Photo-based field work verification (installation QA at point of capture)

Network, energy and building contractors prove completed work with photos, but humans review only a small sample days later, so rework means a second truck roll. A vision model checks every photo on the phone against rules for that job type (right equipment, labels readable, cable routing, closed enclosure, not a reused image) and tells the technician to fix or retake before leaving site. The payoff is first-time-right work, fewer disputes with subcontractors and cleaner as-built records.

Problem & value

• Today: technicians upload photos to a field-service app; back-office QA staff or the operator's own inspectors sample a fraction, often days after the crew has left. Errors found late need a return visit; subcontractor invoices are contested; GIS records drift from reality.
• AI value: 100% of jobs checked within seconds, feedback while still on site, consistent rules across subcontractors, detection of reused, stale or off-location photos, and an evidence trail that can release payment automatically.
• Evidence (vendor/trade press, not independently audited): Swisscom is reported to use Deepomatic to verify work by field technicians across four subcontractors with payments processed automatically once work is verified; CityFibre engineers upload images before closing a job and have them analysed in real time to confirm completion or flag issues. Treat all figures as vendor claims [1][2][3].

Variants

• Fibre/telecom rollout: splice closures, ONT and drop installation, cabinet and duct work, pole and chamber photos; checks tied to operator build standards.
• Energy and utilities: smart-meter installation, solar and heat-pump commissioning, substation maintenance steps, gas and water connections (compare VIS-003 for drone-based asset inspection).
• Construction and facility services: fire-stopping, waterproofing, cable trays and handover evidence before concealment (see VIS-040 for scan-based as-built capture).
• Subcontractor payment gating: photo verdict is the trigger for milestone payment.
• Private/household: homeowner or installer photo proof for subsidy claims (heat pump, PV); lighter checks.

How it works

1. technician takes photo in field app (guided capture: required shots per job type)
2. on-device quality gate (blur, darkness, framing) + metadata check (GPS, time, device)
3. classification: correct photo type for this step?
4. detection/VLM rule checks per job type (component present, labelled, seated, closed, tidy)
5. OCR of labels, serial numbers, port IDs (see VIS-018) matched to work order
6. duplicate / reuse / screen-recapture detection (perceptual hash, embeddings, EXIF)
7. verdict + reasons shown on phone; retake or fix before closing the job
8. results to field-service / GIS / payment system; low-confidence cases to QA reviewer

• Rule-specific detectors: small object detectors and classifiers are trained per rule; a VLM handles long-tail rules ("is the tube sealed and cable bend radius acceptable?") via prompts with reference images, with detector results as the audit-grade signal.
• On-device inference lets technicians get feedback without connectivity; cloud re-checks later and adds fraud screening.
• Human-in-the-loop: QA reviewers handle uncertain or failed jobs, disputes and rule maintenance; safety-critical or high-value sign-off stays with a qualified person.

Data preparation

• Data needed: work orders and job types (field-service/GIS system), the photos technicians already take (field app, phone galleries, mail to the back office), checklists and installation standards (PDF, intranet), and past QA audit results with pass/fail reasons.
• Typical gaps: rules exist only as long PDFs or as tacit knowledge of the QA team, no reference photos for "correct" and "almost correct", photos without reliable geotag or job link, mixed quality from subcontractors, no recorded reasons for rejections, and technician IDs in the photo metadata.
• Preparation steps: (1) turn standards into short, checkable rules per job type with reference photos (PREP-016); (2) export work orders and photo links from the field-service system (PREP-005); (3) harvest past QA decisions as labels (PREP-020) and add hard negatives; (4) assemble a test set of a few hundred photos per rule with QA reviewers (PREP-018); (5) strip or pseudonymise technician IDs and bystander faces in training data (PREP-022) and check purpose and legal basis (PREP-025).
• Tools: Label Studio or CVAT for labelling, field-app export APIs, perceptual-hash scripts for duplicate checks.
• Share of effort: roughly 30 percent; rule definition and reference examples matter more than volume. For Path A, mostly rule configuration.

Implementation paths

Which path for whom

• Enterprise: B - rules per job type, field app and payment release are core processes with high photo volume; own rule models on cloud APIs (with specialist platforms as accelerator) avoid lock-in and per-photo fees.
• SMB: A - a mature EU-hosted specialist platform or an add-on to the existing field-service suite delivers live rules in weeks; building own rule models rarely pays for a small installer.
• Developer: C - open detectors, embeddings and VLMs with on-phone export give control over rules, cost and data location.

Path A — Buy (recommended)

• Specialist platforms: Deepomatic (France, EU; acquired by IQGeo and now marketed as NetLux AI, formerly Deepomatic Lens) is the reference for telecom field verification and integrates with IQGeo GIS and field apps; DAC.digital has published a fibre-installation validation case [4][6].
• Adjacent products: field-service suites (e.g. SAP Field Service Management, Salesforce Field Service, ServiceNow FSM, IFS, Dynamics 365 Field Service) cover scheduling and photo capture; AI checks are mostly add-ons or partner modules. Check EU hosting, on-device mode, rule authoring by non-developers, and API access to verdicts.
• Choose A when: a recognised job type (fibre, meters) is the focus and rules must be live fast.

Path B — Build on cloud APIs

• Multimodal models: Gemini, Claude, GPT or Mistral (EU provider) vision models with structured output (JSON verdict per rule) and reference images; Azure AI Custom Vision/Azure AI Vision, Google Vertex AI, AWS Rekognition Custom Labels for trained detectors, using EU regions.
• Integration: event-driven function triggered on photo upload; verdicts written back to the field app via REST; n8n or Power Automate for routing exceptions.
• Choose B when: you own the field app and need custom rules, with moderate photo volumes.

Path C — Self-host (open source)

• Models: Ultralytics YOLO (AGPL, check licence) or RT-DETR/RF-DETR for detection; DINOv2/SigLIP embeddings for duplicate and similarity checks; Qwen-VL family or Gemma/PaliGemma-class open VLMs for rule questions; PaddleOCR for labels. Serve with vLLM or Triton; export to ONNX/TFLite/Core ML for on-phone inference.
• Hardware: HW2 handles a mid-sized operator's photo stream; HW1 phones/edge for the gate step.
• Effort: a small ML team plus MLOps for versioned rule models.

Path D — Custom training

Worth it for high-volume job types: train detectors per rule on a few hundred to a few thousand labelled photos per rule, including hard negatives (almost-correct work). Reviewer decisions and technician retakes are natural training data. Re-train when hardware or standards change.

State of the art (2026)

• VLMs now answer free-form checklist questions on phone photos zero-shot, shortening rule authoring from weeks to days, but are less reliable on fine spatial details (bend radius, torque marks, small labels) than trained detectors; hybrid designs dominate.
• Small on-device models give instant feedback offline; heavier checks run in the cloud.
• Provenance checks (EXIF/GPS, perceptual hashing, embedding similarity) catch reused or recaptured photos; C2PA content credentials on cameras are an emerging authenticity signal.
• Limits: lighting, dirt, odd angles and occlusion; generalising to new equipment; photos prove appearance, not function (signal level, torque, leak-tightness still need measurement); technicians can stage a correct-looking photo; rule drift when standards change.

Shared resources & combinations

• Same stack as VIS-002 (assembly verification), VIS-023 (damage photos), VIS-037 (condition grading): detectors, VLM, mobile capture, label OCR (VIS-018).
• Anonymisation of faces and plates in site photos: VIS-030.
• Fraud/pattern analytics across subcontractors: ANOM-003-style anomaly scoring on verdicts.
• Scan-based as-built models: VIS-040; drone checks: VIS-003; form data: DOC-003.

Effort, cost & KPIs

• Path A: weeks to pilot; pricing usually per job or per photo plus platform fee.
• Path B: 2-4 months; main cost is rule definition, labelled examples and app integration.
• Path C/D: 4-9 months incl. data collection and MLOps.
• KPIs: share of jobs verified (target 100%), first-time-right rate, repeat-visit/truck-roll rate, QA reviewer hours per job, dispute rate and payment cycle time, false pass / false fail rates per rule, retake rate, duplicate-photo detections.

People & change

• Involve early: management and the operations or network-build lead (process owner), QA reviewers, technicians and subcontractor managers (end users and suppliers), IT for the field app, the DPO, and the works council from day one. See GUIDE-001, GUIDE-009 and GUIDE-010 (vendor choice).
• Works council: yes. The system processes photos tied to technician IDs, time and location and logs pass/fail per person, so it is objectively suited to monitor performance: §87(1) no. 6 BetrVG applies (BAG), plus information duties under §90(1) no. 3. Conclude a works agreement that limits the system to work products, excludes individual performance rankings and sets retention limits; consider an expert under §80(3) (GUIDE-004).
• Data protection: DPIA likely (systematic evaluation of workers, geodata, bystanders), AVV with the vendor, blur third parties and respect customers' premises (GUIDE-005).
• AI literacy: train technicians on guided capture and how to retake or contest a verdict; train reviewers on false-pass and false-fail risks (Art. 4 AI Act; GUIDE-006).
• Pilot and concerns: pilot with one job type and a volunteer team as champions. Typical fear is covert surveillance: communicate openly what is checked and what is not, show reasons on the phone, keep a human appeal path before payment is withheld, and let QA staff own rule changes (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: usually minimal risk (quality inspection of work). It becomes high-risk (Annex III, employment/worker management) if outputs drive decisions on individual technicians (task allocation, performance evaluation, termination), and must avoid Art. 5 practices such as emotion inference at work. Keep the system scoped to work products, not worker behaviour; AI literacy (Art. 4) applies [5].
• GDPR and labour law: photos show faces, house interiors, plates, addresses and location/time data of workers. Use DPIA, blur bystanders (VIS-030), minimise retention, and involve the works council (Germany: BetrVG section 87) before rolling out; avoid using verdicts as covert performance surveillance. Customers' premises need notice.
• Contract/commercial: automated payment release needs an appeal path for subcontractors.
• Mitigations: human review for disputed fails, versioned rules, audit logs, regular sampling of passes to measure false-pass rate.

Getting started

1. Pick one job type with costly rework and write its photo checklist with QA staff.
2. Collect 500-2,000 historic photos with QA outcomes; measure the current sample-audit error rate.
3. Pilot a Path A tool or a VLM prompt on those photos and compare to QA decisions per rule.
4. Add guided capture and on-site feedback for one subcontractor; log retakes.
5. Agree DPIA, works-council terms and an appeals process before linking verdicts to payment.

Sources

1. Connect Fibre chooses Deepomatic — https://www.connectfibre.co.uk/news/connect-fibre-chooses-deepomatic
2. CityFibre partners with Deepomatic — https://dcnnmagazine.com/infrastructure/cabling/cityfibre-partners-with-deepomatic/
3. Fibre Systems, How one AI platform is revolutionising telecom field operations — https://www.fibre-systems.com/article/revolutionising-field-operations-how-one-ai-platform-transforming-telecoms
4. IQGeo to acquire Deepomatic — https://www.fibre-systems.com/article/iqgeo-acquire-deepomatic-ai-fibre-network-digital-twins
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. DAC.digital, automated fibre-optic installation validation — https://dac.digital/case-studies/automated-fibre-optic-installation-validation-with-computer-vision-and-machine-learning/
7. Ultralytics YOLO — https://github.com/ultralytics/ultralytics