Task Shift
Atlas
↖ Sahilverse home

Sahilverse · Work & technology · Published 26 Oct 2026

Will AI take my job? Look at the tasks, not the title.

The short, honest answer: nobody can tell you with certainty, and anyone quoting a precise percentage for your job is guessing. What the evidence does support is that AI changes tasks inside jobs far more often than it erases whole occupations — and that what a technology can do is a different question from what an organisation actually adopts.

So break your work into tasks. Some are codified and low-stakes (likely to be automated). Some get faster with a tool but still need you (augmented). Some depend on presence, trust and care (worth protecting). And some are codifiable but high-stakes — they force the organisation to rethink who checks, who learns, and who is accountable.

  1. Exposure ≠ loss. The ILO's 2025 index finds about one in four workers globally in an occupation with some generative-AI exposure, and stresses this is potential exposure, not job loss.
  2. Expectations overshot. In McKinsey's surveys, 32% of 2025 respondents expected AI-driven headcount cuts; a year later 14% reported them.
  3. The ladder is wobbling. Stanford researchers find no economy-wide displacement, but fewer 22–25-year-olds in highly exposed US occupations — mainly through slower hiring.

Sources and dates for each point are in Evidence and on the Read Further page.

02 / The sorting table

Pick a role. Sort its tasks. Argue with the model.

Each card is one task. Put it in a bin with the buttons (or keys 1–4 when the card is focused, or drag on desktop). The model offers its own suggestion — you can disagree, and the table records where you did. Everything stays in your browser.

Text atlas — every role, task and model suggestion (accessible equivalent of the sorting table)

Frontline retail / kirana counter

A neighbourhood shop or small store: owner, family members or one or two assistants. In South Asia much of this work is informal, cash-and-UPI mixed, and built on credit relationships.

Tasks for Frontline retail / kirana counter
TaskC / R / SModelUncertaintyNote
Answer stock and price questions on WhatsApp3 / 1 / 0automatelowCatalogue bots can draft replies, but regulars often message to be recognised, not just informed.
Reorder fast-moving goods from distributors2 / 1 / 2rethinkcontestedSuggestions from sales history help; the call on credit terms and a distributor's reliability stays with the owner.
Decide who gets goods on udhaar (credit)1 / 3 / 3protectlowA judgment about trust, family circumstance and neighbourhood ties. Scoring it automatically can quietly exclude people.
Record daily cash, UPI and khata entries3 / 0 / 1automatelowHighly codifiable — if the shop has a phone, connectivity and a tool in the owner's language.
Handle a complaint about a short-weight packet1 / 3 / 2protectmediumDe-escalation and goodwill in a face-to-face relationship.
Arrange shelves for festival season1 / 0 / 1augmentmediumPhysical, local and seasonal. Data may hint; the floor decides.

Skills that may travel: Reading customers' unspoken needs → Customer success, community health outreach; Informal credit risk judgment → Microfinance field officer, collections with dignity; Supplier negotiation → Procurement, vendor management. Access: Device sharing, patchy data, tools mostly in English, and thin margins that make subscription software a real cost.

HR generalist

A people-operations role in a mid-sized organisation: hiring, policies, employee relations and records.

Tasks for HR generalist
TaskC / R / SModelUncertaintyNote
Draft job descriptions and offer letters3 / 0 / 1automatelowTemplates and drafting tools do much of this; legal accuracy still needs review.
Screen CVs against a shortlist2 / 1 / 3rethinkmediumAutomated screening can encode bias and exclude non-standard paths. High stakes for the candidate.
Answer routine policy questions3 / 1 / 1automatemediumA policy assistant can answer; HR still owns whether the answer is right.
Investigate a workplace grievance0 / 3 / 3protectlowTrust, fairness and confidentiality. Not a task to delegate to a model.
Run onboarding for a new cohort1 / 2 / 1protectcontestedLogistics can be automated; belonging cannot.
Prepare headcount and attrition reports3 / 0 / 2rethinklowFaster with AI — but the numbers drive decisions about people, so the method needs owning.

Skills that may travel: Structured interviewing → Research interviewing, UX research; Policy interpretation → Compliance, public-sector administration; Conflict mediation → Mediation, community dispute resolution. Access: Data-privacy rules and employee trust limit what can be fed to external tools.

School teacher

A classroom teacher in a government or private school, often with large classes and mixed learning levels.

Tasks for School teacher
TaskC / R / SModelUncertaintyNote
Prepare worksheets and question banks3 / 0 / 1automatelowGeneration is fast; checking level, language and accuracy is the real work.
Grade objective tests3 / 0 / 2rethinklowCodifiable — but marks feed into decisions about children.
Give feedback on a student essay2 / 2 / 2protectcontestedAI comments can scaffold; the teacher knows what this child needs to hear.
Notice a child who is struggling at home0 / 3 / 3protectlowAttention, safeguarding and relationship. Protect.
Run a classroom discussion0 / 3 / 1protectlowLive, social, improvised.
Fill attendance and admin registers3 / 0 / 0automatelowPure overhead; a good candidate to remove entirely — if the system allows.

Skills that may travel: Explaining complex ideas simply → Instructional design, technical writing; Managing groups → Facilitation, training; Reading learning levels → Assessment design, edtech product. Access: Shared devices, school policy on AI tools, and uneven bandwidth across schools.

Graphic designer

Freelance or in-house, producing brand, social and print material.

Tasks for Graphic designer
TaskC / R / SModelUncertaintyNote
Produce size variants of one approved design3 / 0 / 0automatelowResizing and reformatting are highly automatable.
Generate early moodboards2 / 0 / 0augmentlowImage tools speed exploration; taste chooses.
Interpret a vague brief with a client0 / 3 / 2protectlowFinding out what the client actually means.
Decide the visual direction for a brand1 / 2 / 2protectcontestedJudgment, context and accountability for the result.
Check rights and licensing of assets2 / 0 / 3rethinklowGenerated imagery raises open IP questions; this step grows, not shrinks.
Retouch product photographs3 / 0 / 1automatelowIncreasingly tool-driven.

Skills that may travel: Visual judgment → Art direction, product design; Client translation → Account management, UX research; Production discipline → Design operations, print production. Access: Tool subscriptions and hardware are costly for freelancers; client contracts may restrict AI use.

Marketing executive

Running campaigns, content and channels for a brand or small business.

Tasks for Marketing executive
TaskC / R / SModelUncertaintyNote
Write first drafts of social posts and ads3 / 0 / 1automatelowDrafting is fast now; voice and accuracy are the bottleneck.
Segment the customer list3 / 0 / 1automatelowAnalytics tools handle this well when data is clean.
Decide positioning against a competitor1 / 1 / 3augmentmediumStrategy with real downside if wrong.
Interview customers about why they churned0 / 3 / 2protectlowListening; people say different things to a person.
Approve claims in an ad for legal accuracy2 / 0 / 3rethinklowGenerated copy can overclaim; someone has to own it.
Report weekly campaign numbers3 / 0 / 0automatelowDashboards and summaries.

Skills that may travel: Audience insight → Research, product management; Copy judgment → Editorial, communications; Experiment design → Growth analytics, policy pilots. Access: Platform costs and the quality of a small business's customer data.

Data / business analyst

Turning data into decisions — often early-career, often doing the codified work that AI handles well.

Tasks for Data / business analyst
TaskC / R / SModelUncertaintyNote
Write routine SQL and spreadsheet formulas3 / 0 / 1automatelowCoding assistants do much of this — the entry-level learning ladder is affected here.
Clean and reconcile messy data2 / 0 / 2rethinkmediumFaster with tools, but silent errors propagate.
Frame the business question with a stakeholder0 / 2 / 2protectmediumWhat question is worth answering?
Build the monthly reporting deck3 / 0 / 1automatelowTemplated output.
Decide whether a result is causal or coincidental1 / 0 / 3augmentlowMethodological judgment; models produce confident wrong answers here.
Explain a counter-intuitive finding to leadership1 / 3 / 2protectmediumPersuasion and trust.

Skills that may travel: Problem framing → Strategy, policy analysis; Statistical scepticism → Audit, research, journalism; Data storytelling → Communications, teaching. Access: Enterprise data rules may bar external tools; juniors may lose the routine tasks they used to learn on.

Home care worker

Caring for older adults or people with disabilities at home — often low-paid, often women, often informal.

Tasks for Home care worker
TaskC / R / SModelUncertaintyNote
Help a client bathe, dress and move safely0 / 3 / 3protectlowPhysical, intimate, relational.
Write the daily care log3 / 1 / 1automatemediumVoice-to-text logging could save time — if it works in the worker's language.
Remind about medication schedules2 / 1 / 3rethinkmediumReminders can be automated; noticing a missed dose and why cannot.
Notice changes in mood or health0 / 3 / 3protectlowAttention over time.
Coordinate with family and doctors1 / 3 / 2protectmediumMessaging tools help; the relationship carries it.
Plan weekly shopping and meals2 / 1 / 0augmentmediumLists can be generated; preferences are lived.

Skills that may travel: Patient observation → Nursing assistant, community health work; Calm under pressure → Emergency dispatch, front desk triage; Family communication → Social work, patient navigation. Access: Phones may be shared or basic; monitoring tech can become surveillance of the worker rather than support.

Customer support agent

Answering customers by chat, phone or email — a large employer in South Asian BPO and IT-enabled services.

Tasks for Customer support agent
TaskC / R / SModelUncertaintyNote
Answer common FAQ-type queries3 / 1 / 0automatelowChatbots and AI agents handle a growing share.
Summarise a ticket history before a call3 / 0 / 0automatelowA classic augmentation.
Calm an angry customer whose order failed twice0 / 3 / 2protectlowEmotional labour; customers often escalate to reach a human.
Decide whether to grant an exception refund1 / 1 / 2augmentcontestedPolicy plus judgment; rules may be codified, discretion is political.
Spot a pattern pointing to a product bug2 / 0 / 2rethinkmediumAnalytics flags; humans connect it to cause.
Tag and route incoming tickets3 / 0 / 0automatelowHighly automatable.

Skills that may travel: De-escalation → Hospitality, healthcare front office; Product knowledge → Quality assurance, technical writing; Pattern spotting → Operations analysis. Access: Employer decisions dominate; individual agents rarely choose the tools or the metrics.

Bookkeeper / accounts assistant

Keeping books for small businesses, often several clients, handling GST filings and reconciliations.

Tasks for Bookkeeper / accounts assistant
TaskC / R / SModelUncertaintyNote
Enter invoices and receipts3 / 0 / 1automatelowOCR and automation take much of this.
Reconcile bank statements3 / 0 / 2rethinklowMatching is automatable; unexplained entries need a person.
Prepare GST returns2 / 0 / 3rethinklowSoftware helps; legal accountability stays human.
Explain cash-flow trouble to a shop owner1 / 3 / 2protectmediumAdvice delivered in a way the client accepts.
Spot signs of fraud or leakage2 / 1 / 3rethinkmediumTools flag anomalies; context decides.
Chase clients for missing documents2 / 2 / 0protectmediumReminders automate; relationships get the papers.

Skills that may travel: Accuracy and controls → Audit, compliance; Explaining money plainly → Financial literacy teaching; Small-business context → Advisory, business banking. Access: Clients' records are often on paper; tools assume digital-first businesses.

Junior software developer

Early-career engineer — one of the groups most visibly affected by AI coding tools in recent labour-market studies.

Tasks for Junior software developer
TaskC / R / SModelUncertaintyNote
Write boilerplate and CRUD code3 / 0 / 1automatelowCoding assistants are strong here.
Write unit tests3 / 0 / 1automatelowAssistants draft tests; deciding what to test is the skill.
Debug a production incident2 / 1 / 3rethinkmediumTools assist search; accountability and system knowledge matter.
Clarify requirements with a product manager0 / 2 / 2protectmediumAmbiguity resolution.
Review AI-generated code for security flaws1 / 0 / 3augmentlowA newly important task.
Learn the codebase by doing small tickets2 / 1 / 1augmentcontestedIf AI does the small tickets, where does learning happen? Rethink the ladder.

Skills that may travel: Systems thinking → Solutions architecture, operations; Code review → Security, quality engineering; Debugging under uncertainty → SRE, technical support leadership. Access: Entry-level hiring decisions are made by firms, not by individual juniors.

Logistics dispatcher

Coordinating deliveries, drivers and routes for a distributor or courier hub.

Tasks for Logistics dispatcher
TaskC / R / SModelUncertaintyNote
Plan daily delivery routes3 / 0 / 1automatelowRoute optimisation software is mature.
Assign drivers to shifts2 / 2 / 2protectcontestedAlgorithmic assignment affects livelihoods; fairness matters.
Handle a breakdown in the middle of a route1 / 2 / 2protectcontestedImprovisation with people on the phone.
Track and update shipment status3 / 0 / 0automatelowAutomated tracking.
Settle disputes between drivers and shopkeepers0 / 3 / 2protectlowLocal relationships.
Forecast weekly volume3 / 0 / 2rethinklowForecasting tools help; the dispatcher knows the festival week.

Skills that may travel: Real-time coordination → Event operations, emergency management; Local knowledge → Field sales, community programmes; Fair allocation → Workforce planning. Access: Algorithmic management can shift risk onto drivers; data quality varies.

Paralegal / legal assistant

Supporting lawyers with research, drafting and document management.

Tasks for Paralegal / legal assistant
TaskC / R / SModelUncertaintyNote
Summarise case documents3 / 0 / 2rethinklowSummaries are fast; omissions can be costly.
Search precedents2 / 0 / 3rethinklowLegal AI tools can fabricate citations; verification is non-negotiable.
Draft standard agreements3 / 0 / 2rethinklowTemplates plus review.
Interview a client about what happened0 / 3 / 3protectlowTrust, disclosure and confidentiality.
Manage filing deadlines2 / 0 / 3rethinklowAutomated calendars help; missed deadlines are owned by people.
Check a contract for jurisdiction-specific risk1 / 0 / 3augmentlowJudgment and context.

Skills that may travel: Careful reading → Compliance, editorial, research; Confidential handling → Investigations, HR; Procedural knowledge → Court administration, legal-aid work. Access: Client confidentiality rules limit tool use; legal-aid settings have fewer resources.

03 / Worked cases

Three complete walk-throughs

Fictional composites, clearly labelled. They show the reasoning, not anyone's real outcome.

Fictional case · Delhi

Meena's kirana and the WhatsApp catalogue

Situation. A neighbourhood shop owner is offered a free AI catalogue bot that answers price queries and takes orders. Her nephew handles the shop phone after school.

Sorting. Price replies → automate. Reordering → rethink, because the bot's suggestions do not know which distributor extends credit. Udhaar decisions → protect. Khata entries → augment: the tool records, she checks against the drawer.

The mistake avoided. Letting the bot reply to regulars in a flat template. Regulars message partly to be recognised; the shop's edge is the relationship.

Outcome. She adopts the bot for after-hours price queries only, and keeps credit and complaints personal. The access check matters most: one shared phone, data packs, and an interface in English.

Fictional case · Bengaluru

Arjun, a second-year analyst

Situation. His team adopts a coding assistant. The SQL and reporting decks that took his week now take an afternoon.

Sorting. SQL, decks → automate. Data cleaning → rethink: faster, but silent errors now flow straight into decisions. Framing questions and explaining findings → protect.

The mistake avoided. Assuming saved time is safe time. The tasks he used to learn on are gone, so he asks his manager for a deliberate learning ladder: owning one causal question per quarter, reviewed by a senior.

Outcome. A conversation about apprenticeship, not just productivity — matching what the Stanford research flags about entry-level hiring.

Fictional case · Pune

Sunita, a home care worker

Situation. An agency introduces an app with voice care-logs and medication reminders, plus GPS check-ins.

Sorting. Care log → augment (if it understands Marathi). Reminders → rethink: who is accountable when a dose is missed? Bathing, observation, family coordination → protect.

The edge case. The GPS feature adds surveillance, not support — technology can intensify work as easily as lighten it.

Outcome. She uses the conversation guide to ask the agency what the data will be used for and whether time saved on logs becomes time with clients.

Instructive mistakes and edge cases

Treating exposure as loss. Exposure indices estimate what could be done with a technology. The ILO explicitly says its figures are not job-loss forecasts.
Grading the job title. “Teacher” is not one thing. A teacher's register-keeping and safeguarding sit at opposite ends of the table.
Confusing capability with adoption. Survey after survey finds wide AI use but limited enterprise-wide scaling. Cost, data, law, trust and habit all sit between demo and deployment.
“It needs a human, so it's safe.” Protected tasks may still be underpaid, intensified or monitored. Protection is a design choice, not a law of nature.
Forgetting the ladder. When routine tasks vanish, so can the practice that turned juniors into seniors.
Assuming everyone has the tool. A shop with one shared phone and a patchy connection is not a firm with an enterprise licence.

04 / Inspect the model

The rules are short enough to read

Each task carries three editorial ratings from 0 to 3. They are judgments made for this atlas, not measurements.

C — Codifiability
How much of the task follows written rules, patterns or examples a model can learn from.
R — Relational / physical presence
How much the task depends on trust, care, a body in the room or a live relationship.
S — Stakes
How costly an unnoticed error is, and who bears it.
  1. If R ≥ 2 → protect
  2. Else if C = 3 and S ≤ 1 → automate
  3. Else if C ≥ 2 and S ≥ 2 → rethink
  4. Otherwise → augment

Uncertainty rises with every rating that sits on a boundary (C or R of 1–2, S of 1–2). A task with three boundary ratings is marked contested.

What it cannot do: predict employment, wages, or your firm's decisions; account for your country's law, your contract or your bargaining power; or know your actual tasks. Change the ratings in your head, and the bins move. That is the point.

Full method, worked examples and limitations →

05 / Evidence, labelled

What 2025–26 research says — and what it doesn't

Each item is tagged: sourced fact interpretation open question. Figures are reported exactly as each source frames them.

Read Further: annotated sources →

06 / Explore related worlds

Where this question continues in the Sahilverse

Each link below was opened and checked on 26 Oct 2026.