Photo-based condition and damage triage for goods and buildings

In wholesale, haulage and property management, many small decisions hang on a few phone photos: is this return resalable, who pays for the crushed pallet, is the stain in the flat normal wear or damage. Today someone looks at the photos, searches the delivery note or lease, and writes an email. A vision-language model can sort and describe the photos, read the label in the picture, link them to the right document and draft the next step. It cannot decide liability or building damage: it prepares the file, a person decides.

Problem & value

• Today:
• Wholesale: returns arrive as boxes with a slip. Staff open them, judge whether goods are new, resalable or scrap, take photos, and decide on credit and restocking. Estimated effort in the gap analysis: about 14 h per month in a typical wholesaler.
• Haulage: drivers and warehouse staff photograph damaged pallets, but photos end up in chats or galleries. Claims need a precise note on the consignment note and a timely written notice (about 8 h per month).
• Property management: tenants send damage photos by mail or phone; the manager decides which trade to send and, at move-out, whether a deposit deduction is justified (about 5 h plus 2 h for deposit and key return).
• AI value: consistent condition classes, photos that are complete and linked to the right document, ready-to-send claim and credit drafts, fewer disputes because evidence is clean and timestamped. The time saving comes mainly from sorting, searching and writing, not from the image judgement itself.
• Evidence: no independent studies for these SMB workflows. The saving range in the front matter is an estimate from the process shares, to be replaced by your own pilot numbers.

Variants

• Return inspection (wholesale, B2B retail): photo of item, packaging and serial label at the return desk; grade and restocking proposal. Consumer re-commerce grading is VIS-037.
• Transit damage (forwarders, carriers, receivers): photo set at goods receipt or handover (pallet, wrapping, labels, damage close-up) tied to the consignment note and the delivery; claim draft against the carrier or insurer.
• Tenant damage and handover (property management): damage report triage (water, mould, broken fittings), trade suggestion and urgency; room-by-room comparison of move-in and move-out photos for deposit decisions.
• Heavy damage and insurance claims: vehicles and large property claims are handled in VIS-023.
• Private use: a tenant or landlord documenting a flat with timestamped photos; typically through an inspection app.

How it works

1. guided capture (app/link: required angles, label shot, delivery note or room ID)
2. quality checks (blur, light, duplicates, missing views) and metadata (time, ID)
3. read label / delivery note / return slip (OCR)
4. link to order, delivery or lease
5. condition and damage description (VLM + optional trained classifier) with confidence
6. rules: grade, cause hypothesis, deadline check (notice periods), value thresholds
7. draft next step: credit or restocking grade | carrier claim | trade order | question to customer
8. human decision on liability, credit and building damage; evidence pack archived

• Description and sorting: a multimodal model describes what is visible (crushed corner, wet stain, missing part) against a fixed category list; a small trained classifier helps where classes are recurring and the volume is high.
• Linking: OCR on the label, consignment note or return slip finds the order or lease; ambiguity goes to a person.
• Rules instead of guesses: credit amounts and deadlines come from price lists, contract terms and legal periods, not from the model.
• Human-in-the-loop: the model proposes a class and wording; liability, goodwill, deposit deduction and every claim letter are released by a person.

Data preparation

Which data: photos with a reference (return ID, delivery note number, address or room), the matching document and the final decision (credited, scrapped, claim paid, trade sent). The decision is the label.

Where it lives: phone galleries and messenger chats, e-mail attachments, the handheld scanner or driver app, the ERP/warehouse system (e.g. weclapp, Business Central, SAP B1, JTL), the TMS, and the property-management software; often just a shared folder with file names such as IMG_4821.jpg.

Typical gaps: photos without reference, no fixed angles, label not visible, free-text damage notes, no recorded outcome, photos of people, plates or flats' private belongings.

Preparation steps: 1. Define categories before any classification: condition classes, damage types, causes (PREP-019). 2. Introduce capture at source: a guided photo flow with mandatory reference ID (PREP-008). 3. Harvest labels from past credit notes, claims and work orders (PREP-020). 4. Build a test set of 100-200 past cases reviewed by an experienced clerk (PREP-018). 5. Remove or blur faces, plates, name plates before sharing images with a vendor (PREP-022, VIS-030). 6. Restrict access to tenant photos (PREP-023).

Effort: roughly 30 percent of the project, mostly category definition and photo discipline. Without a fixed capture routine the model has little to work with.

Implementation paths

Which path for whom

• SMB — Path B: there is no mature off-the-shelf product for the whole SMB workflow (returns, transit and tenant damage); a small build on cloud vision APIs with the existing ERP or property software is the realistic route.
• Developer — Path B: frontier VLM APIs on EU regions with structured output allow quick prototypes for integrators; self-hosting pays only at high volume.

Path A — Buy

• Inspection and handover apps: digital handover apps for flats (photo, room, signature) and condition-report tools from the vehicle and equipment sector are productised, but mostly capture and document photos; automatic damage grading is limited or an add-on.
• Claims and returns platforms: Tractable-style claims photo AI targets insurers and vehicles; returns platforms for retail target consumers (see VIS-037). For German wholesalers, haulers and property managers there is little ready-made.
• Check: EU hosting, link to your ERP/TMS, whether grading is trainable on your categories, price per case.
• Choose A when: a handover app or returns tool covers capture and you only want a sorting suggestion on top.

Path B — Build on cloud APIs (recommended)

• Models: multimodal APIs (Claude, Gemini, GPT, Mistral) in EU regions with a JSON schema for class, damage type, visible label text and a confidence note; optional trained classifier in a cloud vision service.
• Glue: phone web form or Teams/WhatsApp-free capture link, n8n or Power Automate, connection to ERP/WWS via REST or CSV, e-mail drafts into M365.
• Choose B when: you have a defined case type, a few hundred past cases for testing and someone who owns the process.

Path C — Self-host (open source)

• Open VLMs (Qwen-VL family, InternVL) via vLLM or Ollama on HW2; YOLO-class or RT-DETR detectors for fixed damage types; PaddleOCR for labels. Sensible if tenant photos must not leave the company or volume is high.

Path D — Custom training

Only for large, recurring product ranges (for example a wholesaler with tens of thousands of returns and a fixed grading scheme). Needs thousands of labelled photos and an owner for the grading rules.

State of the art (2026)

• Works today: describing visible damage, reading labels, checking photo completeness, sorting into a short list of classes, drafting texts. Good enough for triage with human check.
• Does not work reliably: judging cause (transport versus handling versus pre-existing), hidden damage, moisture and mould extent from a few photos, old versus new damage, functional faults of returned technical goods, cost estimates without a catalogue.
• Maturity: early-adopter. Vehicle and insurance damage photo AI is mature (VIS-023); the same technology for pallets, returns and tenant damage is used in pilots and niche products, with little published evidence. Treat vendor numbers as unverified.
• Limits: liability and building damage decisions need a human; a model will describe a stain confidently without knowing what caused it.

Shared resources & combinations

• Same stack (vlm + ocr-docparse): VIS-018 (labels and serial numbers), DOC-006 (delivery notes and consignment notes), DOC-003 (forms and claim documents).
• Neighbours: VIS-023 (vehicle and insurance claims), VIS-037 (consumer re-commerce grading), VIS-039 (field-work verification by photo), VIS-001 (defect detection on production lines).
• Upstream/downstream: VIS-030 for blurring people and plates before photos leave the company; the claim and credit drafts feed your ERP and e-mail workflow. A tracker for deadlines and certificates and a rebate and settlement reconciliation agent are candidates in the gap list but not yet in the catalogue.
• One capture flow, three industries: wholesale, haulage and property management can share the same phone-capture form and VLM service; only categories and rules differ.

Effort, cost & KPIs

• Path A: weeks; per-user or per-case subscription for an inspection app.
• Path B: 2-4 months for one case type including capture flow and ERP link.
• Path C/D: 4-9 months; GPU server and labelling.
• Cost drivers: category definition, capture discipline, integration, labelling.
• KPIs: minutes per case, share of cases with complete photo set, agreement with expert grade, share of claims filed within the notice period, credit-note dispute rate, deposit disputes lost, re-request rate for photos.

People & change

Involve early: management and the heads of returns, logistics or property management; drivers, warehouse and caretaker staff as end users; IT for the ERP link; the DPO because tenant photos show homes and belongings; legal for liability wording. Customers and carriers are affected by the result.

Works council: depends. If the app records who photographs what, how fast and how often the AI proposal is overruled, §87(1) no. 6 BetrVG applies; §90(1) no. 3 requires information about AI-supported work (GUIDE-004). Drivers' phones and handhelds are a typical trigger.

Data protection: tenant photos can show private belongings; use data minimisation and an AVV with each vendor (GUIDE-005).

AI literacy: Art. 4 AI Act (softened by the 2026 AI Omnibus to a duty to take measures; verify the current text): train staff on typical errors (shadows, reflections, wetness) (GUIDE-006).

Pilot and concerns: pilot one case type, for example B2B returns in one product group, in shadow mode for 4 weeks (GUIDE-009). Staff fear surveillance and blame; communicate that photos protect them in claims (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: minimal risk for object and building condition triage. Re-assess if outputs are used to evaluate individuals (for example scoring tenants or drivers) or if biometric features creep in. AI literacy (Art. 4) applies; the 2026 AI Omnibus reportedly softened it to an effort-based duty (check the current wording).
• GDPR: photos of flats, belongings, people, plates and EXIF locations are personal data; data minimisation, access limits, retention aligned with claim and limitation periods, EU hosting preferred. Art. 22 GDPR: no solely automated deduction from a tenant's deposit.
• Transport law: under §425 HGB the carrier is liable for loss and damage between taking over and delivery; §438 HGB requires notice of visible damage at delivery and of hidden damage in text form (e.g. e-mail) within 7 days after delivery, sending in time is enough, otherwise the goods are presumed delivered undamaged. A photo triage helps keep these deadlines; ADSp 2017 and the consignment terms add procedures and limits. Specific wording matters: note concrete damage on the consignment note, not only a general reservation.
• Commercial law: §377 HGB: merchants must inspect goods and report defects without delay, otherwise goods count as approved. Photo documentation at goods receipt supports both sides in B2B returns.
• Tenancy law: §548 BGB: landlord's claims for damage expire six months after getting the flat back, so triage and decision have to happen quickly; the burden of proof for damage is on the landlord, which makes dated photos and a signed handover protocol valuable. Deposit rules are in §551 BGB.
• Risks and mitigations: wrong cause attribution (human decision, show annotated evidence), disputes about photo authenticity (capture in app, keep metadata, hash), over-trust in confident text (confidence and checklist), staff pushback (transparent purpose).

Getting started

1. Pick one case type and measure the baseline: cases per month, minutes per case, disputes.
2. Fix the category list and the minimum photo set (angles, label, reference ID).
3. Collect 100-200 past cases with outcome as a test set.
4. Prototype with a cloud VLM and structured output; run in shadow mode next to the clerk for 4 weeks.
5. Compare class and next-step proposal; add deadline checks; then connect to ERP/e-mail drafts.

Sources

1. § 425 HGB — Haftung für Güter- und Verspätungsschäden — https://dejure.org/gesetze/HGB/425.html
2. § 438 HGB — Schadensanzeige — https://dejure.org/gesetze/HGB/438.html
3. Haufe — Die Rügeobliegenheit nach § 377 HGB in der Praxis — https://www.haufe.de/recht/weitere-rechtsgebiete/wirtschaftsrecht/die-ruegeobliegenheit-nach-377-hgb-in-der-praxis_210_679680.html
4. § 548 BGB — Verjährung der Ersatzansprüche und des Wegnahmerechts — https://dejure.org/gesetze/BGB/548.html
5. Pallet-Terminal — Europaletten Schäden dokumentieren — https://www.pallet-terminal.com/blog/europaletten-schaeden-dokumentieren
6. immocloud — Wohnungsübergabeprotokoll 2026 — https://www.immocloud.de/magazin/wohnungsuebergabe-protokoll/
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj