In an industry where “zero harm” is more than a slogan, turning safety ambition into measurable outcomes remains a persistent challenge. At a recent viso.ai webinar, Horizontal Drilling International (HDI) pulled back the curtain on a live deployment of AI Vision across a number of its construction sites. HDI is a horizontal directional drilling specialist within VINCI Construction, established in 1984, with more than two million metres of pipeline installed across 40 countries.
Here are the lessons, outcomes, and implications for HSE, lean, and efficiency leaders in heavy civil engineering.
Why Horizontal Directional Drilling Raises the Safety Stakes
Construction remains among the highest risk industries for serious injury and fatality, and trenchless work concentrates that risk into a small footprint. A horizontal directional drilling spread installs pipelines and cables beneath rivers, motorways, railways, and shorelines, which means rigs of up to 400 tonnes, suspended pipe strings, pressurised drilling fluid, and tracked plant all operating within metres of one another.
The geometry of risk also shifts constantly. Exclusion zones move as the bore advances, the rig repositions, and the pipe string is welded and pulled. Unlike a fixed production line, there is no permanent layout to guard, so computer vision in construction has to adapt to conditions that change between shifts.
This is precisely where AI-driven construction safety is proving most useful, shifting teams from reactive compliance towards proactive prevention. But success is not really about the technology. It depends on making computer vision accessible, reliable, and trusted in unpredictable, real-world environments.
HSE and lean leaders will recognise the familiar tensions: reducing waste, changing behaviour measurably, and rolling out something new across dozens of sites without it becoming another siloed trial. HDI treated AI Vision as a compact safety tool embedded in existing workflows rather than a heavyweight R&D experiment.
Idea to AI vision app in seconds.
HDI’s Starting Point
When HDI first engaged with AI Vision, its safety systems were heavily manual: observations, toolbox talks, audits, incident reports, and reactive follow-up. The company was already certified to ISO 9001, ISO 14001, and OHSAS, so the gap was not process maturity. It was coverage.
Today HDI runs deployments on multiple live sites with five active use cases, using only four cameras and one edge device per site. The motivation was simple but deep: reduce blind spots, make safety visible, and prevent incidents rather than merely record them after the fact.
As their HSE sponsor put it:
“With heavy machinery, moving loads, and shifting site layouts, manual observation just doesn’t cover enough. AI Vision gives us real-time visibility, site-specific adaptation, and the ability to act before things go wrong.”
Five Computer Vision Use Cases on Live Drilling Sites
Four cameras per site feed a single on-site AI device with no cloud dependency. From that lean footprint, HDI runs five core vision applications:
- Person under load detection. Alerts when someone walks beneath a lifted load, such as an excavator bucket or a suspended pipe section.
- Man-machine proximity monitoring. Focused on heavy machinery blind spots, which is critical for preventing collisions and pinch point injuries.
- Housekeeping detection. Hazards like stray construction bags look inconsequential, yet they repeatedly cause blocked routes, trip risks, and lost time.
- PPE compliance at entry points. Gear detection for hard hats, vests, and other equipment as workers cross into active zones.
- Restricted and exclusion zone detection. Confirming that nobody stands in a crane swing radius or inside the drill’s most restricted limits while it is operating.
These systems feed alerts into HDI’s central incident centre, where HSE leads review them. The system is not passive: it actively shapes supervision, escalation, and behaviour on the ground.
Each detection event also creates a verified data point. Analysed over time, those points reveal recurring risks, non-compliance patterns, and hotspot areas across sites. Supervisors use the insight to adjust workflows, retrain crews, and reconfigure layouts so the same incident does not happen twice.
One standout technical decision was to run all inference locally. Edge computing removed cloud dependency entirely, which eliminated latency, reliance on unstable site networks, and burdensome data transfer.
Outcomes, ROI, and Behaviour Change
HDI already operated robust safety policies and a strong incident record, so the value was never about fixing a broken system. Adopting AI Vision let the company standardise safety practice, automate parts of its observation workload, and free teams for higher value activity.
Amplifying existing best practices produced several benefits:
- Near misses are consistently caught before escalation, feeding structured near-miss detection data back into planning
- Supervisor response times improved because alerts route straight to accountable stakeholders
- Comparative performance data between sites became possible, enabling continuous benchmarking of safety effectiveness
- Behaviour shifted, and this matters most: HDI’s people now expect insight from the system rather than tolerating it
Commercially, HDI has moved to multi-year renewal and is expanding towards more than ten sites, which signals confidence in both the technology and the delivery model behind it. For teams building a business case, measurable safety ROI follows the same pattern: fewer disruptions, faster correction, and less time gathering evidence manually.
What Made the Deployment Work
Five factors separated this rollout from the pilots that stall.
| Success Factor | What It Looked Like at HDI |
| True partnership | Weekly alignment, openness to iteration, and executive sponsorship from site level to boardroom |
| Pragmatic technical design | A deliberately lean four-camera, one-device architecture built around limited bandwidth and unstable power |
| Iterative rollout | Stepwise scaling instead of a single large launch, with learning carried between sites |
| Adoption governance | Escalation plans, training, and site-level ownership so alerts actually drive change |
| Environmental realism | Models tuned for scaffolding, dust, fluctuating light, weather, and camera movement |
The last point deserves emphasis. Horizontal directional drilling sites are messy and dynamic, and even building movement or a power outage affects camera stability. Any solution has to be robust, resilient, and adaptable, or the alerts stop being credible.
From Detection to Agentic Computer Vision
HDI’s next step is to extend coverage to additional sites and to widen the use case set towards fall detection, housekeeping change monitoring, and scaffolding safety. Results are already being shared across the parent group, VINCI Construction, through demos and innovation days.
The larger direction of travel is a shift from detection to reasoning. Today’s deployment answers a fixed question set defined at configuration time. Agentic computer vision changes that relationship, because the system can be queried in natural language and reason across sequences of events rather than isolated frames. A supervisor could simply ask why one bore section produced clustered proximity alerts.
That capability matters here. The hazard profile of a bore changes as it progresses, so a system that adapts its own questions beats one needing reconfiguration at every stage. Regulators are moving in step: the UK’s Health and Safety Executive has confirmed that AI sits within existing health and safety law, with risk assessment, transparency, and human oversight remaining foundational.
Lessons for HSE, Lean, and Efficiency Leaders
If you are evaluating AI Vision for a drilling, pipeline, or heavy civil environment, HDI’s experience points to a clear starting position:
- Begin with one or two use cases across up to three sites, then test robustness, workflows, and adoption before scaling
- Design for privacy by design and align with SOC 2, ISO 27001, and GDPR from day one
- Co-design with frontline crews, because buy-in and usability decide whether alerts get acted on
- Assume unreliable connectivity and avoid architectures that depend on it
- Measure behaviour change, not alert counts
- Anticipate scaling constraints across device capacity, camera counts, power, and environmental noise
Treat AI Vision as a lean safety tool rather than another tech project: compact, value-driven, scalable, and integrated into how crews already work.