A job is a list of responsibilities. We look at each one.
AI already does some parts of a job well. Other parts still need a person. HumanLens splits a job into its parts, checks each one, and adds them back up.
Every responsibility, sortedList the job's responsibilities
We start from the U.S. Department of Labor's official list for that kind of job. An executive assistant has 22, from managing the calendar to greeting visitors. If you add a job posting, the parts it stresses count for more.
Ask 4 questions about each one
You'll find them below. Every answer comes with a reason you can read.
Sort each one into 3 groups
AI can do it: AI handles most of it and gets it right. AI helps: AI takes on part of it, or most of it with a person checking the result. Needs a person: AI can't do much of it yet.
Add it up
You see how much of the job's time is on each side, and what handing work to AI really costs once you count the time someone spends checking it.
The 4 questions
Can AI do it?
Can today's AI do this as well as a capable worker? That includes AI agents, which work on their own over many steps with the access a person in the job normally has, like email, calendar, files and the company's software.
Does it get it right?
How often is AI right, and how easy is it to spot its mistakes? An agent working alone over many steps can pile up errors, so long, judgment-heavy work scores lower here.
Must a person sign for it?
Does a law, a license, a signature or being there in person require a human?
Does someone need a person?
Do customers, clients or coworkers need to deal with a real person?
2 responsibilities from the same job
Both belong to an executive assistant. They end up on opposite sides.
What you get
- For a company checking a job: how much of the time still needs a person, which responsibilities fall on each side, and what you'd save each year by handing the AI side to software. For a typical executive assistant, AI could take on about 62% of the time, and 38% still needs a person.
- For a job seeker: a Human Edge score from 0 to 100. It shows how much of the whole job your experience holds in the parts that still need a person, so a job AI could mostly take on can't score high, however strong the résumé. Under it, coverage shows how much of that human part your résumé covers. It looks only at the work on your résumé, never at who you are. With your email you also get a free plan: why you got your score, what to lean into, what to learn next with a step for each, roles your experience fits, and a private link to re-check it as the scores move.
- What the number carries. Human Edge already counts the job's split. Someone who covers all of the human side of a job that is 60% AI scores 40 at most, not 100. A high score means both that the job still needs a person and that your experience is in that part.
Good to know
- This is an early preview. We add new jobs every weekday, based on what employers are posting. Anyone who checks a job can tell us whether they agree with each score, and the scores will shift as that feedback comes in. When a responsibility could go either way, we mark it as a close call.
- AI gets better every month, so we update the scores every month. Every result shows which version it used.
- HumanLens helps people decide. It never rejects, ranks or picks anyone on its own.
The scores move as AI gets stronger.
A score from last year is already out of date. Every month we update the scores using the evidence below, and we note where each change came from.
- How people really use AI at work. When Anthropic publishes new data on how often AI succeeds at real work, the scores move with it. Real use beats predictions.
- How fast AI is improving. In lab tests, AI can finish longer and longer pieces of work on its own. Real jobs change much more slowly than lab tests, so we count only about a tenth of that progress.
- What people tell us. If people who do or hire for a job keep disagreeing with a score, a person reviews it. It never changes on its own.
- What never changes on its own. Legal rules, licenses and the need for a trusted person in the room change only when the law or the job does. We update those by hand.
The full working
The math behind the scores and where the data comes from. Every result also explains each of its own scores.
The formulas
Each responsibility is scored 0 to 100 on Capability (C), Reliability (R), Accountability (A) and Trust (T).
Share AI could take on = C/100 x (1 - max(A, T)/100)
Only the strongest reason a person is needed counts. A responsibility that needs a signature doesn't need a person twice as much because a client also likes a human touch.
AI can do it: share of 60% or more and R of 70 or more
Needs a person: share under 25%
AI helps: everything in between
Job share = sum of (weight x share) / sum of weights
Weights come from the Department of Labor's importance and frequency ratings, adjusted to how much your posting stresses each responsibility.
Savings = H x w - (tools + H x (1 - R/100) x w)
H is the yearly hours AI could take on, w is the hourly cost of the role (median wage x 1.3 for benefits and overhead, over 2,080 hours), and R is the reliability of that work. 1 unreliable hour of AI output costs 1 hour of a person's time to check.
Human Edge = 100 x sum(w x (1 - share) x min(1, e / 0.67)) / sum(w)
Coverage = 100 x sum(w x (1 - share) x min(1, e / 0.67)) / sum(w x (1 - share))
e is the evidence on your résumé for each responsibility: not shown 0, some exposure 1/3, a regular part of a role 2/3, led or deep 1. Responsibilities count by the part of them that still needs a person, and those marked AI can do it count as 0 in both, because the plan tells you to hand them to AI. Doing something as a regular part of a role counts as full evidence, and the fairest score comes from checking against the posting you're applying for.
Sources: O*NET 31.0 job data, BLS wages May 2025, Anthropic Economic Index (June 2026 release), and published studies cited per responsibility.
How the scores are checked against research
Our scores decide which responsibilities AI can take on and why. We then check the overall numbers against the best published research:
- Level: across the first 20 jobs we calibrated, AI could take on 57% of the time on average. That matches McKinsey Global Institute's 2025 estimate that current technology could automate 57% of US work hours.
- Order: where our score for a job disagrees with published research on how exposed that job is to AI (Eloundou and others, 2023), we move it halfway toward the research. Every new job is checked the same way.
- Speed: METR finds the length of work AI can finish on its own doubles about every 4 months. In real use, AI's success rate rose only about 1.4 points in 3 months, so we count about a tenth of lab gains.
- Evidence range: the level and order checks move the scores we drafted for each responsibility. So every job also shows a range. The evidence end uses only what has been measured for each responsibility (real-world success rates from Anthropic's Economic Index) or, where nothing was measured, the lower of our score and a second rating from a separate AI model. The other end is our figure. Where a responsibility rests on judgment alone, it shows its own range too, when the 2 ends differ.
- Possible is not the same as adopted: these numbers show what AI could take on today, not what most companies have done. The U.S. Census Bureau found 19.8% of businesses were using AI as of May 2026.