Every few weeks another study puts a percentage on it. So many percent of jobs are exposed to AI. The figure earns a headline, a round of arguments, and then nothing changes on Monday, because a number about all jobs everywhere tells you nothing about the job in front of you.
AI does not arrive as a verdict on your occupation. It arrives one task at a time. A role is a bundle of tasks, and AI touches some of them heavily, some a little, and many not at all. Until you break the job into its tasks, you are debating an average that describes no one.
That is the problem the three new tools are built to solve. Before the tools, the data.
The number in the headline is the wrong unit
The best public read on how AI is used at work is Anthropic's Economic Index, built from millions of anonymized Claude conversations mapped to the tasks that make up real occupations. It is not a survey of what people think AI might do. It is a record of what they actually asked it to do. A few findings are worth holding onto.
Augmentation leads automation. When you look at how people use AI on a task, the collaborative pattern of asking, refining, learning, and iterating shows up more often than the hand-it-off-completely pattern. On Claude.ai the split runs roughly 52 percent augmentation to 45 percent automation. The base case for most work is a person working with AI, not a person replaced by it.
Coverage is wide but partial. Roughly half of jobs have a meaningful share of tasks, about a quarter or more, where AI can already contribute. That is a lot of work touched. It is not a lot of work finished.
A large share of jobs is barely touched. Close to a third of occupations show almost no measurable AI exposure, mostly the physical, in-person, and hands-on work that a language model cannot reach.
Exposure clusters by the kind of task. Software and writing tasks show high coverage. A programmer's task set can run around three-quarters covered and data entry around two-thirds. Physical and relational tasks run low. The dividing line is the type of work, not the prestige of the title.
None of that is a forecast of employment. Usage is not the same as job loss. The data tells you where AI can help today, not what any given employer will decide to do about it. That distinction is the whole game, and it is exactly the distinction the headline number erases.
Averages hide the thing you actually need to see
Two people can sit inside the same reassuring or scary percentage and be in completely different situations. One spends the day on tasks AI covers heavily and has real reason to reshape how they work. The other spends it on judgment calls and in-person relationships that the same tools barely touch. The average that contains them both is true and useless at the same time.
A business works the same way. Your exposure is not a single company-wide figure. It is the weighted sum of the roles you actually employ, and two firms in the same industry can end up in very different places depending on how their headcount is distributed. You cannot manage what you have only measured at the top line.
Three tools, three questions
We built the tools to answer the three questions people actually ask, each at the level the data can support.
Is my job exposed to AI? The job exposure tool lets you pick an occupation and see how much of its task set AI can already contribute to, and whether the pattern skews toward augmentation or automation. It is a read on the work, not a prediction about your position.
What does AI mean for my team or business? The business impact tool takes the shape of your team and returns a directional picture of where AI is most likely to assist, where it might handle work end to end, and where it will not reach. It points to where the work could shift. It is not a headcount number and not a promise.
Should this task be augmented or automated? The augment-or-automate tool takes a specific task and classifies it: keep a person in the loop, hand it off, or leave it alone. This is the decision that actually moves the needle, and it gets made one task at a time.
All three live at atlasinstinct.com/ai-impact. They are free, they need no login, and they store nothing.
How to read the results without fooling yourself
Tools like these are only useful if you hold them at the right confidence. A few rules keep the readings honest.
Directional, not deterministic. The estimates describe patterns across a large dataset, not a measurement of your specific situation. Use them to aim attention, not to settle a decision on their own.
Coverage is not automation. A task AI can help with is not a task that runs itself. High coverage tells you there is leverage available. What you do with it is a separate choice.
It is a snapshot of a fast-moving field. The underlying usage reflects how people work with today's models. Both the models and the habits are moving, so treat the picture as current, not permanent.
Augmentation is the base case. Because the collaborative pattern shows up more often than full handoff, the honest default reading of wide exposure is work getting assisted, not work getting erased. The scary interpretation is the less likely one, not the more likely one.
The tools describe capability and behavior, not anyone's plans. They cannot see your employer's budget, strategy, or intentions. Verify every reading against your own reality before you act on it.
What a serious leader does with this
The unserious move is to react to the headline, announce a number, and unsettle a workforce over an average. The serious move is quieter and more useful.
Map your real tasks, not your org chart. Decide augment or automate deliberately, case by case, rather than letting the loudest vendor decide for you. Invest first in the augmentation cases, because they usually deliver faster, carry less risk, and cause less disruption than a full automation project. Reskill toward the judgment, relationship, and in-person work that the data shows staying stubbornly low-exposure. And treat the whole thing as an operating-model question, not an HR event.
That is the same discipline behind our AI Readiness work. The tools are the free, self-serve front door to it. They will not make the decisions for you, but they will make sure you are arguing about the right unit.
Frequently asked
Do these tools predict whether I will lose my job?
No. They show how much of a job's task set AI can already contribute to, and whether the usage pattern skews toward augmentation or automation. That is a read on the work, not a forecast of any employer's staffing decisions. Usage is not the same as job loss.
What data are the tools based on?
Anthropic's Economic Index, the largest public dataset on how AI is actually used at work, built from millions of anonymized Claude conversations mapped to the tasks that make up real occupations. Directional estimates in the tools are labeled as such.
Does high AI coverage mean a job gets automated?
No. Coverage means AI can contribute to a share of the tasks. A task AI can help with is not a task that runs itself. On the underlying data the collaborative augmentation pattern shows up more often than full automation, so wide coverage more often points to work being assisted than removed.
Is Atlas Instinct affiliated with Anthropic?
No. Atlas Instinct is an independent firm. It is not affiliated with, endorsed by, or sponsored by Anthropic. The figures in the tools are cited from Anthropic's public Economic Index research, and directional estimates are labeled as such.
Do the tools store my inputs?
No. The tools run in your browser, need no login, and store nothing.
Atlas Instinct is an independent firm. It is not affiliated with, endorsed by, or sponsored by Anthropic. Figures referenced here are cited from Anthropic's public Economic Index research, and directional estimates are labeled as such.