apprun.icuAI tools

Work & law · task analysis

Job description AI impact, task by task

Paste a job description and get each duty labelled Could be automated, Augmented or Human-led, with its reasons. Unreadable lines are marked Not recognised. A rule-based conversation starter about tasks, not a forecast of anyone's job. Your text stays in this tab.

Private runs in your browserNo account nothing uploadedRule-based every reason is shown

Try it

Everything runs in this tab. The tool does not upload what you enter and stores nothing unless you export it.

Job description

0 of 60,000 characters

Up to 60,000 characters, pasted or loaded from a file. Everything is read in this tab; the tool does not upload it.

Add a task that is missing

Optional

Result

Rule-based estimates

Paste a job description and press Analyze. The tool sends nothing anywhere.

Method and sources: read this before you act on a verdict

Each scored line starts at 1, then gets points for its leading verb class, its object noun class and its context cues. A line with no recognised verb or object is not scored: it is marked Not recognised and left out of the shares. A reading mode shifts every score by one point before banding. Scores are clamped to bands 0 to 4: 0 and 1 are Human-led, 2 is Augmented (leans human), 3 is Augmented (leans AI), and 4 is Could be automated. An accountability phrase such as "signs off" or "licensed" caps the band at 2. Confidence counts how many of the verb, object and cue signals point the same way as the verdict; it is not a probability.

What the labels mean. Exposure is a property of tasks. The research behind this idea uses it as a proxy for potential impact and does not say whether the effect is augmenting or displacing. Automation, as Anthropic measures it, is a usage pattern. Displacement is about jobs and people, which a task list cannot show.

What is skipped. Requirements, benefits, equal-opportunity text, company descriptions and job metadata are skipped, and each skipped line is listed in full under "Why N lines were ignored". A benefit under a heading the tool does not recognise can still appear as a line, so check the table.

Without a duties heading, every sentence with an action verb is read as a task, except requirements, benefits, equal-opportunity and company text. The result panel says so. Pasted or loaded text is limited to 60,000 characters, and at most 300 tasks are analysed from one text.

  • Eloundou et al., "GPTs are GPTs", arXiv 2303.10130 (v5, Aug 2023): exposure is "a proxy for potential economic impact without distinguishing between labor-augmenting or labor-displacing effects". The paper says it does not predict "the development or adoption timeline" of LLMs.
  • Anthropic Economic Index, "Cadences" (26 June 2026): theoretical exposure is "an upper bound on what is possible instead of a measure of current use". The text gives no aggregate automation vs augmentation split.
  • ILO Working Paper 96 (Aug 2023): the most important impact is "likely to be of augmenting work", and scores below 0.25 count as "very low exposure". Its figures are potential, not realised.
  • Stanford Digital Economy Lab, "Canaries in the Coal Mine" (revised 12 Aug 2026): "We find no evidence of widespread, economy-wide job displacement". The 19% early-career gap is as of June 2026 (ADP payroll data), and the authors call these "early, descriptive indicators" rather than causal estimates.
  • NBER working paper 31161, "Generative AI at Work" (revised November 2023): a 14% average gain in issues resolved per hour, for 5,179 customer-support agents, with a 34% gain for novice and low-skilled workers.
  • Autor, Journal of Economic Perspectives (2015), abstract only: "Automation does indeed substitute for labor", and automation also "complements labor".

Dates matter. These sources were checked on 17 October 2026; figures from one report should not be read as current for another.

How to use Job Description AI Impact Analyzer

  1. Paste the job description. Include the duties. Headings such as Responsibilities help, but they are not required.
  2. Choose a reading mode. As written is the default. The conservative and generous modes shift every score by one point before banding.
  3. Press Analyze and check the reasons. Open the reasons column and the list of ignored lines. Untick any task that should not count.
  4. Export or copy the result. Download the Markdown report or the CSV, or copy a link that holds only the reading mode.

What this tool measures, and what it does not

The tool reads a job description as a list of tasks. It labels each task by how much of that task current tools could plausibly take over, and it shows the wording that produced the label. It measures the words of a task. It does not measure the work itself, how well anyone does it, or what an employer plans. The guide on which job tasks AI automates or augments explains the idea behind the labels, with sources.

Three terms are easy to mix up, and the research uses them differently. Exposure is a proxy for potential impact. In one widely cited early paper by Eloundou and colleagues, exposure is defined as "a proxy for potential economic impact without distinguishing between labor-augmenting or labor-displacing effects". Automation is a usage pattern. In Anthropic's June 2026 Economic Index report, a conversation counts as automated when its pattern is directive or a feedback loop. Displacement is about fewer jobs or people leaving them, and a task list cannot show it.

The labels describe tasks only. Could be automated means the wording describes routine work that a tool could plausibly do in large part. Augmented means a tool could help with part of the task while a person keeps the decision. Human-led means the wording points to judgement, people, physical work or accountability. Not recognised is a fourth status: the line has no verb or object the tool can read, so it is not scored, has no band, and is left out of the shares. None of the labels says that a job will go, or when.

How the text is read

The tool looks for headings such as Responsibilities, Key accountabilities, What you'll do, Requirements, Qualifications, Benefits, About us and Equal opportunity. Bullets and sentences under a duties heading become tasks. Requirements, benefits, equal-opportunity text, company descriptions and job metadata such as "Reports to:" are skipped. Every skipped line is listed under "Why N lines were ignored", grouped by reason, with its full text. The Markdown report lists them too. A line that holds several sentences or semicolon-separated parts is split, and each skipped part is listed on its own.

Without a duties heading, the tool reads every sentence with an action verb, from every part of the text except requirements, benefits, equal-opportunity text and company descriptions. The result panel says so at the top. A plain line counts as a heading only when the whole line is a known heading, and a short line ending in a colon is taken as a heading too. A heading such as "Perks at Acme" is not recognised, so the lines under it are judged one by one.

Benefits are skipped under a benefits heading and under requirements or equal-opportunity headings. They are also skipped wherever a known perk phrase appears in a line with no action verb in its first three words, such as "free lunch", "holiday allowance", "health insurance", "pension" or "401(k)". A duty that uses a verb the tool does not list and also mentions a perk word is skipped in this way too. The reason on screen says "benefit or perk wording", so it can be found and added back by hand.

Length. Pasted or loaded text is limited to 60,000 characters, counted as characters (an emoji counts as one). A loaded file goes through the same test as pasted text. The tool analyses at most 300 tasks from one text; any further duties are listed under the ignored lines with the reason.

How the scoring works

Every rule is listed here, so you can check any verdict against it. A line with a recognised verb or object starts at 1. A line with neither is not scored and has the status Not recognised. A scored line gets points for its leading verb, its object noun and the context cues in the same sentence. The reading mode then shifts the total by one point before banding, and the result is clamped to bands 0 to 4.

SignalExamples from the listsPoints
Verb: routine information processingcompile, enter, reconcile, update, report+2
Verb: first draft, analysis, monitoring, schedulingdraft, summarize, respond; analyze, forecast; monitor, verify; schedule, remind+1
Verb: advice within known policytroubleshoot, triage, recommend, implement0
Verb: ownership of a standardensure, oversee, govern-1
Verb: decision or sign-off, people work, hands-on, original directionapprove, sign, diagnose; coach, negotiate, hire; install, repair; design, strategize, redesign-2
Object: information, records, numbers, codeweekly report, invoices, forecast, API+1
Object: written textsummary, description, slides0
Object: customer or partner contact, plans or proposalscustomers, carriers; roadmap, proposal-1
Object: people, physical goods, regulated or clinical matters, trust and conflictteam, patients; equipment, samples; contracts, prescriptions; complaints, morale-2
Cues: routine (each distinct term, at most two count)weekly, daily, each week, template, high volume, data entry+1 each, max +2
Cues: human judgement (each distinct term, at most three count)stakeholder, negotiate, licensed, judgement, in person, on site, empathy-1 each, max -3
Cues: uncertainty (each distinct term, at most two count)novel, ambiguous, unclear, strategic, creative direction-1 each, max -2
Accountability capsigns off, signed off, signing off, sign-off, legally responsible, legal responsibility, final decision, final say, final authority, final sign-off, licensed, patient care, safety-critical, accountable forband capped at 2
Reading modeAs written, conservative, generous0, -1 or +1 before banding

Bands: 0 and 1 are Human-led; 2 is Augmented (leans human); 3 is Augmented (leans AI); 4 or more is Could be automated. The lists sit in the page's script file. Each entry is a deliberate choice that you can read there, not the output of a trained model.

Advice without an object. A line whose only recognised word is an advice verb, such as "Recommend process changes that reduce late deliveries.", starts at 1 and scores 0 for the verb. It therefore lands in Human-led at band 1, and the reasons say so.

Worked example. "Compile the weekly sales report from the CRM" has the verb compile (routine, +2), the object report (information and records, +1) and the cue weekly (routine, +1). The start of 1 plus those 4 points gives 5, which is clamped to band 4, Could be automated. The reasons list shows each step.

Accountability cap. If the sentence contains one of the phrases in the table, the band cannot go above 2. The wording names a person who must answer for the result, so the tool does not let the score reach the automated side.

Confidence. The chip counts three separate signals: the verb, the object and the cues. High means all three point the same way as the verdict. Medium means two agree and none disagree. Everything else is low. It counts agreeing signals; it is not a probability. A Not recognised line has no chip.

Reading modes. As written uses the scores as they are. Conservative subtracts one point from every score before banding, so a task can only move down a band or stay put. Generous adds one point before banding, so a task can move up. The reasons show the change for each task.

What each task shows. Under every row you can open three notes: what current tools could plausibly do with the task, what stays with a person, and how the task could change. Each note is conditional: it says what could happen, not what will. They are generic patterns for the kind of verb and object found, not product recommendations.

Limits and what it cannot tell you

  • It reads words, not work. A tidy sentence can hide a task that is mostly physical, political or a matter of trust.
  • The lists are finite. They include general terms from operations, marketing, finance, healthcare administration, software, teaching, trades, customer support, legal administration and sales. A line with no verb or object from the lists is Not recognised: it is listed with its reason, but it is not scored and it does not count in the shares.
  • Some words have two senses. Words such as lead, report and forecast are read by position, so a line can be read wrongly. The reasons column shows which word was taken as the verb.
  • Only the first six words are checked for a verb. A verb later in a long sentence is missed, and the line may then be Not recognised or scored on its object alone.
  • Some benefits can get through. A benefit under a heading the tool does not recognise, worded without a perk phrase from the list, can appear as a line. If it contains a verb, it is scored like a duty. Check the table, untick it, or remove it by writing the real duties by hand.
  • Some duties can be skipped. A duty that mentions a perk word such as pension, and uses a verb the tool does not list, is skipped as perk wording. Its reason is listed under the ignored lines, so it can be found and added back.
  • The reading is capped. At most 300 tasks are analysed from one text, and pasted or loaded text is limited to 60,000 characters. Anything beyond the cap is listed under the ignored lines with the reason.
  • Exposure is an upper bound. The ILO's 2023 working paper describes its figures as "an upper-bound estimate of exposure", and Anthropic calls theoretical exposure "an upper bound on what is possible instead of a measure of current use".
  • The middle is unresolved. The ILO calls one group of occupations "the big unknown" and says its framework and data "do not allow for a clear-cut classification of this group". Middle bands here carry the same uncertainty.
  • Productivity evidence is narrow. An NBER working paper (31161, revised November 2023) studied 5,179 customer-support agents. It found that AI assistance raised issues resolved per hour by 14% on average, with a 34% gain for novice and low-skilled workers. That is one setting, not a rate for every job.
  • Timing is out of scope. Eloundou and colleagues write that they "do not make predictions about the development or adoption timeline" of such models.
  • Early-career findings are descriptive. Stanford's Digital Economy Lab reports that employment of workers aged 22 to 25 in AI-exposed occupations was 19% below a kept-pace counterfactual as of June 2026, using ADP payroll data. It calls these "early, descriptive indicators" rather than causal estimates.
  • Access is not modelled. The ILO says its effects "do not consider infrastructure constraints", and the tool does not know what tools a team can use.
  • Your own context wins. A verdict on the wording cannot see your team, your tools, your industry rules or the way a task is done in practice. Treat each label as a prompt for a conversation.

Two related tools sit alongside this one. To check whether a report's wording goes beyond its evidence, use the Claim–Evidence Map. To draft the steps of a task as an agent specification, use Idea to Agent Spec. Before you paste a job description into any other AI tool, check it with the LLM Context Passport for personal data.

Privacy: what happens to your input

The text stays in this tab. The analysis is JavaScript that runs in your browser, and the page makes no request to a server with your text. The tool does not store anything in the browser. Clear empties the page. The share link holds only the reading mode, never the text. Downloads, the Markdown report and the CSV, are created in your browser from the current result. Delete them when you no longer need them.

Loading a file reads it in this tab in the same way. Neither the file nor the text is part of any request the page makes.

Sources and dates

The rules above are this tool's own. The sources explain the ideas and bound the claims. Figures are dated, and most describe potential rather than realised effects.

  • Eloundou, Manning, Mishkin and Rock, "GPTs are GPTs" (arXiv 2303.10130, version 5, August 2023). It defines exposure, and its rubric includes a time test: whether a model could cut the time for a task by at least half.
  • Anthropic, "Cadences" (Economic Index report, 26 June 2026). It defines observed and theoretical exposure. Its text gives no aggregate automation vs augmentation split. An older September 2025 edition reported that 77% of business API uses showed automation patterns, against about 50% for Claude.ai users; that figure is dated.
  • ILO Working Paper 96 (August 2023). It splits exposure into automation and augmentation potential and publishes the scoring bands used above. Its later ILO Working Paper 166 (March 2026) reports average exposure of about 32% of employment in high-income economies and about 15% in low-income economies.
  • Stanford Digital Economy Lab, "Canaries in the Coal Mine" (revised 12 August 2026). The 19% gap is as of June 2026 (15% at the July 2025 data vintage), using ADP payroll data through June 2026.
  • Brynjolfsson, Li and Raymond, "Generative AI at Work" (NBER working paper 31161, revised November 2023). It studied 5,179 customer-support agents, and its 14% average gain comes from that setting only.
  • Autor, "Why Are There Still So Many Jobs?" (Journal of Economic Perspectives, 2015). The tool relies on the abstract only. It argues that automation substitutes for labor and also complements it, so substituting for some tasks does not by itself mean fewer jobs.
  • O*NET (onetonline.org) describes occupations through more than 19,000 task statements. The tool does not use O*NET data directly; the idea of a task-level unit comes from it.

Questions

Is my data uploaded?

No. The text is read and scored by JavaScript in this tab, including text from a file you load. The tool's code sends nothing to a server and saves nothing in your browser; a file exists only if you download it. The site-wide promo banner is separate and never receives your text. The share link holds only the reading mode, never your text.

How accurate is this?

It is not a measurement of accuracy in any statistical sense. It applies a published rule set: a verb class, an object class, context cues and an accountability cap. It can misread unusual wording, and a task it labels Could be automated may still need a person to check it. Read the reasons, then judge the task yourself.

Does a high AI score mean my whole job goes?

No. A verdict describes one task. The exposure research this tool draws on says exposure does not distinguish augmentation from displacement. Stanford's Digital Economy Lab reports "no evidence of widespread, economy-wide job displacement" and calls its early-career findings descriptive, not causal. Autor's task-based argument is that substituting for some tasks does not automatically mean fewer jobs.

How do exposure and automation differ?

Exposure asks whether AI could speed up or take over part of a task. Automation, in Anthropic's usage measure, is a usage pattern: a conversation counts as automated when it is directive or a feedback loop. The two are different measures and this tool does not measure either one directly.

Why do some tasks land in the middle?

Tasks that mix a routine step with a human judgement often score in the middle bands. The ILO uses the label "the big unknown" for a group of occupations whose exposure its framework cannot classify clearly. A middle label here means the wording points both ways, not that the task is half done by AI.

What does the confidence chip mean?

It counts how many of three separate signals, the verb, the object and the context cues, point the same way as the verdict. High means all three agree. It is not a probability, and a low chip usually means the wording is thin or unusual. A line marked Not recognised has no chip.

Why does the tool ignore some lines?

Benefits, equal opportunity statements, company descriptions, requirements and job metadata are not duties, so they are skipped. Every skipped line appears in full under "Why N lines were ignored", grouped by reason, and in the Markdown report. Perk wording is skipped only when the line has no verb in its first three words, and a benefit under a heading the tool does not recognise can still appear as a task, so check the table.

Can I trust a verdict without reading the reasons?

No. The reasons are the point. Every verdict lists the verb, the object, each cue that counted and any accountability phrase, with the arithmetic that produced the band. If a reason is wrong for your job, the verdict is wrong too, and you should untick the task or rewrite it.