What problem does it solve?
Operators that build and maintain networks through contractors cannot see most of the work. A technician connects a cabinet, a distribution point or a modem, takes a few photos and closes the job. Someone in a back office may look at a sample later. A defect that slips through shows up as a repeat visit, a degraded fiber asset, a penalty from a wholesale carrier or a customer who leaves.
Cell towers have the same problem in another form. They reach up to 75 meters and need periodic checks for corrosion and loose bolts, and the check used to mean climbing the tower or studying many images by eye.
Regulation can add volume. In France, Orange France says ARCEP made photos before and after each fiber intervention mandatory in summer 2021, which means analyzing about 20,000 photos a day. Sampling is not an option when every intervention has to be verified, and a person cannot read that many images.
- IQGeo's case study says that before AI photo checks, Bouygues Telecom reviewed only 5 to 10% of fiber connections, through manual back office audits that ran after the installation.Bouygues Telecom case study, NetLux AI (2026)
- SK Telecom said in 2023 that safety inspectors previously had to study around 100 drone images to complete the inspection of one cell tower.SKT Innovates Cell Tower Safety Inspection with AI (2023)
How does it work?
- Capture. The technician photographs the work with the field app: cabinet, distribution point, splice protection, modem or ONT, a connection cable or a nameplate. For towers, a drone flies a set route and takes images of the structure and fasteners.
- Analyze. An image recognition model compares each photo with the operator's standards for that job type. It flags anomalies such as unconnected fibers, wrong color coding, cable paths that break the rules, missing hardware or loose bolts, and can read text on labels to match it against the job record. Orange France set a target of three seconds so the check runs as the technician is about to leave.
- Feed back on site. The app marks the problem on the image. The technician fixes it and uploads a new photo, or disagrees with the machine and continues. At Orange France the technician can ignore a recommendation.
- Record. Photos, verdicts and timestamps become the job record, which supports contractor billing checks, wholesale carrier disputes and a searchable history of the asset.
- Learn. At Orange France, 50 volunteer technicians tested the first version and their feedback was used to improve the models. Plan the same loop for your own corrections.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Mobile app, Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Accuracy | Too few to pool | 98% | 1 | 1 vendor |
| Cost reduction | Too few to pool | 20% | 1 | 1 vendor |
| Cycle time reduction | Too few to pool | 95% | 1 | 1 organization |
Value drivers: Lower cost to serve, Compliance quality, Speed and cycle time, Customer experience.
Indicative value
A fiber operator whose contractors complete 500,000 installations a year
USD 480,000 to USD 5.1 million
Repeat visit and penalty cost avoided per year
How this is calculated
Formula: installs * defectShare * caughtShare * costPerRepeat. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Contractor installations per year installs, installations per year | 500,000 | 500,000 | The reference operator. |
| Share of installations with a defect that would cause a repeat visit or a penalty if unchecked defectShare, fraction of installations | 0.03 | 0.08 | Editorial assumption, replace with your own repeat visit and rejection rates. The evidence on this page does not publish a defect rate. |
| Share of those defects found on site and fixed before the technician leaves caughtShare, fraction of defects | 0.4 | 0.8 | Editorial assumption. Bouygues Telecom's vendor reports 98% accuracy on visible defects, but not every defect is visible in a photo and not every flag leads to a fix. |
| Cost of a repeat visit or penalty for one defective installation costPerRepeat, USD per defect | 80 | 160 | Editorial assumption, replace with your own cost per visit. |
What it leaves out: Counts repeat visit and penalty cost only. It leaves out billing checks on contractors (IQGeo reports a 20% cut in fiber unit installation costs at Vodafone Spain, a vendor figure with no baseline period), churn avoided, the review time saved, and the cost of the app, labeling and drone work.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
SK Telecom
South Korea · Telecommunications · 2023
SK Telecom built an image analysis model in house that reads drone photos of cell towers, which reach up to 75 meters, and judges the condition of bolts and nuts. Before, inspectors had to climb the towers or study about 100 images per tower by eye. SKT says the model cut the time for that review by 95% and made the analysis more consistent, and plans to add wind pressure and inclination checks and a link to its safety management system.
- Cycle time reduction: 95%
"With the adoption of this AI analysis model, SKT was able to reduce the time required for the process by 95%, while increasing the reliability and consistency of the analysis results."
Claimed by: organization
CityFibre
United Kingdom · Telecommunications · 2022
In February 2022 CityFibre announced a national rollout of Deepomatic's image recognition across its build, upgrade and maintenance work, to be complete by the end of 2022. Engineers upload photos of their work before they close a job, and the images are analyzed in real time to either validate the work or point out issues to fix before leaving the site. CityFibre says this supports its Right First Time installation rates and builds a searchable visual record of the network. The release reports no measured results.
No outcome disclosed.
Bouygues Telecom
France · Telecommunications · 2026
Bouygues Telecom relies on contractors for its fiber rollout and used to check only a small share of installations through manual back office audits. After a proof of concept, technicians now photograph the cabinet, distribution point and modem or ONT with a smartphone, and an image recognition model flags problems such as unconnected fibers, wrong color coding and cable paths that break the rules. IQGeo's case study says the solution is already analyzing more than 3 million installations and reaches 98% accuracy on visible defects. Both figures are IQGeo's own claims.
- Accuracy: 98%
"With 98% accuracy, the AI consistently detects visible defects, setting a new benchmark for quality assurance."
Claimed by: vendor - Interactions handled: at least 3 million, installations the solution is already analyzing, per IQGeo
"Already analyzing more than 3 million installations, the solution evolves alongside Bouygues’ needs, supporting new equipment, tasks, and quality standards."
Claimed by: vendor
Vodafone Spain
Spain · Telecommunications · 2026
Vodafone Spain used to let contractor technicians type the operator's name into the field service app without proof, so it could not confirm that the work matched what was invoiced. After a four week field proof of concept, it rolled out NetLux AI to all contractors: the technician photographs the connection cable during a fiber installation and the system reads the operator's name from the image and cross checks it against the field service data. IQGeo's case study reports a 20% cut in fiber unit installation costs and a billing process Vodafone Spain can verify.
- Cost reduction: 20%
"Reduced fiber unit installation costs by 20%."
Claimed by: vendor
Orange France
France · Telecommunications · 2023
Since summer 2021 the French regulator ARCEP has required photos before and after each fiber intervention, which Orange France says means analyzing about 20,000 photos a day. Orange and the consultancy Artefact built a computer vision model inside the technician's app that checks, close to real time, whether the work on network equipment is conforming and says which area is not. The technician can overrule it. The first version took nine months, was tested by 50 volunteer technicians and is now used by the 10,000 technicians deployed every day in France.
- Users served: 10,000, technicians using the tool every day in France
"It was a real technical and human feat, as the tool has now been adopted by the 10,000 technicians deployed every day in France."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Photos from past jobs with the later outcome, such as repeat visits, rejected builds and penalties
- Written installation standards for each equipment type, turned into checkable rules
- Job records with the operator name, location and work order, to match against the photo
- For towers, drone image sets with engineer notes on each defect found
Systems to integrate
- Field service management and work order systems
- Technician mobile app, with offline capture
- Asset and geographic information systems that hold the network records
- Contractor billing and quality reporting
- Drone flight planning and image storage for tower work
Complexity: Medium
The build needs a labeled library of correct and defective photos for each equipment type (Orange France had 80,000 photos labeled in three months), a field app that returns a verdict in seconds and works offline, and a clear rule set from network engineering. Towers add drone flights, flight permits and an engineer who signs off on structural findings.
- 1
Pick the defects that cost money
Start from repeat visits, wholesale penalties and failed activations, and list the visual causes. Choose two or three checks that a photo can settle, such as splice protection or cable color coding, and leave the rest for later.
- 2
Run a proof of concept on live jobs
Vodafone Spain ran four weeks in the field and Bouygues Telecom ran a proof of concept before rollout. Compare the model's verdicts with an auditor's on the same photos and measure missed defects and false alarms separately.
- 3
Put the verdict inside the technician's workflow
Photo taking should already be part of the job. Return the verdict in seconds, show where the issue is on the image and let the technician continue after a disagreement, with the disagreement logged.
- 4
Close the loop with contractors
Share defect rates by contractor and equipment type for coaching and billing checks, with human review before any penalty, and feed model errors back as labeled examples.
- 5
Extend to new job types and sites
Add equipment types, installs at the customer's home and tower inspection one by one. SK Telecom plans to add wind pressure and inclination checks to its drone inspection later.
Guardrails
- A flagged defect is advice to the technician, who can overrule it; the override is logged for review
- No penalty, ranking or disciplinary action based on model output alone
- Photos of customer premises are limited to what the check needs, with retention and access rules
- Structural findings on towers are confirmed by a qualified engineer before any repair decision
KPIs to instrument
- Share of jobs with a photo check, against the old audit sample rate
- Missed defect rate and false alarm rate against an auditor's review of a sample
- Repeat visits, rejected builds and wholesale penalties per thousand installations
- Time from photo upload to verdict, and technician override rate
Human in the loop
Technicians own the work and can disagree with a flag. Quality engineers review disagreements and a sample of passed jobs each week. Engineers sign off on tower findings, and contractor penalties go through a person.
Common failure modes
- Lighting and angle confuse the model
- Orange France says the conditions in which photos are taken can confuse its own orange cables with competitors' red ones, and it reduced the model's performance so it would not contradict what the technician sees. Test in rain, dark cabinets and cramped cabinets.
- Technicians read it as surveillance
- If the tool is seen as a way to control people, photos get gamed or skipped. Involve technicians from the start, which Orange France says it did, and have volunteers test the first working version, as its 50 volunteer technicians did.
- Photos pass but the work is wrong
- A model that reads one photo can miss a fault that shows only in a test or an inside view. Keep the check next to service activation tests.
- Model drifts as equipment changes
- New cabinets, labels or contractors change what photos look like. Keep a labeling process running after launch.
What are the risks and rules?
EU AI Act
Depends on design
Checking photos of network equipment is normally minimal risk. If the output is used to allocate tasks based on individual behavior or personal traits, or to monitor or evaluate individual technicians or contractor workers, Annex III point 4(b) treats the system as high risk, and under Article 26(7) employers must inform workers' representatives and the affected workers before using it. Keep the model on the work, not the person.
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 4 covers employment and worker management, including task allocation based on individual behavior and monitoring of performance.
- Article 26, obligations of deployers of high risk AI systems (European Union, Europe). Paragraph 7 requires employers to inform workers' representatives and the affected workers before putting a high risk AI system into use at the workplace.
- General Data Protection Regulation (European Union, Europe). Photos of customer premises and of technicians can be personal data, so purpose, minimization and retention rules apply.
Controls to put in place
- Purpose limits for photo data, agreed with employee representatives where required
- Data protection impact assessment for photos taken inside customer premises
- Documented comparison of model verdicts with auditors before any contractor consequence
- Version control and retraining routine for the image models
Frequently asked questions
- How accurate is AI at checking installation photos?
IQGeo says its model reaches 98% accuracy on visible defects at Bouygues Telecom, and that the solution is already analyzing more than 3 million installations. That is a vendor claim and does not say how accuracy was measured. Run your own proof of concept and compare the verdicts with an auditor's on the same photos before you rely on it.
- Why check every installation instead of a sample?
Bouygues Telecom reviewed only 5 to 10% of fiber connections by hand before it adopted AI, and Orange France says French rules require photos for each intervention, about 20,000 photos a day. A model makes full coverage affordable, so the remaining human effort goes to disagreements and structural decisions.
- Does AI replace tower climbers?
It reduces climbing. SK Telecom says its drone and image model lets it minimize the need to go up the towers and cut the review time per tower by 95%, and it describes the system as a complement to existing visual inspection. Engineers still decide what needs repair.
How to cite this page
Blits.ai AI Use Case Library, "AI visual inspection of telecom field work and cell sites", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/telecom-field-work-visual-inspection. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published