ReviewDesk guide · By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
How do automated-decision duties apply to gig platforms and marketplaces?
Generate a free adverse-outcome letter — no sign-up Build a free pre-use notice See pricing
The short answer. When a platform's systems allocate work, set visibility or deactivate accounts, they make consequential decisions about people's livelihoods — continuously and at a scale no human team reviews case by case. The ADMT pattern still applies, and platforms are structurally best placed to honour it: notice inside the app, explanations generated from the decision record, correction for the telemetry errors platforms generate, and human review with authority to reinstate. ReviewDesk's workflow model maps onto platform operations with the clocks doing the heavy lifting.
What is ADMT on gig platforms?
Gig-platform ADMT covers automated systems that materially influence a worker's access to income on a platform: work allocation and dispatch priority, search and visibility ranking, performance scoring that triggers consequences, and automated or semi-automated deactivation. Employment and livelihood decisions sit within the consequential areas the tracked regimes address; the precise employment classification questions are for counsel — the notice, explanation and review pattern is the operating answer regardless.
Which platform systems carry the pattern?
Apply the discretion test per system, measured rather than asserted. Dispatch that consistently offers the best work to the highest-scored workers is allocation with livelihood consequences. Ratings aggregation that triggers warnings, throttling or deactivation below thresholds is a decision pipeline. Fraud and safety systems that suspend accounts automatically are the sharpest case: the consequence is total and immediate, and the affected worker's need for a fast human route is at its maximum. Pure matching with no consequence gradient — showing all available work to all eligible workers — sits outside, and the worksheet should record why.
Platforms should expect the measurement to be unflattering in places. A 'suggestion' ranking followed by ninety-eight per cent of workers is a decision. The scope exercise in the scope-test guide, run on platform telemetry, usually reclassifies at least one system the organisation had filed as informational — and that reclassification, self-discovered, is the cheapest version of the lesson available.
How does notice work inside an app?
As a product surface, versioned and logged per worker: the systems that affect allocation and standing described in plain terms — what is measured, what it influences, what it never decides alone — presented at onboarding and re-presented when material changes ship. Platforms have a distribution advantage no other sector enjoys: every affected person opens the channel daily. They also have a distinctive failure: notice buried in terms-of-service updates that no worker can reconstruct later. Keep the standing notice at a stable in-app destination and a hosted page — the pattern of the hosted-notice guide — with versions in the record.
State the human architecture in the notice itself: which consequences always get human review before they land (deactivation is the obvious candidate), which the worker can appeal after, and the response times the platform works to. Platforms live on worker trust economics; a notice that reads as a genuine operating description is a retention asset, and workers compare platforms on exactly this once they have been burned by an opaque one.
What do explanations look like at platform scale?
Generated, specific and fast — the letter pipeline pattern from the fintech playbook, transposed. A deactivation or throttling explanation must name the triggering pattern in world terms: the events, the dates, the threshold, the data as recorded — because platform data is telemetry, and telemetry has known failure modes (GPS drift, duplicate reports, sensor gaps) that workers can often identify instantly when the record is shown to them. 'Violation of community standards' is to platforms what 'did not meet our criteria' is to lenders: a non-explanation that guarantees escalation.
Speed is a design requirement, not a courtesy. A deactivated worker is losing income daily; the explanation and the review route must move in days, with internal clocks set far inside any statutory frame — the tempo argument from the healthcare guide applies with equal force. Platforms that can dispatch in milliseconds can explain in a day; choosing not to is a prioritisation the record will eventually describe.
How does correction work with platform telemetry?
By giving workers a readable copy of their own record and a route that accepts their evidence: trip logs, delivery confirmations, communications that contradict a reported pattern. Telemetry corrections have a distinctive shape — the platform's sensors are usually the wrong party, so verification is often a data-join exercise inside the platform's own systems, fast when the tooling exists. Build the correction route into the same case docket as review, per the correction guide, and let corrected records re-run the affected scoring with the sequence appended, so a worker's standing is restored by the system that damaged it, visibly.
Aggregate the corrections as platform intelligence. A sensor or join defect that mis-scores forty workers in a city will present as forty individual appeals unless the corrections log is read in aggregate — platforms, of all organisations, should recognise a data-quality incident when one files itself forty times. The corrections review belongs in the platform's reliability rhythm, beside the incident reviews, because for the workers affected it was an incident.
What does human review require on a platform?
Reviewers with reinstatement authority, the full record (not the score summary), and service levels matched to income loss. Platform review has one structural advantage: the file is complete by construction — every event the system used is in the platform's own logs — so a reviewer can genuinely re-derive the decision, meeting the standard in the review checklist without the archaeology other sectors face. The corresponding temptation is review-by-replay: re-running the same pipeline and calling the identical output a review. Guard it by requiring the outcome note to engage the worker's own submission, which a replay cannot produce.
Publish the review statistics to workers in aggregate — volumes, median times, outcomes changed — as platforms publish other trust metrics. The aggregate honesty does double duty: it disciplines the operation (numbers watched are numbers managed) and it differentiates the platform in a labour market where experienced workers actively price opacity. A platform whose review system is real can afford to describe it; the description then becomes a promise the metrics must keep, which is a healthy loop to live inside.
How do platforms sequence the build?
Start at the sharpest consequence: deactivation. Notice, generated explanation, correction and review for deactivation decisions first — the workflow where harm is total and the pattern is most testable — then allocation and scoring systems in consequence order. Instrument traces per consequential event from the start, because platform volumes make retrospective reconstruction impossible: a million allocation decisions cannot be re-derived, only recorded. The trace discipline from the pack guide is, for platforms, simply the decision to keep the logs they already generate in evidence shape.
Staff for the surge pattern platforms know well: consequences cluster (a model change, a city launch, a policy update) and so do requests. The bench model — named owners, weighted queues, escalation on read — is in the assignment guide. Platforms already run incident response for systems; decision review deserves the same operational seriousness, because to the worker on the other end, it is the incident.
Your five actions for ADMT on gig platforms
- Run the scope test on platform telemetry per system — allocation, scoring, deactivation — and file measured determinations.
- Keep the standing notice at a stable in-app and hosted destination, versioned, stating the human architecture plainly.
- Generate explanations from the decision record naming events, dates and thresholds; set internal clocks in days for income-stopping decisions.
- Give workers their record and a correction route that re-runs scoring on corrected data, with corrections read in aggregate as data-quality intelligence.
- Staff review with reinstatement authority and require outcome notes that engage the worker's own submission — never review-by-replay.
Worked example: a deactivation reversed on the platform's own logs
A fictional delivery platform — SwiftCart — automatically deactivates a courier whose account shows three 'failed delivery' events in a week, a pattern its standards treat as abandonment. The generated explanation names the three events with timestamps. The courier's review submission identifies the pattern instantly: all three are the same apartment complex whose gate system logged his arrivals under a neighbouring address — his own delivery photos carry the correct geotags. The reviewer verifies from the platform's logs in one sitting, the record is corrected, standing and the week's earnings treatment are restored, and the case closes in three days.
The aggregate read matters as much: the corrections log shows eleven couriers flagged at the same complex. The address-matching defect is fixed as a data incident, the eleven files are corrected proactively — before several of those couriers had even appealed — and the platform's worker-facing trust report gains a story it could not have bought. One worker's three timestamps, taken seriously, repaired a defect the platform's own monitoring had filed as eleven separate character judgements.
Frequently asked questions
Are gig workers' platform decisions covered by ADMT frameworks?
Livelihood and employment decisions sit within the consequential areas the tracked regimes address. Classification questions are for counsel, but the operating pattern — notice, explanation, correction, human review, records — is the answer platforms should build regardless.
Which platform systems should be scoped first?
Deactivation and suspension systems, then allocation and scoring with consequence gradients. Measure real influence from telemetry rather than accepting 'informational' labels — a followed suggestion is a decision.
How fast should platform reviews move?
In days for income-stopping decisions. Set internal clocks far inside statutory frames; a deactivated worker loses income daily, and platforms that decide in milliseconds can explain and review at human speed by choice.
What is review-by-replay and why is it banned?
Re-running the same pipeline and presenting its identical output as a human review. The outcome note must engage the worker's own submission — evidence and circumstances the replay cannot weigh — which is the test that keeps review human.
Is this guide legal advice?
No. ReviewDesk is templates and workflow tooling, not legal advice. The pattern is an operating model drawn from the regimes in our research; consult counsel for your platform's classification and duties.
Keep reading: Why does ReviewDesk render missing evidence as 'not recorded' — and never invent it? · How do you integrate decision traces through the ReviewDesk workspace API? · Who should own an ADMT programme — owner, reviewers and record-keepers? · What are the key ADMT deadlines — Illinois now, Colorado and California 2027, the EU in December 2027? · What changed with Colorado SB 26-189 — and what does it keep from the earlier AI Act approach? · All ReviewDesk guides · ReviewDesk home · Try the free desk · Free letter generator · Free notice builder · Pricing
ReviewDesk is built by Ingenious AI. Templates and workflow tools — not legal advice.