The 15-minute method, step by step
Each step has one output, so you can stop at any point with something useful. Occupational databases such as O*NET list the tasks for each occupation, so a posting is a short, employer-written version of that list. The split below is a suggested timing for a posting of ordinary length. It is not a measured figure from any study.
| Minutes | Step | What you write down |
|---|---|---|
| 0–2 | Strip boilerplate: benefits, equal-opportunity text, perks and company marketing | Only the lines that describe work |
| 2–5 | List the tasks: one duty per line, starting with the responsibilities section | A plain list of duties |
| 5–7 | Tag the verb and the object: the main action, and what it acts on | Two words per duty, such as “compile · report” |
| 7–9 | Check accountability and presence: sign-off, licence, legal responsibility, on-site or hands-on work, stakeholder meetings | Yes or no for each duty |
| 9–11 | Mark routine or judgement: does the duty repeat on a fixed pattern, or does it need judgement about something new? | Routine, judgement or mixed |
| 11–13 | Decide the label: human-led, augmented or could be automated | A label and one reason |
| 13–15 | Write the questions: one for each duty you are unsure about | Three to five questions |
Keep each reason short and checkable. “Compile plus weekly report, so routine” is a reason you can test against the posting. “Seems routine” is not. If you would rather start from a first pass, paste the posting into the Job Description AI Impact Analyzer and then check its reasons against your own tags.
A worked example with an invented operations analyst
The duties below are invented for this example and do not come from a real employer. We tagged them by hand with the method above, to show how each label is reached. The hand labels come from reading each row by the method above; the analyzer column is its real output, checked 18 Oct 2026. For the background on why tasks, not jobs, are the unit, read the guide to which job tasks AI automates, augments or leaves alone.
| Duty (invented) and verb · object | Hand label | Analyzer output, checked 18 Oct 2026 |
|---|---|---|
| Compile the weekly stock-variance report from the system export compile · weekly report | Could be automated | Could be automated, band 4, high compile +2, weekly stock variance report +1, weekly +1 |
| Reconcile supplier invoices against purchase orders each week reconcile · invoices | Augmented (leans AI) | Could be automated, band 4, high reconcile +2, supplier invoices +1, each week +1 |
| Sign off budget variances above the agreed limit sign off · budget variances | Augmented (leans human) | Human-led, band 0, low sign -2 (judgement), budget variances +1 |
| Explain variances to department heads in person at monthly reviews explain · variances | Human-led | Augmented (leans human), band 2, low explain 0, variances +1, monthly +1, in person -1 |
| Redesign the reporting approach when a data source changes redesign · reporting approach | Human-led | Human-led, band 0, low redesign -2 (creative), reporting +1 |
| We are an equal opportunity employer none | Ignored | Ignored (company, benefits, EEO or job-metadata text) |
“Could be automated” is a label on one task under a fixed rule set, not a finding about the analyst’s job. The hand labels and the analyzer agree on rows 1 and 5 and disagree on rows 2, 3 and 4. Row 2 is the clearest case: the analyzer counts “each week” as a routine cue (+1) and gives the highest band, where the hand reading said augmented. Row 3 is the reverse: “sign off” scores as a judgement verb (-2), which pulls the duty to Human-led, where the hand reading said augmented. Row 4 counts “monthly” as a routine cue, which the hand reading did not, and the in-person cue (-1) does not outweigh it, so the analyzer gives augmented, leaning human, where the hand reading said Human-led. Five rows are not a test of the tool. They show where a fixed rule set and a careful reader part ways, and why the reasons matter more than the label.
Where the analyzer and a hand reading differ
The Job Description AI Impact Analyzer applies a fixed, published rule set, not a judgement about the job. It looks for duties in responsibility-like sections. Benefits are skipped when they sit under a recognised benefits or requirements heading, or contain a known phrase such as health insurance. A benefit under an unrecognised heading can be counted as a task, so check the task table. If a posting has no duties heading, it reads sentences with an action verb instead, and the page says so. For each duty it reads the main verb, the type of object, and the context and autonomy cues, then sets a band. An accountability phrase such as “signs off”, “final decision” or “licensed” caps a duty at augmented. It never classifies a duty from the job title alone.
It then shows the reasons behind each label and a confidence chip that reflects how many separate cues agree. The table above sets hand labels against the analyzer’s output, and the two can differ. A word list can miss a phrase it does not contain, and a verb it does not list can be read as the main action. Compare the reasons with your own tags, not only the labels, and check the posting wording where they split. A vague posting gives fewer cues, so its labels deserve less weight. A line with no recognised verb or object is marked Not recognised, gets no band, and is left out of the shares.
What to do with the result
Use the labels to decide where to look, not to judge the role. Three uses are reasonable.
- Double down on human-led duties. These are the duties where the posting asks for judgement, sign-off or presence. Name them in your own review, and keep examples of the decisions you made. Those examples are evidence you can show a manager.
- Learn the tooling for augmented duties. For a duty that leans toward AI, try the step where a tool helps and check its output yourself. The ILO describes augmentation as automating some tasks within an occupation while leaving time for other duties.
- Ask the team concrete questions. Which of these duties does the team already do with AI, and who checks the output? Which duties need sign-off, and by whom? Which routine duties would the team want to change first, and why?
If a headline claims that a job is about to disappear, check the claim before you repeat it. The Claim–Evidence Map lists a paper’s claims beside the sentences that support them. It maps wording and structure, and it cannot tell you whether a claim is true.
Honest limits of a 15-minute audit
One posting is one employer’s wording of a role. It can be out of date, leave out duties the team actually does, and say nothing about how a particular person spends a week. A 15-minute pass is a sorting exercise. Its routine and judgement tags are a reading of text, not a measurement.
It is also not the time test used in the research. Eloundou et al. ask whether a model could cut the time for a task by at least half, a question about time that a posting cannot answer. Anthropic’s June 2026 report calls theoretical exposure “an upper bound on what is possible instead of a measure of current use.” A list of duties an AI could do is therefore not a list of duties people already hand over to it.
Timing is outside the method. Eloundou et al. write, “We do not make predictions about the development or adoption timeline of such LLMs”, so no audit of one posting can supply a timeline. The analyzer’s own framing says the same thing: exposure is about tasks, not jobs, and none of it predicts layoffs or timing.
Common mistakes. Avoid these:
- Reading a label as a verdict on a person, rather than a note about one task.
- Counting benefits, perks or equal-opportunity text as duties.
- Tagging from the job title instead of from the duty wording.
- Applying the NBER customer-support result, a 14% average gain for 5,179 agents, to other roles.
- Treating any label as a timeline.
Common questions
How long should it take to audit a job description for AI exposure?
About 15 minutes for one posting of ordinary length. The split in the method table is a planning guide, not a rule. A long posting, or one with unclear wording, will take longer. It is fine to stop after the labels and write the questions later.
Should I audit the job posting or my own list of duties?
Use both if you can. The posting shows what the employer wrote, which may leave out work the team actually does. Your own list shows what you do each week. Tag both the same way and compare them. The gaps between the two are often the most useful questions to bring to a manager.
Does a human-led label mean AI cannot help with that task?
No. The label means the wording points to a person leading the duty, usually through judgement, sign-off or presence. Parts of such a task can still be supported by AI, for example drafting, summarizing or checking. Read the reasons behind the label, not only the label itself.
Is this the same as an AI job-loss score?
No. It labels the duties in one posting with a fixed rule set. It makes no prediction about jobs, layoffs or timing. The research uses exposure to mean potential, and exposure is not the same as displacement. A high-exposure duty is a reason to ask a question, not a conclusion.
Can an employer use this to decide who to hire or keep?
It is not built for that. It classifies duties in a posting, not people, and the sources do not validate this kind of rule against later employment outcomes. Decisions about people involve employment and equality law, which this method does not cover. This is general information, not legal advice.
What if the job posting is vague?
Then the labels will be less certain, and the result is still useful. Vague wording gives fewer cues, so the reasons are thinner. Turn each vague duty into a question for the hiring team, such as what a typical week looks like and which decisions the role makes without a second check.
Sources and further reading
- Eloundou, Manning, Mishkin and Rock, “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” — arXiv 2303.10130, v5, August 2023
- Anthropic Economic Index report — Anthropic, June 26, 2026
- Gmyrek, Berg and Bescond, “Generative AI and jobs” — ILO Working Paper 96, August 2023
- Gmyrek, Viollaz and Winkler, “Disruption without dividend” — ILO Working Paper 166, March 2026
- Brynjolfsson, Li and Raymond, “Generative AI at Work” — NBER working paper 31161, revised November 2023
- Tomlinson et al., “Working with AI: Measuring the Applicability of Generative AI to Occupations” — arXiv 2507.07935, v6, December 2025
- O*NET OnLine, job duties search — National Center for O*NET Development (undated)