Forecasts Missed Both Ways. Track Your Own Task List.

Why a Job Survives Long After Most of Its Tasks Don't

A job is a bundle of tasks, and the tasks that hold out longest are usually the ones with no clean substitute: reading what a client actually wants under what they said, deciding what to leave out, being the name someone can hold accountable when a call goes wrong. Those tasks rarely make up most of the job's hours. They still keep the whole role alive, because ownership doesn't divide into fractions the way task-hours do.

Indeed's GenAI Skill Transformation Index rated almost 2,900 work skills against current models and mapped them onto more than 53.5 million US postings: 46% of the skills in a typical posting now fall into a hybrid or fully transformed category ¹. At the same time, only 19 of those skills, 0.7% of the whole set, were rated very likely to be fully replaced ². Close to half of a typical posting's skills are rated as changing shape. Almost none are rated as vanishing whole, and the gap between those two numbers is the actual reason a job-level date keeps missing.

The Burning Glass Institute found the same split running inside single occupations rather than across them: skills exposed to automation are 16% more likely than baseline skills to see posting demand fall, while skills exposed to augmentation are 7% more likely to see it rise ³, and the occupations doing the most of one usually do the most of the other too. A role rarely sorts cleanly into safe or exposed. It carries both at once, which is exactly the part a single forecast date cannot represent.

Run This Review Every Quarter

Once each quarter, sort your own task list into three columns: what you still do yourself start to finish, what you now check rather than produce, and what the model still gets wrong in your specific field in a way a stranger wouldn't catch. Track which column each task sits in over time. The count in any single column on any single day tells you less than the direction it's moving.

• What moved from "I do it" to "I check it" this quarter
• What the model still got wrong in your domain, and why
• What that shift changes about where your next quarter of learning goes

Take a role like financial analysis or paralegal review. A first draft of a summary memo might already sit almost entirely in the "I check it" column, while spotting the one number in a client's file that doesn't reconcile with anything else still sits in "I do it." Neither column is fixed. What moves between them, and how fast, is the only forecast you actually need, because it is built from what happened in your own work rather than from what a model could theoretically do somewhere else.

Treat any published skill-change forecast as exactly that, a forecast, including the most repeated one: LinkedIn's Work Change Report projects that 70% of the skills used in most jobs will change by 2030 ⁴. It's a projection the company has not published the model behind, not a measurement anyone outside LinkedIn can check against results. Your own quarterly log is the version of that number you can actually verify, one task at a time.

What to Watch in Your Own Field

Watch what your field's job postings start asking for before you watch what commentators predict. Indeed Hiring Lab counted AI-related language in 6.3% of US postings in August 2026, up from a prior peak of 3.3% in 2022 ⁵, still a minority share and the fastest-moving signal you can check yourself without waiting for anyone's survey.

What employers actually ask for when a posting names an AI skill is worth reading before you assume the language showing up in your own field's postings means what you think it means; the wording changes faster than the underlying task does. Reading a posting for this signal is different from skimming it: a tool name buried in a "nice to have" line means little, while the same tool named as part of a specific deliverable, in the required section, is the posting telling you what the role's task list has already become.

One pattern worth tracking rather than trusting outright: Stanford's Digital Economy Lab found entry-level employment growing more slowly in occupations built on codified knowledge, the kind that lives in a textbook or a written procedure, while occupations built on tacit knowledge, learned through practice and mentorship, grew faster for mid-career and senior workers ⁶. The researchers call this a raw, descriptive gradient, not a proven mechanism, and it names something specific worth adding to your own review: which of your tasks depend on steps a model could read, and which depend on judgment nobody wrote down.

Employers doing a version of this same exercise are told to pull the occupation's published task statements and mark what a model can already draft before writing a job post. That is the same list, read from the hiring side, which is worth knowing the next time a posting in your field changes shape.

Should You Still Plan Around a Date?

No. Plan around the columns instead. A date is a single number that has to be right about a hundred different tasks at once, and the evidence above shows it usually isn't. A quarterly review updates with what you're actually seeing, which is the only input that has been reliable in this territory so far.

If the review keeps turning up the same column growing quarter after quarter, that's information a date was never going to give you: not when your job disappears, but where your next raise or your next application should be aimed.

Two consecutive quarters where the same task moves from doing to checking is worth an actual response: ask for time on the tasks in the other two columns, the ones a model still can't carry, and let that be where you spend the next stretch of learning. That response is available to you whether or not anyone ever names a year for your occupation.

See the benchmarks >