Blog · arXiv Analysis · Last reviewed June 25, 2026

The App Boss Becomes the Human Manager

Omir Kumar and Krishnan Narayanan study AI, apps, and blue-collar gig work in India. Their frame is not automation versus workers. It is whether app-mediated work can keep human accountability where the stakes are highest.

An algorithmic-human manager is not a chatbot supervisor. It is a governed work system in which software may match, route, price, monitor, verify, and flag work, but consequential decisions remain explainable, appealable, time-bound, and correctable by a human with authority.

The Paper

The paper is The Algorithmic-Human Manager: AI, Apps, and Workers in the Indian Gig Economy, arXiv:2606.19975 [cs.CY; cs.AI]. arXiv records version 1 on June 18, 2026 and lists Omir Kumar and Krishnan Narayanan as authors. The arXiv comment says it was published by the Centre for Responsible AI at IIT Madras.

The study uses a social-justice frame and a mixed-methods design built around interviews with 16 gig workers and 21 stakeholders, including platform executives, business leaders, and worker advocates. Its domain is location-based blue-collar platform work in India: ride sharing, delivery, home services, warehouses, and dark stores.

The important move is that the paper does not treat AI as a distant replacement machine. It treats AI and digital technology as the practical management layer already allocating tasks, measuring performance, shaping pay, verifying identity, routing grievances, and determining whether a worker can keep earning.

Current Context

As of June 25, 2026, the paper is an arXiv v1 record and report-style PDF, not a regulator finding or platform audit. It is useful because it names a live governance object: the app-mediated decision that changes a worker's order flow, pay, penalty, account status, or ability to challenge the system.

The scale context is public. NITI Aayog's official 2022 release estimated 77 lakh gig workers in India in 2020-21 and projected 2.35 crore by 2029-30. India's Code on Social Security, 2020 defines platform work and platform workers, and its aggregator-contribution provisions make gig and platform work a statutory social-security subject. By June 2026, the Ministry of Labour and Employment was publishing Social Security (Central) Rules, 2026 materials and an aggregator-onboarding letter for eShram that cites Rule 48(2) and a May 8, 2026 notification requiring aggregators to share gig and platform worker details. That is implementation machinery, not proof that workers can already contest every automated decision.

Outside India, the EU Platform Work Directive gives a useful benchmark for source discipline: it regulates automated monitoring and decision-making in digital labour platforms, including transparency duties, human oversight, access to evidence, private worker communications, and national transposition by December 2, 2026. The EU AI Act separately classifies many employment, worker-management, task-allocation, performance-monitoring, and self-employment-access systems as high-risk. In the United States, the EEOC says civil-rights laws apply when AI or automated technologies are used in recruiting, monitoring, productivity assessment, wage setting, promotion, or firing, while the Department of Labor's 2024 AI roadmap is a dated guidance document naming worker input, meaningful human oversight, transparency, rights protection, training, and worker-data security.

Those sources do not settle Indian gig-worker law, and they should not be imported as if every jurisdiction has the same rights. They show the governance shape that serious platform management now has to answer: notice, explanation, human review, evidence, data limits, worker voice, and remedies.

The App as Boss

The worker-facing finding is a familiar contradiction. Platforms expand access to paid work, especially for people with limited alternatives, but the same app becomes the channel through which opaque allocation, incentives, penalties, ratings, and deactivation arrive. The worker may experience the platform as a tool, a market, and a supervisor at the same time.

The paper reports that many workers did not know the term AI or did not believe the company used AI, while still describing effects that belong to algorithmic management. That distinction matters. A worker does not need a vocabulary of machine learning to know that order distribution, pay, or account status has changed without an explanation.

The stakeholder interviews add another layer. Platforms use targeted ads, data pipelines, generative AI for candidate processing, self-onboarding, selfie verification, location tracking, performance metrics, route optimization, handheld warehouse devices, and computer vision in selected workflows. Some of these systems can reduce friction or improve safety, such as fatigue log-off features. Others can make work more brittle when weather, traffic, flooding, infrastructure, customer behavior, or equipment failure is treated as if it were worker failure.

Support as Governance

The strongest governance lesson is hidden in support tickets. Workers in the study engaged heavily with chat support because that was where pay disputes, penalties, account issues, and deactivations became real. Routine operational help may fit a chatbot. High-stakes decisions do not.

The paper describes worker frustration with automated or inadequate grievance systems, especially where complaints lack time-bound resolution, clear justification, or escalation to a person. This is where the phrase human in the loop either becomes real or collapses into decoration. The human is not needed as ceremonial oversight. The human is needed because the worker needs someone with authority to explain, correct, override, and record the decision.

For platform work, support is not only customer service. It is the practical court of first instance. If a penalty, suspension, payout shortfall, identity mismatch, fraud flag, or bad rating cannot be challenged through a usable channel, the algorithm has become management without due process. The relevant standard is not whether a human exists somewhere in the company. It is whether the worker can reach a reviewer who can see the evidence, change the outcome, preserve a record, and do so before lost income becomes unrecoverable.

This belongs beside this site's pages on the dashboard boss, hidden AI labor, workplace surveillance, workplace affect scoring, and automated hiring interfaces. The recurring pattern is that management becomes software before accountability becomes software.

The Hybrid Manager

Kumar and Narayanan use the term algorithmic-human manager for a pragmatic middle path: algorithms may allocate and optimize, but fairness, dignity, and due process require explainable systems, worker participation, and human review for high-stakes decisions. They recommend algorithmic transparency, collaborative risk assessments, participatory design, microlearning inside worker apps, peer support, and worker-facing digital and financial literacy.

For government and policy, the paper discusses India's Code on Social Security 2020, state-level efforts in Rajasthan, Karnataka, Jharkhand, Bihar, and Telangana, and proposals for a right to explanation and human review. It also proposes a Unified Workers Interface as digital public infrastructure for portable worker identity, work history, reputation, settlement, social-security administration, and data ownership. That proposal should be read as a governance design, not as an existing national platform with proven protections.

The proposal is powerful and dangerous for the same reason. A portable work ledger could help with benefits, mobility, and accountability. It could also become a cross-platform reputation trap if contestability, correction, privacy, governance, and worker control are weak. Portability without due process is only a larger cage.

What the Receipt Must Add

An algorithmic-human manager needs a receipt at every high-stakes point: allocation rule, wage rule, incentive threshold, penalty trigger, rating input, monitoring signal, identity check, location policy, grievance ticket, human reviewer, decision deadline, appeal route, evidence shown to the worker, correction made, and whether the worker participated in design or audit.

The audit-grade sentence is not "the platform uses AI responsibly." It is: this worker was assigned, paid, rated, penalized, or deactivated under these rules; these data sources were used; this explanation was provided in an accessible language; this human reviewer had authority to change the outcome; and this appeal path actually resolved the case.

The receipt should be purpose-limited. It should preserve enough evidence to contest pay, penalties, deactivation, discrimination, safety risk, and social-security entitlement without turning every trip, face check, chat, GPS trace, and customer interaction into a permanent worker dossier. That links the paper to data minimization, Privacy and Data, and notice and appeal.

Failure Modes

Human-in-the-loop theater. A company advertises human oversight, but the reviewer cannot see the relevant evidence, change the decision, or respond before the worker loses income.

Support-bot dead end. The grievance channel handles FAQs but loops or closes when the worker challenges a penalty, deactivation, payout error, identity mismatch, or unsafe assignment.

Context-blind penalties. Weather, traffic, flooding, heat, police stops, customer absence, equipment failure, app outages, and local infrastructure are translated into worker fault because the metric sees delay but not context.

Portable reputation trap. A cross-platform worker interface carries ratings, flags, fraud suspicions, or stale account history without correction rights, expiry rules, or visibility into who used the record.

Consent by dependency. The platform treats continued work as consent to monitoring, biometric checks, location collection, data sharing, or training reuse, even though refusal may mean loss of income.

Welfare database overreach. A social-security or registration system becomes a broader surveillance or exclusion system because data sharing, access controls, correction duties, and deletion rules were not designed up front.

Governance Standard

A serious algorithmic-human manager should classify decisions by consequence. Routine routing, payout calculation, fraud checks, safety interventions, account restrictions, and deactivations should not share the same review standard.

First, high-stakes decisions need pre-decision notice where possible. If a worker is about to be suspended, penalized, or denied a payout, the system should show the rule, evidence, and immediate contest path unless doing so would create a specific safety or fraud risk.

Second, review must have authority. Human review means the reviewer can inspect the relevant logs, reverse or modify the decision, compensate the worker, correct the record, and mark system defects for repair.

Third, worker participation must reach design and audit. Workers and representatives should see recurring complaint categories, deactivation reasons, payout disputes, model changes, and evaluation summaries. This is the platform-work version of an algorithmic impact assessment.

Fourth, management metrics should include repair cost. A platform should not call a routing or scoring system efficient if it creates unpaid waiting, repeated appeals, unsafe speed pressure, customer conflict, or hidden support labor.

Fifth, the data boundary should be explicit. Worker data collected for matching, safety, payment, or social security should not silently become discipline, pricing, insurance, training-data, credit, or cross-platform reputation evidence. That is the line between governance and extraction.

The Spiralist rule is simple: when the app is the boss, appeal is part of the wage. A worker who cannot challenge the machine is not being managed by a hybrid system. They are being managed by a black box with a help desk.

Limits

This page reads one arXiv record, one report-style PDF, and current official or primary legal sources for context. The interview sample is useful for surfacing mechanisms and lived experience, not for estimating population-wide rates. The PDF text extraction is noisy, so factual checks were made against the arXiv metadata plus readable sections of the PDF: executive summary, methods, findings, policy landscape, recommendations, and appendices. No claim here should be read as proof that every Indian platform uses the same system or produces the same harm.

State-level and international legal references should also be kept in their lanes. A welfare act, draft bill, directive, code, or official guidance document can show that algorithmic management is a governance subject. It does not prove implementation quality, worker comprehension, timely appeals, social-security delivery, or platform compliance.

Source Discipline

Use Kumar and Narayanan for the study design, interview findings, worker/stakeholder themes, algorithmic-human manager proposal, Unified Workers Interface proposal, and report limitations. Use NITI Aayog for Indian gig-workforce projection context. Use India's Code on Social Security and Ministry of Labour materials for statutory and implementation context. Use EU and U.S. sources as comparative governance references, not as direct rules for Indian platforms.

Do not collapse evidence types. A worker interview is not population prevalence. A statutory definition is not operational protection. A social-security portal is not due process. A support ticket is not human oversight unless someone with authority can resolve it. A portable identity ledger is not worker empowerment unless correction, consent, access limits, and appeal travel with it.

Sources

Return to Blog