New: Networks — CCTV surveys, condition grading and renovation programmes for water and sewer networks, in the same service. Explore Networks →
What Pipelytica changes in inspection work. Measured figures are marked as measurements. Other figures are examples, and your own result depends on your videos and workflow.
0.7 ×
The inspector’s time goes into confirming the findings the AI proposes. The report is built from the confirmed findings.
2nd pair of eyes
In our measurement (9/2026, 34 videos the model never saw) the AI produced a finding at 69% of the defect spots an inspector had marked, and named the right defect family on 60% of its findings. The AI can be wrong, so the inspector confirms every finding before the report.
Your numbers
The saving depends on your site volume and labour cost. Estimate your own with the ROI calculator.
EN 13508-2
A PDF report in one click. The inspector confirms every code before the report.
Consistent
The AI’s proposals, over 20 architecture checks and report templates give a consistent starting point the inspector finishes.
Capacity
How many more sites the same team can handle depends on your workflow. Estimate it with the ROI calculator.
We do not publish validation scores from our own training data. We measure the finished product on videos the model has never seen, against the labels of an inspector.
September 2026 · 34 field videos · labelled by an EN 13508-2 inspector
| Measure | Result |
|---|---|
| Right defect family in the customer report The name of the finding (e.g. corrosion or sediment) matches the label the inspector gave. 95% confidence interval 51–69%. | 60% |
| Exact defect type right The same, with deposit sub-types (sand, scale, sludge) told apart. Inspectors often disagree on these too. | 39% |
| Marked defect spots found Share of the defect spots the inspector marked where the AI produced a finding. The rest are left for the inspector to find. | 69% |
| Analysis time relative to video length Measured on the production server: a 67-second video analysed in 45 seconds. | 0.7 × |
34 field videos were kept entirely out of model training. An EN 13508-2 inspector marked the defect spots, and every AI finding was compared with those marks. The confidence interval is computed per video.
In published studies even experienced inspectors miss about a quarter of defects and code the same defect differently (Dirksen et al. 2013). The AI is a second pair of eyes, not a judge.
AI can make mistakes – Always verify findings before delivering a report to a customer.
| Task | Without Pipelytica | With Pipelytica |
|---|---|---|
| Documentation/site | 4h | 45min |
| Reporting/site | 3h | 15min |
| Work hours/year | 3500h | 500h |
| Labor costs | €175 000 | €25 000 |
Example calculation: ~€150,000 savings/year
+ ~3,000 work hours freed up (a model calculation based on assumptions, not a measurement)
Try it with your own videos free for 14 days.