In India, deactivating a platform worker is now a regulated decision, not an operations setting. Karnataka's gig workers Act requires valid reasons in writing and fourteen days' prior notice before an aggregator restricts a worker's access to the platform. Telangana's 2026 Act requires written reasons and seven days. If a score drives that decision, the score has to be defensible.
That is a measurement problem wearing a compliance costume. What follows walks through six decisions a team faces when an internal "reliability score" starts switching people's livelihoods off. The scenario is illustrative — a composite, not a client — but every legal requirement cited is real and sourced.
The situation
An aggregator operating in Bengaluru and Hyderabad has around 40,000 registered workers across delivery and home services. Access to work is governed by a composite score: task acceptance rate, cancellation rate, average customer stars over the last 50 jobs, and complaint count. Below a threshold, the account is blocked. Nobody wrote down why the threshold is where it is.
The scale is not hypothetical. NITI Aayog's 2022 policy brief estimated that "in 2020-21, 77 lakh (7.7 million) workers were engaged in the gig economy", expanding to "2.35 crore (23.5 million) workers by 2029-30". It put about 47% of gig work in medium-skilled jobs, 22% high-skilled and 31% low-skilled — a workforce whose access to earnings is mediated almost entirely by software.
Decision 1: Name the instrument
The score is a selection procedure: it takes observations about a person, combines them by a rule, and decides whether they can work. The Standards for Educational and Psychological Testing and the SIOP Principles apply the same logic whatever the label; the US EEOC's guidance on tests and selection procedures treats any device used to make an employment decision as a test to be justified.
Karnataka's Act closes the obvious escape route. Section 2(a) defines automated monitoring and decision-making systems as "systems which make decisions by automated means with or without human intervention". A human clicking "confirm" on a model's output does not convert the model into a judgement call.
Decision 2: Read what the statute actually demands
This is where teams over-panic and under-prepare at the same time.
Section 13(1) of the Karnataka Platform Based Gig Workers (Social Security and Welfare) Act, 2025 requires the aggregator to inform the worker, "in simple language and in Kannada, English or any other language listed in the Eighth Schedule" known to them, of "the procedure to seek information in respect of the automated monitoring and decision making parameters" that affect working conditions — the Act's own examples are "fares, earnings, customer feedback and allied information". Section 12(1) of the Telangana Platform Based Gig Workers Act, 2026 mirrors this, in Telugu, Hindi or another Eighth Schedule language.
Read it carefully: the published duty is the route to an explanation, not the weights. But the moment a worker uses that route, you need an answer that holds together. The practical consequence is a parameter register — every input, its source, its weight, its refresh rate, and the date each changed.
Decision 3: Let the notice clock redesign the policy
Karnataka
- Section 14(2): no termination or deactivation "without giving valid reasons in writing and with prior notice of fourteen days and following the principles of natural justice".
- Section 14(1): the contract must contain "an exhaustive list of grounds for termination".
- Proviso to 14(2): where the ground "includes cases of bodily harm", action can be immediate, with a prescribed right of appeal.
- Section 2(o) defines termination broadly — "materially restricting" access, including blocking and suspension. A shadow throttle is a termination.
Telangana
- Section 14: termination only "after due enquiry", with "valid reason/s in writing and with prior notice of seven days"; immediate action is allowed where a physical or mental threat to the end consumer is foreseen.
- Section 13(2): at least fourteen days' notice of "any material change in the automated monitoring and decision-making system".
That last clause is the one that reshapes engineering practice. A weekly retrain that moves thresholds is a notifiable change. The design answer is to split the pipeline in two: a safety path that can act immediately on defined, serious grounds, and a performance path that runs on a slow, documented, noticed cycle. Mixing them is what produces the fourteen-day letter nobody can write.
Decision 4: Decide what the appeal actually hears
Karnataka routes grievances about "payouts, deductions, termination etc" to an internal committee or web portal under section 22, with a fourteen-day completion window and a written Action Taken Report, escalating to the Board. Telangana gives a Grievance Redressal Officer thirty days, an Appellate Authority ninety days to appeal, and requires an Internal Dispute Resolution Committee for aggregators with at least 100 registered workers.
An appeal is only meaningful if the evidence still exists — which means storing a decision snapshot: the inputs as they stood, the threshold in force, the model version. It also means facing the measurement reality. A composite built on 50 customer ratings carries a wide error band, and two workers either side of a cut-off are often indistinguishable. Deactivating at exactly the threshold asserts a precision the data does not have. Build a band, review inside it, act only outside it.
Decision 5: Test the discrimination duty with numbers
Karnataka section 13(2) and Telangana section 12(2) both require measures to prevent discrimination by these systems on grounds of religion, race, caste, gender, place of birth or disability. "Measures" without evidence is a policy document.
The Fairwork India Ratings 2024 from IIIT Bangalore scored 11 platforms out of 10 points across five principles. Seven of the eleven earned the first Fair Management point — due process plus appeal channels covering "penalties, low ratings, and ID blocks", available to workers "who have been deactivated". Only four earned the second, which required platforms to have "institutionalised the conduct of regular, external audits to check for biases in their work allocation systems". The study interviewed 440 workers across five cities and is explicit that the sample is not statistically representative, but the gap between seven and four is the honest state of practice: appeal channels are becoming normal, bias testing is not.
The test itself is unglamorous: compute deactivation rates by group on the decision as it is actually made, not on the model in isolation, and review the gaps annually.
Decision 6: Move the measurement to the front door
Most deactivation volume is a screening decision deferred by six weeks. If complaints cluster around judgement under pressure — a refused handover, an argument at the door — then that is what onboarding should measure, before the person has earnings at stake. AI-graded scenario assessments test exactly that kind of judgement, and AssessAll's pay-as-you-go credits at ₹30 (US$0.50) make it viable to run them across tens of thousands of onboardings without a seat licence.
There is a second-order benefit. Both statutes give every onboarded worker a Unique ID portable across platforms (Karnataka section 10(4), Telangana section 10(4)), and NITI Aayog's brief recommended that platforms enable "Skill Certificates" or "Skill Passports". A verified assessment record attached to that identity is worth more to a worker than a star average they cannot carry anywhere.
The checklist
- A written parameter register for every input to the access decision.
- The information-request procedure published in the languages the statute names.
- An exhaustive grounds list in the contract, matching what the system can actually trigger.
- Two separate paths: immediate safety action on defined grounds, slow noticed action on performance.
- Stored decision snapshots — inputs, threshold, model version, date.
- Reasons-in-writing templates a non-lawyer can complete correctly.
- An appeal body with the authority and the evidence to reverse a decision.
- An annual adverse-impact review of deactivations, not of the model.
- A change-notification process triggered by threshold and weighting updates.
One caveat on currency: Karnataka notified its Rules on 19 November 2025 and, per reporting on the February 2026 notification, set the welfare fee at 1% of each payout with per-transaction caps. Telangana's Act takes effect on a date the state notifies. Confirm the operative position with counsel before changing policy — the measurement work underneath it is the same either way.
Takeaway
The new state statutes have quietly turned platform access scores into assessment instruments that must be explained, noticed, appealed and audited. Treat the score like a test — documented inputs, honest error bands, a real review path — and the compliance work is mostly already done.