How the check works
GraveJob answers one narrow question: of the work your job is made of, how much can AI products that have actually shipped do today? It doesn't predict whether you'll be fired. Employers, regulation, and cost decide that.
1. Your title becomes an occupation
We search 51,506 job titles from O*NET 31.0, the U.S. Department of Labor's occupation database, and map yours to one of 922 standard occupations (an O*NET-SOC code). When several fit, an AI model picks the closest one, using your resume if you added it.
2. The occupation becomes tasks
O*NET lists the tasks each occupation does, marks each as core or supplemental, and rates how important it is (1 to 5). We judge a job on its core tasks, most important first, up to twelve. Supplemental tasks are duties only some people in the job do (a paralegal “calling witnesses to testify”), so they are added only when a job lists fewer than eight core tasks, and only if they rate 3.5 or higher.
O*NET is government data and updates slowly, so we also read current job ads for the job. Up to three day-to-day duties that the ads ask for and O*NET's list misses (new tools, work that moved into the job) are added as tasks and scored the same way; one most ads share counts as much as a core task. Each is marked “from current job ads” and links to the ads. The tasks AI already does or helps with become the stones on your island, most affected first.
3. Each task is checked against shipped AI
Live web search (Parallel) finds pages about AI products that do this kind of work. An AI model (DeepSeek V4.1 Flash) reads those pages and marks each task:
- Done by AI: a product anyone can buy today does the task end to end, and a person at most reviews the result.
- AI-assisted: products do a substantial part, like drafting, triage, or analysis, but a person still does the core judgement, the physical work, or the accountable sign-off.
- Human: no shipped product meaningfully does it yet.
Demos, research papers, waitlists, and announcements don't count. A product only counts as sourced when the page we link to actually names it; otherwise it shows without a link and the check's confidence drops. Audits are shared per occupation and refreshed after 21 days.
4. The score
Exposure is the importance-weighted share of tasks covered: a task done by AI counts fully, an AI-assisted task counts 0.45, a human task counts zero. Each task's weight is its O*NET importance minus one, so central tasks matter more than occasional ones. The result runs from 0 to 100.
Because an AI-assisted task counts 0.45, a job where every task is assisted scores 45. The labels are cut to match: Very High from 70 (real end-to-end automation), High from 45, Moderate from 25, Low from 10, and Very Low below that.
With a resume, we count only the tasks you actually do, plus up to three duties from your resume that the standard list misses. Your resume is read once to make a short summary of your work. Neither the file nor its text is stored, and the summary is deleted when the check finishes.
5. Time to impact
The years in which roughly half of the job's weighted task time is likely handled by AI at a typical employer. Higher scores start sooner. Rising layoffs or falling hiring in recent reporting (including news about your employer) pull the range a year earlier, and so does a job market where a third or more of the open ads already ask for AI skills; strong hiring pushes it a year later. The burial date on a result is a third of the way into its range: the end of a range is mostly the margin we add for uncertainty.
Every range carries a confidence level. With no live sources on shipped AI products, the check says “not enough data” and gives no range at all. Otherwise it earns points for tasks backed by a sourced product, for official BLS figures, and for recent hiring, pay, and layoff reporting. The percentage next to it is how much of that evidence was found, from 30% to 90%. It never reaches 100%, because no forecast is certain. Medium confidence adds a year to the end of the range; low confidence adds two.
| Score | Range starts | Range ends |
|---|---|---|
| 75 and up | this year | +1 year |
| 60–74 | this year | +2 |
| 45–59 | +1 | +3 |
| 30–44 | +1 | +4 |
| 15–29 | +2 | +6 |
| under 15 | +4 | +9 |
6. Hiring, pay, and layoffs
Official figures come from the Bureau of Labor Statistics (OEWS): how many people hold the job nationally and in your state, and the median pay. Web search adds reporting from the last 18 months: projections, posting trends, and layoff news for the occupation, your state, and your employer. A model summarises only what those sources say, and marks a signal "not enough data" when they don't say. A number such as “−6%” is shown only if that exact number appears in a cited source, so the model can't invent one. Every signal links to its sources.
7. Safer moves
O*NET lists occupations related to yours by shared skills. We score them the same way, using a quick check of their top five tasks unless someone has already run a full check, and show the ones that are meaningfully less exposed without being a big step down in required preparation. For each, we show how many people hold that job in your state (BLS) and open postings we found near you. Postings must be individual, open job ads from the last 60 days; salary guides and search pages don't count.
8. Checking a company
Type a company instead of a job and we identify the jobs that make up most of its workforce, map each to an O*NET occupation, audit them the same way, and weight the scores by rough share of staff. Company signals are its own hiring, layoffs, and announcements of AI replacing work. Redeploy paths are safer moves for its most exposed role.
9. Keeping score
Every check writes its forecast to a public track record, including the ones we declined for thin data. Job forecasts are checked against later BLS releases; company forecasts against news at the end of the range. The hit rate is shown there as forecasts come due.
10. The method underneath
GraveJob is built by the team behind Planetarium and follows the evidence rules of Mimi, Planetarium's simulation pipeline. Mimi turns real public evidence into calibrated predictions of how audiences respond, and it is held to the same rules used here:
- Every claim is grounded in evidence that can be shown, never in what a model “knows”.
- Results are ranges with an honest confidence, not single numbers.
- It abstains when evidence is too thin to call, instead of guessing.
- Predictions are recorded before the outcome and scored against what really happened.
Mimi itself simulates audiences, not labour markets, so its accuracy figures don't carry over to job forecasts. GraveJob earns its own record on the track record page.
What this can't tell you
- Whether your employer will adopt these tools, or when.
- New tasks AI creates in your job. The score only looks at existing ones.
- How a specific team will change. Company checks work at the level of the main occupations it employs.
AI models make mistakes, including about which products exist. That's why every claim links to a source. If something looks wrong, the source link is the place to check.
