires a capability at a level that AI has already reached, the remaining gap in that domain is zero. If the occupation requires more, the difference is a residual gap. Add those gaps carefully, and one obtains an indicator of potential exposure. The care matters because the result is only as meaningful as the dimensions being compared and the importance assigned to each one.
Nine ways of being capable at work
The OECD framework begins by refusing to treat “AI capability” as a single substance. It separates Language, Social Interaction, Problem Solving, Creativity, Metacognition and Critical Thinking, Knowledge, Learning and Memory, Vision, Manipulation, and Robotic Intelligence. The first six describe forms of cognitive or social performance; the final three capture perception, physical handling, and the ability to operate in an environment through coordinated action.
This vocabulary changes the question. A language model may be close to the requirements of a role that involves drafting, summarising, classifying, or retrieving information. That same system may be much farther from the requirements of a role in which the decisive work is persuading a distressed person, making a context-sensitive judgement, handling an unpredictable physical environment, or coordinating fine movements. Calling both roles “exposed to AI” without identifying the relevant capability hides the mechanism that makes the comparison useful.
The domains are not nine independent departments inside a job. They overlap in real work. A manager may need language to explain a decision, problem solving to weigh alternatives, metacognition to assess whether the decision is sound, and social interaction to bring other people along. A nurse may combine knowledge, perception, manipulation, judgement, and interpersonal understanding. A production worker may use vision and physical coordination while still making decisions when a process deviates from its usual pattern. The point of the framework is not to pretend that these capabilities occur separately, but to make their different contributions visible before combining them.
That breadth also prevents a familiar mistake in exposure analysis. If the only measured frontier is the one associated with language and information processing, routine clerical work will be represented reasonably well, while physical or socially situated work may appear less exposed simply because important capabilities were not included. A broad capability framework can ask a more complete question: how close is AI to the configuration of abilities this occupation requires, not merely to the most visible type of task it contains?
The occupation side of the comparison is grounded in O*NET descriptors. For example, the Language indicator draws on oral and written comprehension and expression. Social Interaction draws on capabilities and skills such as coordination, instructing, negotiation, persuasion, service orientation, and social perceptiveness. Problem Solving combines several reasoning and decision-related descriptors. The same logic is used for the remaining domains: multiple occupational measures are brought together rather than treating one isolated skill as the whole requirement.
This is important for interpretation. The index is not asking whether a tool can perform one activity that appears somewhere in a job description. It is estimating how much of the occupation’s measured capability profile lies close to the current frontier. That is a more demanding question than “Can AI help with this task?” and a less sweeping one than “Can AI do this job?”
The flow runs from two inputs to a summary, but it should not be read as a machine that produces a job verdict. Occupational demand and AI capability are first placed on comparable scales. The remaining distance is then weighted according to how important that capability is to the occupation. Only after those steps is a total calculated. The total is a convenient summary of a multidimensional profile, not a replacement for the profile.
From a domain gap to a weighted total
For a capability domain i and occupation j, the logic can be stated in plain language: take the level the occupation needs, subtract the level AI has reached, and treat a negative result as zero. If an occupation requires Language at level 2 and current AI is assessed at level 3, the language gap is zero. If it requires Vision at level 5 and AI is assessed at level 3, the vision gap is two. The zero does not mean that the occupation has disappeared; it means only that this particular capability requirement is no longer the measured bottleneck.
The OECD’s worked example makes the asymmetry clear. In the example occupation, AI meets the requirements for Language, Social Interaction, Creativity, Knowledge, Learning and Memory, and Robotic Intelligence. Gaps remain in Problem Solving and Metacognition, each by one level, and in Vision and Manipulation, each by two levels. If every domain counted equally, that would already be a useful profile. But occupations do not depend equally on every capability, so the index applies importance values as well.
The importance values are derived by linking each AI indicator to relevant O*NET abilities or skills and synthesising those measures into a standardised score. In the example, the remaining Problem Solving gap is multiplied by an importance value of 0.6, the Metacognition gap by 0.5, the Vision gap by 0.7, and the Manipulation gap by 0.8. The weighted contributions are therefore 0.6, 0.5, 1.4, and 1.6, producing a total gap index of 4.1. The arithmetic is not the difficult part. Its value is that it prevents a capability that is peripheral to a role from counting as much as one central to carrying it out.
Notice what the importance weight does—and what it does not do. It says that a capability is more or less central to the occupation’s measured performance requirements. It does not say that the capability is easy or difficult to automate, that workers spend a particular share of their day on it, or that an employer will permit a system to take it over. A highly important capability with a large remaining gap can be a strong constraint on whole-role substitution while still leaving room for assistance elsewhere. A low-importance capability with a zero gap can be technologically tractable without being economically consequential.
This distinction becomes clearer when two occupations have the same total. Suppose one has a large residual gap in Manipulation and a small gap in Language, while another has a large residual gap in Social Interaction and a small gap in Vision. Their totals might be similar, but the practical questions would not be. The first points toward dexterity, equipment, and adaptation in a physical setting; the second toward context, trust, interpretation, and responsibility. A manager testing tools, a worker deciding which tasks to monitor, and a policymaker considering training would need different evidence in each case. The composite can put the occupations on a common comparison scale, but it cannot make their remaining problems interchangeable.
The measure also describes occupational requirements, not the capability of every person holding the job. A score is not an assessment of whether an individual worker is intelligent, creative, socially skilled, or physically able. Nor does a high occupational demand imply that every task or every work setting calls on the highest level represented in the profile. It is a structured approximation of what the occupation, as currently defined, requires. That distinction matters for career decisions: a worker may move toward a role whose exposed tasks are likely to be assisted while developing complementary judgement or domain knowledge that the score does not capture as a personal guarantee.
Consider what would happen without the weights. A job that occasionally requires visual checking could look as constrained as a job whose defining activity is operating in a visually complex environment. A role that uses conversation only to exchange routine information could appear equivalent to one in which trust, persuasion, or emotional judgement is the work. Importance weighting does not solve every problem in occupational measurement, but it makes the comparison more faithful to the way roles are structured.
The composite still has limits. The underlying scales are judgement-based, and a one-level movement does not necessarily represent the same substantive distance in every domain. Moving from one level to the next in Language may not mean the same thing as moving one level in Manipulation or Metacognition. The total also compresses unlike forms of distance. Two occupations can have the same overall index while one has a remaining social-interaction gap and the other a remaining physical one. Their likely technological paths, training needs, and reasons for retaining human responsibility could be very different.
Some domains also overlap conceptually. Problem solving, metacognition, language, social understanding, and robotic intelligence can each touch aspects of judgement or context. Adding them can partly count related requirements more than once in some occupations. That is not a reason to discard the index. It is a reason to treat the total as a comparative summary and inspect the domain-level contributions before drawing a conclusion.
The same caution applies to the range of the score. A theoretical maximum based on the scales is not the same as the range that can arise under current AI capability levels and realistic occupational importance values. Current systems already perform above the bottom of the capability scales in every domain, and occupations rarely assign maximum importance to every capability at once. A number therefore has meaning relative to the construction of the index and the comparison set; it is not a universal unit of “replaceability.”
A zero gap deserves special attention because it is the easiest result to overread. It says that current assessed AI capability is at or above the requirement in that domain. It does not say that AI has been adopted in the workplace, that all tasks in the occupation are covered, or that substitution is economically or institutionally feasible. A medical transcription role may have very small gaps in structured language and information handling while still involving quality control, exceptions, accountability, or workflow decisions that the occupation-level score does not spell out. A judge may have some language tasks close to AI while the broader role retains gaps in contextual reasoning, responsibility, and interpretation.
Conversely, a large total gap is not a promise of stability. It can coexist with meaningful assistance in the domains where AI is already capable. A healthcare professional may use a system for documentation or information retrieval even if the full role remains far from the frontier because it also requires social interaction, judgement, physical examination, and action in changing circumstances. The index identifies the distance that remains in the measured profile; it does not predict how a workplace will partition tasks around that distance.
This is why the phrase “measure the gap, not the job” is more than a warning about terminology. It identifies the proper unit of interpretation. The gap is between a capability demand and a capability level. The job is a social and economic arrangement that bundles tasks, responsibilities, tools, incentives, and rules. The first can be represented by a transparent calculation. The second changes only through choices and processes that the calculation does not observe.
For a worker deciding where to build durable value, the score can indicate which parts of a role deserve closer inspection: routine information processing, structured analysis, perceptual inspection, physical manipulation, or socially accountable judgement. For a manager, it can suggest where to test assistance and where to preserve review, discretion, or human responsibility. For a policymaker, it can point toward occupations whose capability profiles may need task-level and training analysis. None of those uses requires treating the index as a prediction of employment outcomes.
The next interpretive problem is therefore not mathematical. It is methodological: who decided what each occupation requires, how consistently those judgements were made, and how a map covering hundreds of occupations can be extended without disguising uncertainty as precision. The calculation is explicit. The occupation-side inputs are where the map’s human and model judgements become consequential.