The short version
We look at every job the same way. Four steps.
1
We start with the official record.
Every job in Welica is matched to a real occupation in the US and UK government job databases: how many people do this job, how fast it is growing over the next ten years, what it pays, and what you need to get in.
2
We break the job into its tasks.
A job is five or six things you actually do in a week. For each we ask: how much of this does AI do now? Four-point scale, from mostly done by AI to barely touched, plus one line on what has changed and what is still yours.
3
We check what is happening to new hires.
The question for a student is not "will this job exist" but "can I get in". Government data on junior postings, job by job, year on year, separates a shrinking job from one that is fine but harder to start in.
4
We combine the three signals into one plain-English call.
AI impact, four-year outlook and entry competition go through a fixed table, so two jobs with the same ratings always get the same verdict, written in words a sixteen-year-old can read.
EXAMPLE OF ONE CELL IN THE TABLE
AI impact: automates parts+Outlook: steady or growing+Entry competition: high=Squeezed at entry, growing top
The field is growing, but the junior work is the part AI now does, so the first job is harder to land. Different from cooling, where the whole occupation contracts, and from reshaped, where the tasks changed but junior work still exists.
What we do not do
We do not use press releases, paid market reports, or recruitment company blogs. Every fact in every job record links to a source we opened and read. Every job carries a confidence rating and says whether the evidence is about this exact job or the wider field. Every job is checked again every month.
935
sources opened and read
Monthly
every role re-checked
The four-point scale, across every task we track
Mostly done by AI · 204 tasks
Sped up a lot · 806 tasks
Lightly changed · 752 tasks
Barely touched · 662 tasks
Each role is broken into five or six tasks; each task is rated on this four-point scale. Reviewed September 2026.
For the curious: the full method
How the data is actually put together
What we review
Six kinds of evidence, in order of how much weight they carry.
Official occupation data
Every role is matched to a code in the US O*NET database and the BLS Occupational Outlook Handbook, and where it exists the UK National Careers Service profile: task list, titles, the 2025–35 projection, employment, openings, median pay, entry education. We read the sub-line for the specific occupation, never the group headline.
The BLS AI-exposure table
In August 2026 the BLS published an AI-exposure quartile for 831 occupations. Top quartile is high, the middle two medium, the bottom low. No judgement involved.
Entry-level hiring data
The UK government’s entry-level hiring snapshot and Indeed Hiring Lab. The only data that separates "the job is shrinking" from "the first rung is going".
Regulators and professional bodies
For what changed in the job’s rules. Never used for a demand or salary claim, because several of them sell certification.
Named primary studies
ILO, OECD, peer-reviewed papers. Used for how work is changing, with the date read from the page itself.
Trade press, checked
Staffed publications with named bylines. Used for what is happening, never for a number. Every page checked for sponsorship labels.
What is banned
Newswires and their mirrors, paid market-research firms, vendor blogs, recruiter content marketing, and any title invented by a vendor survey. If a claim only exists in a press release, the claim goes.
How we put it together
1
Anchor the role.
Match to an occupation code. Compare our task list to the official one; if they do not match, re-anchor. Confirm employers post the title. If not, the record says so: "you get in as an accountant and grow into this".
2
Break the job into tasks.
Five or six verb-led tasks that cover the real week. Each gets one of four AI ratings, plus what changes and what is left. The rating must be supported by what is left.
3
Set the three signals, each on its own evidence.
AI impact from the task pattern; four-year outlook from the official projection sub-line; entry competition from junior posting trends and how exposed the junior tasks are.
4
Combine through a fixed table.
The same three inputs always give the same call, so verdicts are consistent across 406 roles.
5
Grade the evidence.
Every role carries an evidence tier, role-specific, role-family or method-only, and a confidence level. A low-confidence honest record beats a confident invented one.
6
Verify everything.
Every citation fetched and read. Dates from the page body. One bad source triggers a sweep of the whole field.
7
Refresh monthly.
Every role re-checked against the newest official data. News items carry expiry dates. A rating from last cycle is never presented as current.
What is judgement and what is not
Two of the signals are mechanical: the exposure band comes straight from the BLS table, and the outlook straight from the projection. The AI-impact label and the entry-competition rating are structured judgement: an analyst reads the evidence and applies the rules above, and that judgement is visible in the record. Every task rating carries its own justification, every claim its source. We think that is more defensible than a formula that hides the same choices inside a weighting.
Sources we use
Sources we don’t use: newswires and their mirrors, paid market-research reports, vendor blogs and surveys, recruiter content marketing, and any statistic whose only home is a press release.
Limits
The occupation records are US and UK, because those governments publish task-level data with projections. For India, the UAE and other markets, the read on demand and entry-level hiring comes from current postings and named studies, and each record says which. Where the evidence is about the wider field rather than this exact job, the record says so and the confidence is lower. Found a source we should have read, or one we should not have? Tell us.