AI Damage Detection for Car Rental: Why Human-in-Loop Is the Only Safe Design

The operational workflow for computer-vision damage detection at a mid-market car rental return lane has three stages: scan, compare, and flag. At the pre-rental stage, the vehicle drives through — or is walked around by — a scanning system from a vendor like Ravin AI. Drive-through scans operate at up to 30 km/h; mobile 360° scans complete in 30 seconds to 1 minute. Each scan produces a timestamped photographic record of the vehicle's current condition, including any pre-existing damage already in your system from previous rental cycles. This record is written to the vehicle's PMS record and linked to the active rental agreement. At the return stage, the same scan is performed and the CV model compares the return condition to the pre-rental baseline. Any anomaly detected by the model — a new scuff, a dent, a broken mirror — is flagged with photographic evidence and a confidence score. The flag goes to a human reviewer, not directly to an invoice. The reviewer looks at the pre/post comparison, confirms whether the anomaly represents genuine new damage or a pre-existing condition that was not logged (or a lighting artefact), and either approves a charge or dismisses the flag. If a charge is approved, the evidence package — pre-rental scan, post-rental scan, timestamps, reviewer identity, and sign-off — is attached to the invoice and available for any customer dispute. The speed improvement is significant at high-volume return lanes: drive-through scans at up to 30 km/h, or 30-60 seconds for a 360° mobile scan, versus 8-15 minutes for a thorough manual walkaround means a single return lane can process multiple vehicles in the time a manual check takes for one. But the speed improvement is only commercially valuable if the workflow is correctly designed — specifically, if the human review step is built into the process rather than treated as optional. The Hertz/UVeye situation is the industry case study for what happens when that step is bypassed.

The Hertz/UVeye situation is the car rental industry's most important AI implementation case study of the past decade. During FY24, Hertz integrated UVeye's automated vehicle scanning system at scale. After rollout, the FY24 Hertz/UVeye situation, widely covered in industry press, was followed by a class-action lawsuit and FTC scrutiny. Customers reported receiving damage charges — some for damage they did not cause, and in some cases for damage that predated their rental — with limited recourse and minimal human review in the charging workflow. The case settled in 2025, per published reports of the class-action proceedings. The core issue was not the scanning technology itself. UVeye's computer vision capability is technically capable — the vendor serves automotive OEMs and large fleet operators — and the photographic evidence it produces is high quality. The issue was the workflow design: damage flags were moving from AI detection to customer charge without a consistent human review step confirming the damage was new, genuine, and attributable to the current rental. For mid-market operators evaluating AI damage detection in 2026, the practical takeaway is not avoid the technology but design the workflow correctly. BVRLA's Code of Conduct for UK operators establishes that damage charges must be evidenced, communicated clearly to the customer, and subject to a dispute process. ACRA provides equivalent guidance for US operators. Both frameworks assume a human decision point between the AI flag and the customer charge. Any AI damage detection system that routes directly from scan to invoice — without a named human reviewer confirming each charge — is non-compliant with both standards and creates the same legal exposure that Hertz encountered. The settlement is the industry's current benchmark for the legal and reputational cost of getting this wrong.

A compliant workflow has four mandatory steps, in sequence. Step 1: pre-rental scan, confirmed by a human agent who signs off that the vehicle's condition at rental start is accurately recorded and the customer has had the opportunity to review and dispute any pre-existing damage logged. Step 2: post-rental scan, performed at return, with the AI comparison model generating a flagged anomaly report within seconds. Step 3: human review of every flagged anomaly — a named agent looks at the pre/post photographic comparison and makes a binary decision: this is new damage attributable to the current rental, or this is pre-existing / artefact / unclear, dismiss. No charge is raised without this step. Step 4: if a charge is approved, a structured customer notification is sent with the photographic evidence, the charge amount, and the dispute process — in writing, before the charge is applied to the payment method. This four-step design is not bureaucratic overhead — it is the minimum that BVRLA and ACRA standards require, and it is what the Hertz settlement made explicit as an industry standard. The practical implementation for a 30-200 vehicle operator means your counter team needs a workflow tool — not just a camera — that enforces the sequence. The scan generates the evidence; the tool routes it to a reviewer; the reviewer approves or dismisses; the system records the decision. Any configuration that skips step 3, or makes it optional, creates legal exposure. Vectimo's design for damage detection implementations builds this sequence into the PMS integration rather than treating it as a training recommendation. The human review step is not a policy that can be bypassed by a busy return lane; it is a system gate that cannot be bypassed. The evidence package — pre-rental scan, post-rental scan, reviewer identity, sign-off timestamp, and customer notification record — is stored in the vehicle's PMS record for a minimum of 12 months for dispute reference. For operators on TSD, RentWorks, Coastr, or RENTALL, the integration layer writes these records automatically.