Industries / Construction and property / Elevator and escalator services
Breakdowns and callouts
Breakdowns and callouts are the unplanned side of the trade: freeing people trapped in stopped cars, answering alarm calls, and getting broken lifts and escalators running again.
6
Milestones achieved, from 1999 to 2018.
1
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1999
Forty-one hours in a car
On a Friday night in October 1999, Nicholas White, a production manager at Business Week, was trapped for 41 hours in an express elevator in the McGraw-Hill Building in New York. The New Yorker reports that the intercom gave no answer, the alarm bell on the car roof could not be heard far, and eight security guards came and went without noticing him on the camera monitor; he never learned why the car stopped, though there was talk of a power dip.
LessonAn alarm helps only if it reaches someone who is listening.
Credit
reported by Nick Paumgarten in The New Yorker
2005
A quake stops 64,000 lifts
On 23 July 2005 an earthquake centered in northwest Chiba stopped about 64,000 lifts maintained by Japan's five largest service firms, almost all of them halted by their earthquake control devices, which take the car to the nearest floor and open the doors. People were trapped in 78 cars in the Tokyo region, 73 of which had such a device, and were freed on average just under 50 minutes after the call, up to 170 minutes, by the Japan Elevator Association's count as published by the transport ministry (MLIT). The ministry thought a door safety stop had acted first. A government council in 2006 called for devices that sense the first, weaker P-wave of a quake, and earthquake control devices became required on new lifts from 28 September 2009.
LessonA device that parks cars at a floor does not stop trapping if another safety device stops the car first.
Credit
Japan Elevator Association and Japan's Ministry of Land, Infrastructure, Transport and Tourism
2015
Escalator floor plate gives way
On 26 July 2015, at the Anliang department store in Jingzhou, Hubei, a loose cover plate at the top of an escalator flipped as a woman stepped on it, and she fell into the machinery space below and died. The technical investigation found that the maker, Shenlong Elevator, had designed plates that could loosen and lift, and that three plates did not match their drawings. Shenlong said 4,648 escalators of the same model in 31 provinces would get four added brackets under each plate by 31 August, and Hubei, Shanghai, Guangzhou and Inner Mongolia ordered Shenlong escalators stopped for checks.
LessonA design fault in one part is present in every unit of that model, so one failure calls for a check of the whole fleet.
Credit
the Hubei accident investigation, reported by Xinhua through The Paper
2016
The city fixes neglected elevators
Under New York City's Local Law 101 of 2015, in effect from 1 July 2016, the Department of Buildings (DOB) refers elevators that owners leave broken after a hazardous-condition violation to the housing department (HPD), which can hire a contractor to repair them and bill the building; unpaid bills become a tax lien. In the year to June 2025, by HPD's count, 773 apartment buildings received immediately hazardous elevator violations, 176 were referred, owners fixed 134 of those an average of 56 days after referral, and HPD repaired 3 itself for $64,960.
Credit
New York City Council, HPD and DOB
2017
An escalator runs backward
On 25 March 2017 a long escalator carrying shoppers up at the Langham Place mall in Mong Kok, Hong Kong, reversed and injured passengers. Hong Kong's Electrical and Mechanical Services Department (EMSD) found that the main drive chain had broken from metal fatigue and that the device meant to detect a broken chain did not apply the backup brake, because sticky grease had built up in it and one of its two springs had been locked by a lock nut. EMSD had all 64 escalators in Hong Kong with a rise of 15 m or more specially inspected and suspended an engineer and a worker for six months.
LessonA backup safety device has to be tested in the state it will be used in, because a jammed one gives no warning until it is needed.
Credit
EMSD's technical investigation
2018
Osaka quake traps 339 cars
On 18 June 2018 an earthquake in northern Osaka stopped about 66,000 of some 122,000 lifts maintained by Japan Elevator Association members in five prefectures, and trapped people in 339, of which 155 had earthquake control devices, by the transport ministry's count. Most of those cars were stopped by the main shock before they reached a floor. Rescue took about 80 minutes on average and up to about 320 minutes, as jammed phone networks and traffic slowed mechanics. In response the industry restores one lift per building before moving on, and trains firefighters and building staff to free people where that is safe. A 2021 review by Satoshi Fujita and colleagues gives 346 trapped lifts and about 63,000 stopped.
LessonIn a large quake, rescue time depends on how fast anyone can reach the building, not on the lift.
Credit
Japan Elevator Association and MLIT
Watch for
Cameras spot trapped riders
The National Capital Region Transport Corporation (NCRTC), which runs India's Namo Bharat regional trains, says an AI system it built in-house watches lift cabins through its existing CCTV cameras, detects an entrapment within a minute, alerts the control center and station staff, sounds an alarm at the lift, and raises the alert to "critical" if nobody is freed within 60 more seconds. No figures on entrapments caught, missed or falsely flagged have been published.
Achieved whenNCRTC or an outside party publishes, for at least several months of service, the number of entrapments detected and missed and the number of false alarms, checked against the lifts' fault logs.
Claimed by
NCRTC, reported by the Press Trust of India, April 2026
Open
01
Staged lift alarm calls
A public set of at least 1,000 staged lift alarm calls, recorded by volunteer actors through a lift emergency telephone on a bench and passed over a phone line, in at least five languages. Each call is labeled: trapped and calm, trapped and panicking, medical emergency, no speech (a caller who cannot talk), misuse or prank, or an accidental press. The set is used to score speech-to-text and call-sorting software on how often it misses a real emergency.
The briefCredit · unclaimed02
A ledger of quake trappings
A public table of every earthquake for which a government or industry body has published lift figures: lifts stopped, lifts with people trapped, average and longest rescue times, lifts damaged, and the area and companies counted, with each number linked to its page and its definition written down.
The briefCredit · unclaimed03
How long outages last
An estimate of how long each reported elevator outage in New York City lasts, built from public records joined by building number: 311 complaints and their closing dates, DOB complaint outcomes, violations and when they were cleared, and device filings. The estimates are then checked against outages whose start and end dates are known.
The briefCredit · unclaimed
Open 01 · the brief
Staged lift alarm calls
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 1,000 staged lift alarm calls, recorded by volunteer actors through a lift emergency telephone on a bench and passed over a phone line, in at least five languages. Each call is labeled: trapped and calm, trapped and panicking, medical emergency, no speech (a caller who cannot talk), misuse or prank, or an accidental press. The set is used to score speech-to-text and call-sorting software on how often it misses a real emergency.
Why it is still open
The closest work, a 2021 dataset from Duke Kunshan University and KONE, holds 3,724 audio and video clips from 266 actors, but only in Mandarin and built around one call-for-help word, and the papers do not say it can be downloaded. Bosch Service Solutions offers AI translation of lift emergency calls and publishes no test of it.
Built with what exists
A new lift emergency telephone of the kind sold for retrofits, wired to a test phone line on a desk, never fitted to a lift; open speech recognition such as OpenAI's Whisper, and ElevenLabs' Scribe v3 (see the Note "Fewer word errors"); a model that gives probabilities for fixed answers, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"); a signed consent form for every actor.
First step
Write 40 scripts across the six labels and record them with ten volunteers through one emergency telephone on a desk.
Done looks like
At least 1,000 labeled calls published under an open license, with a scorecard for at least three systems that gives, by language and by label, the share of real emergencies missed and the share of false alarms raised.
Open 02 · the brief
A ledger of quake trappings
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of every earthquake for which a government or industry body has published lift figures: lifts stopped, lifts with people trapped, average and longest rescue times, lifts damaged, and the area and companies counted, with each number linked to its page and its definition written down.
Why it is still open
The figures sit in separate reports and do not agree. For the 2018 Osaka quake, the transport ministry counts 339 entrapments among Japan Elevator Association members in five prefectures, while Fujita and colleagues give 346 trapped lifts. Their 2021 review covers Japan only, and Gu and colleagues (2023) estimate trapping on one university campus by simulation rather than by collecting records.
Built with what exists
Reports from MLIT, the Japan Elevator Association, Hong Kong's EMSD and published field surveys from other countries; PyMuPDF and pdftotext to pull text from PDFs; machine translation; a spreadsheet with one column per definition; an AI agent that cites the page behind each value, such as one built with Cursor's agent kit (see the Note "Findings that cite notes"), checked by hand.
First step
Enter the 2005 Chiba and 2018 Osaka figures from the ministry's reports, with both 2018 counts, and list the published reports for the 2011 Tohoku and 2016 Kumamoto quakes.
Done looks like
At least 12 earthquakes from at least three countries, every value linked to a page and checked by a second person, and a published list of every disagreement between sources.
Open 03 · the brief
How long outages last
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An estimate of how long each reported elevator outage in New York City lasts, built from public records joined by building number: 311 complaints and their closing dates, DOB complaint outcomes, violations and when they were cleared, and device filings. The estimates are then checked against outages whose start and end dates are known.
Why it is still open
Elevator Uptime, a site that already lists 311 and DOB elevator complaints for each building, says it does not publish uptime or time to fix where inspection records are its only source, and calls the complaints "unverified signals". HPD's yearly report gives an average repair time (56 days) only for the 134 buildings fixed after referral in the year to June 2025. The New York City Housing Authority reports an average outage of 6 hours 52 minutes for its own public-housing elevators in the first four months of fiscal 2026, but nothing comparable exists for private buildings. A 2026 GitHub project analyzes 22,504 elevator complaints made through 311 by borough and season, not by how long the outages lasted.
elevatoruptime.comnyc.govdata.cityofnewyork.usnyc.govgithub.com
Built with what exists
NYC Open Data's API for the 311 Service Requests and DOB NOW Elevator Safety Compliance datasets; DuckDB or pandas; a mapping library such as Leaflet.
First step
Download the 21,424 elevator complaints made through 311 in 2025 (13,153 for a building's only elevator and 8,271 for buildings with several, by a query of the city's data for this file) and join them to the device records by building number.
Done looks like
An estimated outage length for every 2025 elevator complaint, a published method, and an error rate measured against at least 50 outages whose dates are known from court filings, news reports or HPD records.