An AI job risk calculator that hands you a percentage and calls it your probability of being replaced is doing something it cannot do. No published model predicts an individual's employment outcome, and presenting a score as though it did would be a guess dressed up as analysis. This page therefore does something narrower and defensible: it builds a task-exposure profile from a description of your own week, using the framework that automation research actually uses, and shows its working.
Arb Digital publishes a free tools library, and this page sits in its AI section next to the automation savings calculator, which prices the hours and money a specific process would reclaim if automated. That tool is about a process and its economics. This one is about the composition of a role, and it deliberately produces no financial figure at all.
What Exposure Actually Means
The distinction this page rests on comes from the research itself, not from caution on our part.
The most widely cited task-level study of language models, GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models by Eloundou, Manning, Mishkin and Rock, estimates that around 80% of the US workforce could have at least 10% of their work tasks affected by language models, and roughly 19% of workers could see at least 50% of their tasks affected. Those are large numbers, and they are frequently reported as though they were forecasts of job losses. They are not. The authors are explicit that they measure exposure — whether a model could reduce the time taken to complete a task at equivalent quality — and that they make no predictions about development or adoption timelines.
The gap between those two things is enormous and is where almost all of the practical outcome lives. A task can be exposed and never automated, because the tooling never gets built, because verifying the output costs more than doing the work, because a regulator requires a named human, because the organisation never reorganises around it, or simply because the work was never the bottleneck. Exposure is a measurement of technical overlap. Employment is a consequence of economics, regulation, institutions and management decisions, none of which a task-overlap score contains.
It is also worth knowing that studies in this field disagree with each other. Earlier occupation-level work produced very different headline figures from later task-level work, largely because the unit of analysis changed from whole jobs to individual tasks, and the choice of unit changes the answer more than the technology does. Treat any single number in this area — including the one this page produces — as one framing among several.
What This AI Job Risk Calculator Does
It asks you to split your working week across five kinds of task, applies a published-research-informed exposure weight to each, and then adjusts the result for two factors that consistently keep exposed work with people.
The five categories are deliberately coarse, because fine-grained self-assessment is unreliable. The two adjustments are accountability — how far a named human must legally or professionally sign the output — and verification cost, meaning how expensive an undetected error is. Both reduce effective exposure, and both are the reason a great deal of technically exposed work has not moved.
The bars show which categories contribute the exposure, which is the genuinely useful output. A profile that is 47% exposed because of routine administration is a different situation from one that is 47% exposed because of drafting and analysis, and they point toward different responses.
How to Use It
- Describe last month, not your job title. Titles are almost useless here. Two people with the same title can have completely different task mixes, which is precisely why task-level analysis replaced occupation-level analysis.
- Make the five figures add to 100. The tool normalises if they do not, but an accurate split gives a more honest profile.
- Be strict about the creative category. Work with a known good answer is not creative judgment even if it feels skilled. Reserve it for decisions where reasonable experts would differ.
- Set the two adjustments from your industry, not your preference. Regulated professional work sits high on accountability; internal drafting usually sits low.
- Read the bars, then ignore the headline. The composition tells you something actionable. The single percentage does not.
The Formula / How It's Calculated
The weights below are this page's own, chosen to reflect the ordering that task-level research consistently finds rather than to reproduce any specific study's coefficients. They are published here so you can disagree with them, which you should feel free to do.
Routine cognitive and administrative — weight 0.90. Writing, research, coding and analysis — 0.75. Creative and expert judgment — 0.45. Interpersonal, care and negotiation — 0.20. Physical, manual and on-site — 0.10.
Unadjusted exposure is the weighted sum: Σ (share of time × weight), expressed as a percentage. The two adjustments then apply multiplicatively: adjusted = unadjusted × (1 − 0.20 × accountability ÷ 10) × (1 − 0.15 × error cost ÷ 10). At their maximums the two together reduce exposure by 32%, which is a deliberate design choice — accountability and verification cost slow adoption substantially but do not eliminate it, and a model that let them drive exposure to zero would be as misleading as one that ignored them.
Worked example, matching the values this page loads with. A week split 30% routine, 25% writing and analysis, 15% creative judgment, 20% interpersonal and 10% physical. The weighted sum is (0.30 × 0.90) + (0.25 × 0.75) + (0.15 × 0.45) + (0.20 × 0.20) + (0.10 × 0.10) = 0.270 + 0.1875 + 0.0675 + 0.040 + 0.010 = 0.575, or 57.5% unadjusted exposure. With accountability at 5 the first multiplier is 0.90, giving 0.5175, and with error cost at 6 the second is 0.91, giving 47.1% adjusted task exposure. Routine administration alone contributes 27 of those 57.5 points, which is the finding that matters — not the number itself.
Why Task Composition Beats Job Title
The most useful thing this exercise does is force you to look at your week as a set of tasks rather than as a role, and that shift explains most of what is confusing about public discussion of AI and work.
Occupations are bundles. The US Department of Labor's O*NET OnLine database, which underpins a great deal of this research, describes over 900 occupations in terms of their detailed work activities precisely because the occupation label carries so little information on its own. A single job title can contain tasks with exposure weights spanning the full range, and the mix varies enormously between two people doing the same nominal job at different employers.
This is why headlines about "the most at-risk jobs" mislead so reliably. They rank titles, and titles are not the unit at which anything happens. What actually happens is that particular tasks get faster, the role reorganises around the tasks that did not, and the job either absorbs the change or does not depending on factors specific to the organisation.
It also explains an effect that surprises people: task-level research consistently finds that higher-paid knowledge work often shows greater exposure to language models than lower-paid manual work, which is the opposite of what earlier automation waves produced. Writing, summarising and analysis are exactly what these systems do; unloading a van is not.
What Reduces Exposure, and What Only Feels Like It Does
Three things genuinely reduce the effective exposure of work, and two commonly cited ones do not.
Verification cost genuinely reduces it. When checking an output takes nearly as long as producing it, automation saves little. This is the single strongest brake in practice, and it is why exposed tasks in high-stakes domains have moved so slowly.
Accountability genuinely reduces it. Where a named person must sign, the person remains in the loop even if a model does most of the work. The task changes shape rather than disappearing.
Context that lives outside any document genuinely reduces it. Work that depends on knowing this organisation, this client, this history is hard to hand over, because the inputs were never written down.
Two things that feel protective but are not: difficulty — hard-for-humans and hard-for-models are different axes, and some genuinely demanding cognitive work sits high on exposure — and seniority, since senior roles often contain a great deal of drafting, summarising and review, which are among the most exposed activities of all.
If you want to look at this from the organisation's side rather than the individual's, the automation savings calculator quantifies the hours a specific process would reclaim, and the AI vs human cost calculator compares per-task cost including oversight time — which is where verification cost stops being an abstraction and becomes a line item.
Arb Digital designs and builds fast, dependency-free web interfaces around models — dashboards, calculators and tools that load instantly and rank.
See Web Design Services Talk to Arb DigitalCommon Mistakes to Avoid
- Reading the percentage as a probability — it is a share of tasks with technical overlap, not a likelihood of anything happening to you.
- Scoring your job title instead of your week — titles bundle tasks with wildly different exposure, which is why task-level analysis exists at all.
- Treating one study's headline as settled — occupation-level and task-level work produce very different figures, and the choice of unit changes the answer more than the technology does.
- Assuming difficulty protects you — hard-for-humans and hard-for-models are unrelated axes, and demanding analytical work is often highly exposed.
- Ignoring verification cost — where checking an answer costs as much as producing it, exposure rarely converts into adoption.
Related Free Tools From Arb Digital
The automation savings calculator prices the hours a specific process would reclaim, the AI vs human cost calculator compares per-task cost including oversight, and the AI ROI calculator covers payback on a whole project. The percentage calculator handles the supporting arithmetic if you want to recompute the weighted sum by hand. Everything else is in the free online tools hub.
Frequently Asked Questions
No, and no tool can. It measures task exposure — the share of your week made up of tasks that language models can currently assist with or perform. Whether a job continues, changes or ends depends on economics, regulation, management decisions and adoption timelines, none of which a task-overlap score contains.
Exposure means a model could plausibly reduce the time a task takes at equivalent quality. Replacement means an organisation actually restructures work and removes a role. Most exposed tasks never make that journey, because tooling is never built, verification costs too much, accountability requires a person, or the work was never the constraint.
They are this page's own, chosen to reflect the ordering that task-level research consistently reports rather than to reproduce any specific study's coefficients. They are printed in the formula section so that you can examine and disagree with them, which is the only honest way to publish a scoring model of this kind.
Mostly because they use different units of analysis. Occupation-level studies rank whole jobs and task-level studies rank individual activities, and the choice of unit changes the headline figure more than the underlying technology does. Different definitions of exposure and different time horizons widen the gap further.
Because language models are good at exactly the activities that fill knowledge work — writing, summarising, translating, drafting code and first-pass analysis — and poor at unstructured physical tasks. This reverses the pattern of earlier automation waves, which displaced routine manual work first.
High verification cost, formal accountability requiring a named human signature, and dependence on context that was never written down anywhere. Difficulty and seniority do not, because hard-for-humans and hard-for-models are unrelated, and senior roles often contain a great deal of drafting and review.
Use the composition rather than the total. Knowing that most of your exposure sits in one category tells you where a tool would change your week and where your time is best spent, which is a planning input. The single percentage is not a decision-making figure, and this page does not offer career advice.
This tool is an educational task-composition exercise, not a forecast, an assessment or career advice. It does not predict employment outcomes for any individual, no published model does so, and decisions about your career should be made with people who know your circumstances.
