AI companionship is already widespread though I’m very curious to see how it appears at the task level since it doesn’t map to an occupation. Looking forward to your article!
The interesting part is that this data can be useful even if you are not trying to predict job loss.
Knowing which tasks are already being automated vs augmented can help you decide what to learn, what to build, and even which workflows are worth experimenting with.
The task-vs-title split does more work than it looks like at first. It also gives you a way to route decisions, not just learn from them: if a task in your own stack is directive plus a tight feedback loop, that is a deterministic problem, hand it to a workflow tool. If it needs iteration, learning, and judgment along the way, that is an agent or a reasoning call, not a script. The tool decision follows the collaboration pattern, not the other way around, which is usually where people guess wrong first.
Observed use tells us where people reach for help, but it doesn't tell us whether the task deserves to survive. Before choosing a pilot, I'd ask who checks the output and what decision changes when it's right. If neither has an answer, automating the disliked task may only make unowned work arrive faster.
True, yet it slightly misses the scope of this, the Index data doesn't necessarily mean any task should become a pilot project for a team starting to implement AI or a product
Dieter's right that the ownership question comes before the automation call, not after. We'd add one thing to it: naming a checker isn't enough if nothing ever forces a failure to surface. We've seen automations run for months with someone technically on the hook who never once looked at a bad run, because there was no log making the bad run visible.
The task worth automating is the one where a wrong output actually breaks something people notice, not just the one everyone already avoided doing.
Daria, having the Economic Index searchable inside Claude changes it from something you read about to something you can actually interrogate for your own role, which is a much more useful thing. I'd be interested in what it says about legal and financial services specifically, since that is where most of my clients keep asking what is coming for them.
I dodgy get any results for the mindfulness industry but did help me find interesting data for a research project we are working on the mindful use of AI in business schools. Thank you for starting these ways to use this feature. I'm still exploring the other uses and will share what I find.
Oh, that's because they correlate it with jobs. I see that in ONET they vaguely associate it with the "27-2099.00 Entertainers and Performers, Sports and Related Workers, All Other" occupation: https://www.onetonline.org/find/result?s=mindfulness
I like the task-vs-title distinction. Marketers writing light Python scripts, for example. Classify that by task alone and it counts as developer automation, which isn't what's happening.
Yep, exactly. I’d be very curious to see how the tasks associated with each job title change in ONET over the years as more roles take on technical work and incorporate AI into what they do.
That is exactly the gap task-only classification leaves open. The Index also tags a collaboration pattern on top of the task, directive, feedback-loop, task iteration, learning, validation, which looks at how the work actually happened, not just which bucket the task falls into. A marketer running one directive script and a developer running an iterative build loop can land in the same task category and still be doing very different work.
Worth checking the pattern label too before trusting the automation-vs-augmentation split.
Thanks for this, I just used it to check on which finance jobs people hand over to AI now. IIt is not surprising to see that the handle turning jobs to produce the report/docs are going. BUT choosing how to produce it has not moved.
Es un enfoque interesante. Analizar la adopción de la IA en función de tareas concretas, en lugar de por puestos de trabajo, hace que el debate resulte mucho más útil. La verdadera pregunta no es qué puestos de trabajo desaparecerán, sino qué aspectos de los mismos cambiarán primero.
Looking at how Claude is actually used across occupations is useful data. The more interesting question is which version of the tool different groups get. The public systems still refuse entire subjects, return inconsistent answers, and force the informed user to rewrite large sections. Cleaner capabilities remain reserved. I mapped the longer pattern of that control here:
I wasn't sure if I would get numbers on AI companionship, but I did, and an article will be coming soon! Thanks, Daria!!
AI companionship is already widespread though I’m very curious to see how it appears at the task level since it doesn’t map to an occupation. Looking forward to your article!
It is a rudimentary look at the data, but it is rather telling. Thanks again for sharing the index; otherwise, I would never have known about it!! https://www.aibutintimate.com/p/i-asked-claude-to-prove-ai-wants
The interesting part is that this data can be useful even if you are not trying to predict job loss.
Knowing which tasks are already being automated vs augmented can help you decide what to learn, what to build, and even which workflows are worth experimenting with.
for sure! that's why all the use cases I shared, the data can be used in so many ways
The task-vs-title split does more work than it looks like at first. It also gives you a way to route decisions, not just learn from them: if a task in your own stack is directive plus a tight feedback loop, that is a deterministic problem, hand it to a workflow tool. If it needs iteration, learning, and judgment along the way, that is an agent or a reasoning call, not a script. The tool decision follows the collaboration pattern, not the other way around, which is usually where people guess wrong first.
Super interesting and thank you for breaking it down by examples, Daria! Love that Anthropic connected it to Claude for convenience.
Way easier to access it from inside Claude & glad the examples helped 🤗
Observed use tells us where people reach for help, but it doesn't tell us whether the task deserves to survive. Before choosing a pilot, I'd ask who checks the output and what decision changes when it's right. If neither has an answer, automating the disliked task may only make unowned work arrive faster.
True, yet it slightly misses the scope of this, the Index data doesn't necessarily mean any task should become a pilot project for a team starting to implement AI or a product
Dieter's right that the ownership question comes before the automation call, not after. We'd add one thing to it: naming a checker isn't enough if nothing ever forces a failure to surface. We've seen automations run for months with someone technically on the hook who never once looked at a bad run, because there was no log making the bad run visible.
The task worth automating is the one where a wrong output actually breaks something people notice, not just the one everyone already avoided doing.
Great practical breakdown using real AI usage data at the task level makes the future of work much easier to understand.
Daria, having the Economic Index searchable inside Claude changes it from something you read about to something you can actually interrogate for your own role, which is a much more useful thing. I'd be interested in what it says about legal and financial services specifically, since that is where most of my clients keep asking what is coming for them.
Thanks, Mel, I'm sure you'll find lots of data in there about legal and financial services. And btw, Google also recently published a similar report if you want to look into that one too https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/
I dodgy get any results for the mindfulness industry but did help me find interesting data for a research project we are working on the mindful use of AI in business schools. Thank you for starting these ways to use this feature. I'm still exploring the other uses and will share what I find.
Oh, that's because they correlate it with jobs. I see that in ONET they vaguely associate it with the "27-2099.00 Entertainers and Performers, Sports and Related Workers, All Other" occupation: https://www.onetonline.org/find/result?s=mindfulness
Maybe you can ask about that one
Appreciate you looking for it. I'll check it out.
I like the task-vs-title distinction. Marketers writing light Python scripts, for example. Classify that by task alone and it counts as developer automation, which isn't what's happening.
Yep, exactly. I’d be very curious to see how the tasks associated with each job title change in ONET over the years as more roles take on technical work and incorporate AI into what they do.
That is exactly the gap task-only classification leaves open. The Index also tags a collaboration pattern on top of the task, directive, feedback-loop, task iteration, learning, validation, which looks at how the work actually happened, not just which bucket the task falls into. A marketer running one directive script and a developer running an iterative build loop can land in the same task category and still be doing very different work.
Worth checking the pattern label too before trusting the automation-vs-augmentation split.
Thanks for this, I just used it to check on which finance jobs people hand over to AI now. IIt is not surprising to see that the handle turning jobs to produce the report/docs are going. BUT choosing how to produce it has not moved.
A good example of what human-AI collaboration looks like in finance. Btw you might find these tools useful for finance research: https://aiblewmymind.substack.com/p/ai-finance-research-workflows
Es un enfoque interesante. Analizar la adopción de la IA en función de tareas concretas, en lugar de por puestos de trabajo, hace que el debate resulte mucho más útil. La verdadera pregunta no es qué puestos de trabajo desaparecerán, sino qué aspectos de los mismos cambiarán primero.
Looking at how Claude is actually used across occupations is useful data. The more interesting question is which version of the tool different groups get. The public systems still refuse entire subjects, return inconsistent answers, and force the informed user to rewrite large sections. Cleaner capabilities remain reserved. I mapped the longer pattern of that control here:
https://sleuthfox.substack.com/p/why-your-ai-always-leaves-fingerprints