Skip to content
SME AI Impact
Innovation ExcellenceTransport & Mobility· Computer vision

GetGo Technologies

GetGo uses computer vision to attribute car damage to the right booking, recovering S$20,000 in two months.

Principal published result

53% reduction

Damage-to-invoice lead time

Evidence grade

BCanonical source supported

Grade B — Canonical source supported

Research coverage

95% coverage

Ten evidence dimensions, not an AI maturity score

Sources

2

2 sources · 1 beyond T0

Open getgo.sg ↗Live site, not an official page of this showcase.

Share on LinkedIn

01The official record

Award citation, verbatim in substance
Homepage of GetGo Technologies, captured 2026-10-18
GetGo Technologies's homepage, captured 2026-10-18. Shown for identification and comment; it may differ from the award-winning implementation.Open live site ↗

Provides on-demand carsharing on a pay-as-you-drive basis

Published key challenges

  • Impractical manual screening. Around 7,000 daily bookings generate about 56,000 end-trip photos, making it impractical to manually inspect every booking for damage
  • Delayed damage detection. Vehicle damage could go unnoticed until reported by users or discovered during workshop inspection days or weeks later
  • Difficulty recovering damage costs. Late detection made it difficult to attribute damage to the responsible booking, resulting in cost of repair that cannot be recovered

Project Desmond: AI-powered Vehicle Damage Detection

  • AI vehicle damage detection: Analyses end-trip photos to detect billable damage such as scratches, dents, cracks and dangling parts
  • Booking-level damage attribution: Compares consecutive bookings to determine when damage first appeared and link it to the relevant booking
  • AI-powered image matching: Uses image embeddings to determine whether detected damage is new or recurring across adjacent bookings
  • Human-validated damage workflow: Routes AI-flagged cases to human reviewers for validation before confirmed cases proceed for follow-up and billing
  • Continuous AI model improvement: Retrains the model using real-world failure cases such as rain and reflections to improve precision and reduce false positives

02SCQA case study

396 words
SSituation

GetGo Technologies runs an on-demand carsharing service on a pay-as-you-drive basis. A vendor case study published by Zendesk in 2024 records that the fleet exceeded 2,400 cars across more than 1,500 locations. Every rental begins and ends with photographs of the vehicle, taken by the customer. Those photographs are the only record of the vehicle's condition at the moment of handover.

CComplication

The citation states that around 7,000 daily bookings generate about 56,000 end-trip photographs. Inspecting every booking manually is not feasible at that ratio. Damage could therefore go unnoticed until a customer reported it or a workshop found it days or weeks later. Late detection broke the link between damage and the booking that caused it, so the cost of repair could not be recovered from the responsible renter. The company carried the loss, and the invoice for damage arrived long after the rental ended. The company also carried the cost of disputes, because a charge raised weeks later is harder to defend than one raised at handover.

QQuestion

How could GetGo inspect tens of thousands of handover photographs a day, decide which booking caused any damage, and bill it accurately without accusing the wrong customer?

AAnswer

GetGo built Project Desmond, a damage detection system that combines computer vision, image matching and human validation. It analyses end-trip photographs and detects billable damage such as scratches, dents, cracks and dangling parts. It then compares consecutive bookings to establish when damage first appeared, and uses image embeddings to judge whether a detected defect is new or a recurring one already on file. Flagged cases route to human reviewers, who validate before any case proceeds to follow-up and billing. The model is retrained on real failure cases, including rain and reflections, to raise precision and cut false positives. The citation reports a 53% reduction in the lead time from damage to invoice, 60% less manual review work, and S$20,000 recovered in two months. The design keeps a person between the model and the customer, which matters when the output becomes a charge. The citation does not publish the model's precision or the dispute rate, so the commercial effect of a false positive is not documented. The recovered amount is a floor rather than a ceiling, because it measures only damage billed within two months, and says nothing about damage prevented, disputes avoided, or repairs found before a vehicle left the lot.

Implementation takeaway

When AI output becomes a bill, the review step is not overhead. It is the control that makes the automation usable.

03Verified business results

Company-reported
Published business outcomes for GetGo Technologies, as reported in the award citation.
MetricPublished resultAttribution
Damage-to-invoice lead time53% reductionCompany-reported; published in the IMDA/SBF award citation
Manual review workload60% lessCompany-reported; published in the IMDA/SBF award citation
Damage costs recoveredS$20,000 within two monthsCompany-reported; published in the IMDA/SBF award citation

The award citation publishes relative changes and absolute volumes, not paired before-and-after values. Baseline figures are therefore not shown, and no improvement percentage has been calculated from them.

04Technology stack

Evidence status per layer
Technology stack for GetGo Technologies, with evidence status for each layer.
LayerImplementationVendor / in-houseEvidenceSource
AI modelComputer vision model for damage classification, retrained on failure casesIn-house, unnamedin-houseStatedT0-AWARDS
MatchingImage embeddings for new-versus-recurring damage comparisonIn-house, unnamedin-houseStatedT0-AWARDS
ProcessingBooking-level attribution across consecutive rentalsIn-house, unnamedin-houseStatedT0-AWARDS
IntegrationBooking records linked to damage cases and billingNot disclosedunknownStatedT0-AWARDS
GovernanceHuman validation before confirmed cases proceed to billingIn-house, unnamedin-houseStatedT0-AWARDS

A layer marked inferred includes a documented basis. Layers without a named vendor are left unnamed rather than filled with a plausible product.

05Implementation diagrams

Documented vs illustrative
Documented implementationEvidence-bounded
End-tripphotographsDamagedetectionImageembeddingmatchBookingattributionHumanreviewerInvoice andfollow-upFailure-caseretrainingINPUTPROCESSAIHUMANOUTPUT

Components named in the award citation.

Illustrative reference architectureIllustrative only
GuidedphotocaptureImagequalitygatePanelsegmentationDamageclassifierComparewithprevious…Agentreview anddecisionCharge orcloseINPUTPROCESSAISYSTEMHUMANOUTPUT

A feasible build for another fleet operator. This is not GetGo's production system.

06Unconfirmed — technical inference

Not attempted

No public technical details were identified beyond the official award description. This case is not supported well enough to justify a technical inference, so none is offered.

07How another SME could build something similar

Advisory, not the winner's method

This section is guidance for other companies. It is not a description of what GetGo Technologies did, and it is not endorsed by them.

Suitable for

  • · Carsharing and rental fleets
  • · Logistics fleets
  • · Equipment hire
  • · Property handover inspections

Minimum requirements

  • · Standardised photo capture at handover
  • · A labelled set of damage and no-damage images
  • · A review team with authority to waive a charge

Three implementation paths

Level 1

Guided capture and manual review

Standardise photographs and compare them side by side in a review tool.

Use when: Fleet is small and disputes are the main problem.

Level 2

Assisted detection

Use a managed vision API to pre-flag likely damage, with every case reviewed by a person.

Use when: Photo volume exceeds manual capacity.

Level 3

Attribution and billing

Add embedding-based comparison across handovers and integrate decisions into billing.

Use when: Detection precision is proven and recovery value justifies the build.

Suggested sequence

  1. 1Standardise the capture protocol before modelling anything
  2. 2Label a few thousand images for the damage types you will bill
  3. 3Run detection in shadow mode and compare with human findings
  4. 4Add attribution once detection precision is acceptable
  5. 5Pilot billing on one city or vehicle class

Tools and alternatives

  • · Vision model or managed vision API
  • · Image embedding store
  • · Booking or asset management system
  • · Review queue

Risks and PDPA considerations

  • · False positives that charge innocent customers
  • · Poor photos at night or in rain
  • · Bias against particular vehicle colours or lighting

KPIs to track

  • Precision and recall by damage type
  • Dispute rate
  • Days from damage to invoice
  • Recovery value per month
  • Reviewer minutes per case

08Research gaps

What remains unknown
  • No model, cloud or vendor names are published.
  • The 56,000-photo daily figure is stated without a measurement period.
  • Precision and recall of the damage model are not disclosed.

Are you this company? Submit documentation to upgrade your evidence grade.

If you have published technical documentation, vendor whitepapers, or third-party audits, submit documentation to our research desk. We update evidence grades upon verification.

09Sources

2 records
  1. Infocomm Media Development Authority (IMDA) and Singapore Business Federation (SBF) · 16 October 2026 · Retrieved 18 October 2026 · T0 — Canonical awards source

    Canonical source. Page last updated 16 October 2026. Publishes one profile per winner: company description, three key challenges, named AI solution, solution components and three headline metrics. No individual winner URLs are published.

  2. Zendesk · 4 January 2024 · Retrieved 18 October 2026 · T2 — Named implementation partner

    Vendor case study naming GetGo. States GetGo operates more than 2,400 cars across more than 1,500 locations. Concerns customer support, not damage detection.

Researched 18 October 2026 · Updated 18 October 2026 · Award citation retrieved 18 October 2026.

GetGo Technologies is a winner of the SME AI Impact Awards 2026, organised by IMDA and SBF. This case study is independent analysis and is not an official company publication.

Related case studies

Similar technology and sector implementations

Homepage of TTS Group, captured 2026-10-18
Adoption ExcellenceTransport & Mobility
Winner · Award Cohort 2026

TTS Group

TTS Group replaced manual traffic counts with computer vision, processing about 50 times more sites a month.

Traffic-data acquisition capacity
About 98% increase
Computer visionBCanonical source supported
Homepage of Greenphyto, captured 2026-10-18
Innovation ExcellenceAgriculture
Winner · Award Cohort 2026

Greenphyto

Greenphyto runs a 23-metre indoor farm on computer vision, cutting crop inspection time tenfold.

Crop inspection time
10x faster
Computer visionAIndependently supported
Homepage of Accountingbpo, captured 2026-10-18
Innovation ExcellenceBusiness Services
Winner · Award Cohort 2026

Accountingbpo

Accountingbpo turned a headcount-bound accounting back office into a supervised AI workflow.

Accounting document processing time
Up to 67% reduction
Document processingBCanonical source supported