Ferrovial Deploys AI Agents to Automate Heavy Construction Machinery
Ai4
Deploying proprietary AI agents across zero-tolerance infrastructure assets allows builders to operate autonomous machinery fleets and preempt mechanical failures across global work sites.
## Background
Published on YouTube by Ai4 on August 22, 2026, this session features Jason Abbruzzese of NBC News interviewing Dimitris Bountolos, Chief Information and Innovation Officer at Ferrovial. The program explores how AI is transitioning from pilot experiments to frontline operations across the full infrastructure lifecycle, examining live data integration, safety, and operational efficiency within complex physical environments.
## TL;DR
Ferrovial demonstrates how AI has shifted from theoretical hype into heavy civil engineering, deploying edge computer vision on high-speed rail inspection trains and orchestrating autonomous heavy machinery swarms on airport runways. Rather than relying purely on off-the-shelf models, the company utilizes thousands of specialized agents fine-tuned on seven decades of proprietary operational data to detect 3D BIM design clashes and analyze massive telemetry streams from global tunnel boring machines. Operating under a strict zero-tolerance safety mandate, these deterministic workflows use geofencing and synthetic rehearsal environments to safeguard physical sites. Frontline adoption has surged as field superintendents experience immediate operational relief, evolving from cautious end users into passionate advocates for site-level AI integration.
## Atmosphere & style
The session has a grounded, pragmatic, and intellectually rigorous atmosphere that deliberately steers clear of Silicon Valley buzzwords. Abbruzzese anchors the conversation with incisive journalistic skepticism, while Bountolos responds with steady, authoritative poise rooted in heavy civil engineering and real-world assets. The exchange feels refreshingly tangible and forward-looking, replacing abstract AI rhetoric with vivid operational imagery of autonomous bulldozers, pitch-black tunnel simulations, and sensor-packed boring machines.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=gDONu82P5NM&t=4s) Ferrovial has integrated artificial intelligence into physical infrastructure for over a decade, deploying computer vision and edge inference to assets long before the generative AI boom.
- [00:04](https://www.youtube.com/watch?v=gDONu82P5NM&t=4s) Edge vision systems mounted on high-speed rail inspection trains reduce herbicide consumption by 80% by executing real-time selective spraying at operational speeds of 70–80 km/h.
- [00:04](https://www.youtube.com/watch?v=gDONu82P5NM&t=4s) Generative AI creates synthetic training data for extreme edge cases, allowing inspection models to function accurately in harsh environments like pitch-black tunnels and severe weather.
- [04:17](https://www.youtube.com/watch?v=gDONu82P5NM&t=257s) Because long-term infrastructure concessions span up to 40 years, predictive AI models are vital for forecasting demand, structural degradation, and upfront multi-billion-dollar capital costs.
- [04:17](https://www.youtube.com/watch?v=gDONu82P5NM&t=257s) Autonomous agents deployed at the edge turn remote, solar-powered utility assets into self-supervising systems, conducting real-time operational triage without on-site technicians.
- [07:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=426s) Autonomous heavy machinery swarms, such as excavators and bulldozers, can operate continuously 24/7 on construction sites within strict digital guardrails.
- [07:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=426s) Proprietary design-auditing agents parse complex Building Information Modeling (BIM) files to automatically detect 3D spatial clashes across designs submitted by separate engineering firms.
- [11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) While generic foundation models are commodities, sustainable competitive differentiation relies on proprietary agents fine-tuned on decades of historical physical-world data.
- [11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) Operating in zero-tolerance civil engineering environments requires deterministic safety boundaries, automated geofenced fail-safes, and continuous human-in-the-loop oversight.
- [16:47](https://www.youtube.com/watch?v=gDONu82P5NM&t=1007s) Deploying autonomous construction machinery is substantially easier to regulate than public robotaxis because heavy equipment operates inside restricted, 3D-mapped zones with instant shutdown triggers.
- [16:47](https://www.youtube.com/watch?v=gDONu82P5NM&t=1007s) Even though construction projects are unique prototypes, shared machine fleets—like tunnel boring machines generating 1 TB of telemetry per hour—allow centralized predictive maintenance across global sites.
- [20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) The primary strategic objective for enterprise AI in infrastructure is moving beyond showroom prototypes to industrialize scalable workflows across hundreds of active sites.
- [20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) Frontline field teams and site superintendents become the most vocal advocates for AI once real-time automated data eliminates administrative burdens and sharpens site management.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=gDONu82P5NM&t=4s) Bringing AI from Hype to the Physical World
Jason Abbruzzese of NBC News opens the session by noting that while much of the AI discourse at the conference revolves around esoteric concepts, Ferrovial operates in the realm of building massive physical infrastructure—including major undertakings like the new JFK Airport terminal. He presses Dimitris Bountolos, Chief Information and Innovation Officer at Ferrovial, on how the company moves past the technology hype to deploy AI for concrete efficiency gains and specific operational tasks on the ground.
Bountolos explains that well before the "ChatGPT moment," Ferrovial spent over a decade applying artificial intelligence directly to physical assets across highways, airports, nuclear, solar, and water facilities. A primary operational showcase is their maintenance protocol for the high-speed rail network in Spain and other international corridors. Ferrovial deploys proprietary inspection trains running daily to monitor rail geometry, detect track hazards, and identify weed overgrowth that threatens train speeds. Six years ago, Ferrovial introduced edge-deployed vision models on these trains to scan the railbed at speeds of 70 to 80 km/h (50 to 60 mph). When vegetation or hazards are detected, the system selectively triggers herbicide spray nozzles on the back of the train within one second, cutting herbicide consumption by 80%.
When asked why modern AI outperforms older computer vision or human inspection, Bountolos highlights how generative AI and agentic workflows have replaced static scripts and manual data labeling. Rather than relying on traditional OCR or standard machine learning that struggled with unusual environments, Ferrovial now uses generative models to synthesize training environments—such as pitch-black tunnels or adverse weather—and deploys agents that interpret context on the fly to self-correct and autolabel edge cases.
- Inspection trains scan high-speed rail tracks at 70–80 km/h (50–60 mph) to detect geometric instability and track hazards.
- Precision herbicide spraying achieves an 80% reduction in chemical volume through real-time edge vision inference.
- Generative AI produces synthetic datasets to train vision systems for extreme edge cases like low-light tunnels and severe storms.
---
[04:17](https://www.youtube.com/watch?v=gDONu82P5NM&t=257s) Evolution of AI: From Pre-LLM Predictive Models to Autonomous Agents
Reflecting on the distinction between traditional machine learning and the current era of generative AI and transformers, Bountolos explains that Ferrovial has relied on predictive algorithms for more than fifteen years. Because the company finances, designs, builds, and operates long-term concession assets for up to 40 years with upfront capital commitments in the billions, accurately forecasting variables like demand, degradation, and material costs has always been essential. In the pre-LLM era, data science teams manually developed predictive models against structured datasets.
The major leap forward in recent years stems from handling unstructured data and deploying autonomous agents that act as real-time proxies for human domain experts. Ferrovial now equips remote field assets—such as isolated valves, sensors, and utility meters operating on solar power with limited network connectivity—with edge intelligence. These systems perform autonomous supervision and operational triage that previously required a technician on-site or continuous teleoperation.
Bountolos reflects on the industry's trajectory since the advent of transformer architectures, noting that what was once barely conceivable five or six years ago has matured into a production reality. The core strategic challenge today is not basic feasibility, but orchestrating and industrializing agent adoption across hundreds of decentralized facilities worldwide.
---
[07:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=426s) Applying AI Across the Entire Lifecycle: Ideation, Design, and Autonomous Swarms
Applying AI across the five stages of the infrastructure lifecycle—ideation, financing, design, construction, and operations—creates a compounding advantage. In the ideation phase, agents analyze market demand, cost variables, and environmental topography in real time to simulate thousands of scenarios, such as mapping the optimal route for a high-voltage transmission line through the Amazon rainforest. In the financial modeling phase, proprietary agents scan market fluctuations to refine bill-of-quantities projections when bidding on multi-billion-dollar tenders.
During active construction, AI expands from back-office tasks to frontline operations. At an airport runway expansion project in Puerto Rico, Ferrovial operates autonomous 24/7 heavy machinery, including excavators and bulldozers:
These machines run on specialized software directed by autonomous agents functioning like a coordinated "swarm," operating strictly within digital guardrails defined by project engineers.
In engineering design, Ferrovial deploys proprietary auditing agents to parse Building Information Modeling (BIM) files containing millions of structural components and bespoke ontologies. The agents automatically detect 3D spatial clashes and design misalignments across plans submitted by disparate engineering firms, replacing tedious manual engineering reviews and coordinating corrections directly with external design partners.
---
[11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) Agent Ecosystem, Proprietary Data, and Safety in Zero-Tolerance Environments
Addressing the scale of their internal systems, Bountolos references Satya Nadella's recent projection of organizations eventually running tens of millions of agents. While Ferrovial does not deploy millions, they operate thousands of specialized agents tailored to distinct business domains. Bountolos emphasizes that while generic foundational AI components are consumed off the shelf, capabilities delivering competitive differentiation are trained and fine-tuned on Ferrovial's proprietary operational data accumulated over 70 years. He stresses that long corporate history should not be treated as a legacy liability, but as a rich source of differentiated metadata, inspection imagery, and physical-world experience that generic commercial AI models cannot replicate.
When Abbruzzese asks about safety and automation risks in high-leverage infrastructure, Bountolos insists that civil engineering operates under zero-tolerance constraints: "failure is not an option." He explains that unlike software environments where non-deterministic outputs can be tolerated, physical construction enforces strict deterministic boundaries:
- Operations are governed by precise geofenced zones with automated fail-safes and shutdowns.
- Workflows strictly adhere to human-on-the-loop and human-in-the-loop supervision depending on operational risk.
- High-fidelity synthetic training environments and video game-like simulations allow workers and autonomous agents to rehearse hazardous operations before physical deployment.
---
[16:47](https://www.youtube.com/watch?v=gDONu82P5NM&t=1007s) Autonomous Heavy Machinery and Scaling Across Global Prototype Projects
Abbruzzese asks how local regulators and government bodies, such as in Puerto Rico, react to autonomous heavy equipment on site. Bountolos points out that society has already normalized commercial robotaxis (like Tesla demonstrations and Waymo vehicles) operating in chaotic, unpredictable city streets. In contrast, autonomous construction machinery operates within tightly bounded, 3D-mapped, and geofenced zones with restricted access. The machinery features instant automated shutdown triggers for boundary crossings, unexpected obstacles, or low-visibility conditions, making the operational envelope significantly more controlled and safer than public autonomous driving.
During the audience Q&A, an attendee asks how large operators can scale AI solutions across hundreds of distributed locations—such as retail rollouts—given the bespoke nature of building sites. Bountolos acknowledges that every construction project is fundamentally a "series of one" prototype. However, Ferrovial scales by identifying shared technical foundations across its ~1,500 active global projects spanning Sydney to Toronto.
As a primary example, Bountolos points to five major subway construction projects currently underway in Sydney, London, Paris, Porto, and Toronto. Tunnel boring machines (TBMs) operate as mobile underground factories, each equipped with over 10,000 sensors generating roughly one terabyte of telemetry data per hour. Ferrovial links these TBM fleets with centralized agent architectures to analyze performance patterns in real time and forecast component failures before they cause costly underground disruptions.
- Tunnel Boring Machines (TBMs) across five world metros generate 1 TB of telemetry data per hour across 10,000 embedded sensors.
- Centralized predictive agents analyze sensor streams across international sites to anticipate mechanical breakdowns and optimize tunneling performance.
[20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) Scaling Beyond Prototypes and Empowering Field Teams
Dimitris Bountolos emphasizes that the ultimate objective of their technology strategy is achieving enterprise-wide scale rather than remaining trapped in endless experimentation. While prototypes serve a critical purpose in transferring confidence to the final user and proving out concepts, the real "obsession is to identify processes that could scale," ideally deploying across up to 1,000 active construction sites. Rather than letting innovations linger as isolated showroom experiments, the organization focuses on standardizing tools so that every project can systematically benefit from them.
An audience member follows up by asking whether these analytical capabilities are strictly reserved for the corporate level to forecast long-term project success, or if they are placed directly into the hands of field teams—such as site superintendents and operational managers—for day-to-day coordination, site scheduling, and routine micro-decisions.
Bountolos explains that frontline workers are actually the ones increasingly "raising their hands and demanding live information, precise information." Project champions who incorporate AI into field operations quickly demonstrate how automated support augments human judgment and sharpens site steering. Describing the psychological shift on the ground, Bountolos notes that "it's like an addiction": once field supervisors experience being enhanced by high-fidelity data that eliminates tedious administrative overhead and frees them to focus purely on high-value operational tasks, they become completely hooked. As a result, the operational and field teams evolve from skeptical end users into the most vocal advocates and promoters of the technology across the entire company.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=gDONu82P5NM&t=4s) ChatGPT moment — Dimitris Bountolos
> [04:17](https://www.youtube.com/watch?v=gDONu82P5NM&t=257s) The core strategic challenge today is not basic feasibility, but orchestrating and industrializing agent adoption across hundreds of decentralized facilities worldwide. — Dimitris Bountolos
> [11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) failure is not an option — Dimitris Bountolos
> [11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) Long corporate history should not be treated as a legacy liability, but as a rich source of differentiated metadata, inspection imagery, and physical-world experience that generic commercial AI models cannot replicate. — Dimitris Bountolos
> [16:47](https://www.youtube.com/watch?v=gDONu82P5NM&t=1007s) series of one — Dimitris Bountolos
> [20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) obsession is to identify processes that could scale — Dimitris Bountolos
> [20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) raising their hands and demanding live information, precise information — Dimitris Bountolos
> [20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) it's like an addiction — Dimitris Bountolos
## Mentioned
Ai4, Amazon rainforest, Building Information Modeling (BIM), ChatGPT, Dimitris Bountolos, Ferrovial, Jason Abbruzzese, JFK Airport, London, NBC News, Paris, Porto, Puerto Rico, Satya Nadella, Spain, Sydney, Tesla, Toronto, Tunnel Boring Machines (TBMs), Waymo
## Video description
Jason Abbruzzese — Assistant Managing Editor, NBC News
Dimitris Bountolos — CIIO, Chief Information and Innovation Officer, Ferrovial
The infrastructure industry is at a turning point, as AI begins to reshape how projects are planned, delivered, and maintained–and how next-generation mobility systems are designed.
This session will explore how AI is being applied across the full project lifecycle of major infrastructure initiatives, from predictive planning and demand forecasting to real-time visibility and decision-making. The conversation will explore how AI is moving beyond pilots into day-to-day operations and what it takes to integrate these technologies into complex, legacy-heavy environments.
Attendees will leave with a clear understanding of how AI is being used on the ground today and how live data is improving safety, efficiency, and overall project performance.
Ferrovial Chief Information and Innovation Officer Dimitris Bountolos contends that artificial intelligence delivers practical value in civil infrastructure only when grounded in proprietary operational data, edge computing, and strict deterministic safety controls across an asset's multi-decade lifecycle.
Edge Vision and Predictive Maintenance Operations (0:04)
Ferrovial has applied predictive algorithms to physical infrastructure for more than fifteen years, managing long-term concession assets spanning up to 40 years. In high-speed rail corridors in Spain and internationally, the company runs daily inspection trains at speeds of 70 to 80 km/h (50 to 60 mph) equipped with edge-deployed computer vision models. These systems monitor track geometry, identify hazards, and trigger herbicide nozzles within one second of detecting weed overgrowth, reducing herbicide consumption by 80%.
Modern generative AI and transformer models have largely replaced static rule sets and manual labeling. Ferrovial synthesizes training environments for difficult physical conditions, such as low-light tunnels and severe storms, allowing models to interpret edge cases dynamically. Additionally, the company deploys autonomous agents directly onto remote solar-powered utility meters, valves, and field sensors with constrained network connectivity, enabling localized operational triage that previously required on-site technical crews.
Lifecycle Orchestration from Design to Autonomous Swarms (7:06)
Bountolos outlines how deploying specialized agents across all five infrastructure lifecycle stages—ideation, financing, design, construction, and operations—compounds operational efficiency. During early ideation and bidding, agents simulate thousands of geographic and cost variables in real time, such as routing high-voltage transmission lines through the Amazon rainforest or refining bill-of-quantities estimates for multi-billion-dollar tenders.
During engineering design, proprietary auditing agents parse Building Information Modeling (BIM) files containing millions of structural elements, automatically identifying 3D spatial clashes across plans submitted by separate engineering contractors. In active construction, such as an airport runway expansion in Puerto Rico, Ferrovial operates autonomous bulldozers and excavators 24/7. These machines function as coordinated swarms governed by engineering guardrails and real-time site telemetry.
Deterministic Safety Controls and Proprietary Data Defensibility (11:06)
Because civil infrastructure operates in zero-tolerance environments where failure is not an option, Bountolos emphasizes that construction cannot tolerate the non-deterministic outputs common in conventional software. Instead, physical automations rely on strict geofenced boundaries, automatic fail-safes, and tiered human-in-the-loop or human-on-the-loop supervision. Bountolos argues that operating autonomous machinery in closed, mapped, and monitored construction sites presents a significantly more controlled risk profile than deploying commercial robotaxis on public city streets.
To power these specialized systems, Ferrovial avoids relying solely on generic commercial foundation models for its competitive advantage. The company fine-tunes thousands of task-specific agents on its proprietary archive of operational data, structural metadata, and physical inspection imagery gathered over 70 years of operating international highways, airports, and energy facilities.
Standardizing Solutions Across Bespoke Global Projects (16:47)
Although every construction project is inherently a bespoke prototype, Ferrovial achieves enterprise scale by identifying common technical foundations across roughly 1,500 active global projects. For instance, across five major subway tunneling projects in Sydney, London, Paris, Porto, and Toronto, tunnel boring machines (TBMs) operate with over 10,000 embedded sensors that produce approximately one terabyte of telemetry per hour. Centralized predictive agents analyze these live streams to anticipate mechanical breakdowns and schedule preventative interventions before costly underground stalls occur.
According to Bountolos, organizational adoption is sustained by operational field staff rather than dictated solely by central corporate management. Once site managers and superintendents see that automated agents remove administrative reporting burdens and supply accurate, live operational guidance, field personnel shift from technology skeptics to its most active advocates.
Ferrovial's core operational priority has moved from proving technical feasibility in isolated prototypes to industrializing standardized agent workflows across up to 1,000 decentralized global project sites, ensuring consistent adoption under strict physical safety constraints.
## Background
本内容由 YouTube 频道 Ai4 于 2026 年 8 月 22 日发布，记录了 NBC News 助理执行主编 Jason Abbruzzese 与全球基础设施巨头 Ferrovial 首席信息与创新官 Dimitris Bountolos 的深度对话。视频重点探讨了 AI 如何走出概念炒作，深度融入大型基建项目的全生命周期，通过实时数据与智能体技术重塑施工与运维现场。
## TL;DR
对话深入拆解了 Ferrovial 如何将 AI 与生成式智能体应用于高铁轨道巡检、亚马逊雨林线路规划、BIM 协同设计以及机场跑道自主施工等真实物理场景。Dimitris 强调，企业 70 年沉淀的一线专有数据是构建 AI 护城河的核心资产，在追求效率的同时必须遵循严格的物理护栏与人机协同机制。尽管每个基建项目都是独特的“原型”，但通过提炼底层设备与流程的“最大公约数”，AI 工具正在赋能一线工程师并实现跨千座工地的规模化扩张。
## Atmosphere & style
整场对谈务实而接地气，褪去了常见的科技布道与空泛词藻，充满硬核工业实践的厚重感。主持人提问直截了当、直击落地痛点，嘉宾的回答则冷静从容且极具现场洞察力，既有对“失败绝不是一种选项”的严谨坚守，也有谈及无人施工与智能除草时的生动幽默。观众能真切感受到尖端数字智能与传统重型机械在物理世界激烈碰撞出的扎实生产力。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=gDONu82P5NM&t=4s) 基础设施巨头 Ferrovial 在西班牙高铁巡检列车上部署边缘端小型视觉模型，能在 70 至 80 公里时速下以 0.1 秒的反应速度识别杂草并精准定点喷洒，将除草剂用量大幅削减 80%。
- [03:22](https://www.youtube.com/watch?v=gDONu82P5NM&t=202s) 生成式 AI 彻底颠覆了传统的 OCR 与人工标注流程，使系统能够在深夜、隧道等极端工况（Edge Cases）下生成高拟真合成环境，实现全仿真训练与自适应算法迭代。
- [03:22](https://www.youtube.com/watch?v=gDONu82P5NM&t=202s) AI 的应用已从过去 15 年用于风险与财务测算的结构化数据，全面拓展至非结构化工程现场；AI 智能体作为“专家代理”被部署在偏远弱网的边缘物理监测点，实现高度自治的现场监督。
- [07:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=426s) AI 贯穿了基础设施全生命周期并形成复合效应，在前期构思选线、工程造价清单敏感性分析、BIM 模型冲突检测、现场自主施工机群调度以及运维健康监测中全面释放价值。
- [11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) 传统工业企业的历史资产并非“技术包袱”，Ferrovial 过去 70 年沉淀的经过物理世界检验的专有数据与元数据，是微调垂直智能体、抵御通用模型竞争的最深护城河。
- [13:53](https://www.youtube.com/watch?v=gDONu82P5NM&t=833s) 在物理基础设施中“失败绝不是一种选项”，AI 引入的目的并非冒险，而是通过超大规模变量模拟为现实除险（De-risk），并依托严格的地理围栏与回路中（In-the-loop）人机协同确保安全。
- [16:47](https://www.youtube.com/watch?v=gDONu82P5NM&t=1007s) 相比于开放道路上的自动驾驶，施工现场具有明确界定的三维开挖体积与物理隔离，配合详尽的自动化紧急停机机制，无人重型机械作业的安全可控性与确定性更高。
- [19:31](https://www.youtube.com/watch?v=gDONu82P5NM&t=1171s) 面对工程行业“每一个工地都是独立原型、批次永远只有一件”的定制化难题，破局关键在于提取“最大公约数”——如统一解析悉尼、巴黎等多国盾构机每小时 1 TB 的传感器数据以预测故障。
- [20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) 企业技术落地不能止步于局部概念验证（PoC）的“盆景式”展示，必须聚焦于提炼具备跨场景延展能力的标准化流程，并将其大规模复制推广到上千座施工现场。
- [21:21](https://www.youtube.com/watch?v=gDONu82P5NM&t=1281s) 一线现场总监和工程师在尝到实时流数据（Live Information）减少低价值琐事消耗的红利后，迅速转变为技术落地的内生驱动力，自下而上主动推动 AI 工具的普及应用。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=gDONu82P5NM&t=4s) 走出炒作：实体基础设施中的 AI 落地与铁路杂草清除实战
在对谈开场时，主持人 Jason Abbruzzese 坦言，当天他已经听了太多关于 AI 的空泛与玄妙探讨，但基础设施行业面对的是切切实实的物理世界建造，而 AI 最核心的承诺莫过于提高效率与颠覆过往流程。作为全球顶尖的基础设施巨头，Ferrovial 正在负责如纽约甘迺迪国际机场（JFK）新航站楼等重大项目。Jason 抛出首个问题：抛开行业炒作，Ferrovial 在具体任务中是如何把 AI 真正带入日常运营的？
Ferrovial 的首席信息与创新官 Dimitris Bountolos 回应道，过去几年整个社会经历了巨大的 AI 炒作，但如今已进入将技术落地产生实质价值的阶段。早在 ChatGPT 爆火之前的 10 到 12 年间，Ferrovial 就已在高速公路、机场、核电、太阳能及水务等各类垂直领域深度探索 AI。
```mermaid
flowchart LR
accTitle: 高铁轨道智能维护流程
subgraph S1["轨道边缘感知"]
A["高速检测列车<br/>(70-80 km/h)"] --> B["边缘端小型视觉模型<br/>(Small Vision Model)"]
end
subgraph S2["实时分析与决策"]
B --> C["0.1秒识别几何隐患与杂草"]
end
subgraph S3["精准喷洒执行"]
C --> D["尾部喷嘴毫秒级定点喷洒除草剂<br/>(用量减少 80%)"]
end
```
Dimitris 分享了一个长达数年的典型应用案例：Ferrovial 负责维护西班牙及其他国家的高铁网络，每天都有专属的检测列车在线路上巡视，排查轨道几何形变或影响行车安全的障碍杂草。
- 大约六年前，他们开始在检测列车上架设摄像头，利用视觉模型在列车以 70 至 80 公里（50 至 60 英里）时速飞驰时，对铁轨进行边缘端实时扫描。
- 系统在识别出杂草后，仅需 0.1 秒即可向车尾喷嘴发出指令，进行极其精准的定点喷洒。
- 该系统直接将除草剂消耗量削减了 80%，并大幅提升了喷洒的准确性与环保表现。
### [03:22](https://www.youtube.com/watch?v=gDONu82P5NM&t=202s) 生成式 AI 与智能体进化：告别传统 OCR 与手动标注
当被问及“为什么 AI 必然优于旧有系统甚至人类操作”时，Dimitris 指出，生成式 AI 的加入彻底改变了模型的迭代效率。过去团队主要依赖传统的机器学习、人工标注或特定脚本识别图像；而现在，系统能够利用已有数据实现自动化标注，并引入智能体（Agents）进行实时纠错与算法优化。
在复杂极端条件（Edge Cases）下，如列车在深夜、恶劣天气或隧道中运行时，团队甚至可以通过生成式 AI 生成合成环境（Synthetic Environments）来进行全仿真训练。这种基于上下文理解的自适应优化，是传统 OCR 或规则脚本完全无法企及的。
谈及 AI 在大语言模型（LLM）爆发前后的演变，Dimitris 强调：
- 在过去 15 年间，Ferrovial 一直使用机器学习和深度学习处理海量“结构化数据”，用于风险预测与财务预测。基础设施通常是耗资巨大且运营周期长达 40 年的长线重资产投资，前期必须极其精密地预测各种变量。
- 如今，这一能力已全面外推至非结构化数据领域。AI 智能体作为“专家代理（Proxy of Experts）”被实时部署到工业现场。
- 无论是在偏远野外只靠太阳能板供电、网络信号极弱的边缘端，还是在每一个阀门、仪表等物理监测点，智能体都能实现高度自治的现场监督与边缘推理，而此前这必须依赖人类亲临现场或通过复杂的远程遥控来维持。
### [07:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=426s) 贯穿建设全生命周期：从亚马逊雨林选线到无人施工机群
关于 AI 在整个项目全流程中的应用与挑战，Dimitris 解释道，基础设施生命周期的五大阶段（从构思、设计、融资、建造到运营）存在明显的复合效应（Compounding Effect）。不同阶段正在被 AI 深度重塑：
| 阶段 | AI 落地场景与技术手段 | 核心价值与业务收益 |
|:--- |:--- |:--- |
| **项目构思 (Ideation)** | 部署多智能体对海量供需数据进行全天候扫描与实时分析，模拟数千种空间走廊方案（如在亚马逊雨林中设计输电线路路径）。 | 自动规避生态与地质风险，在极高不确定性下锁定最优建设路线。 |
| **工程造价 (Cost Estimation)** | 专有智能体深入供应链市场，敏感性分析各材料的通胀走势，精细化预测工程量清单（Bill of Quantities）。 | 锁定前期数十亿美元重资产投入的财务风险，提高成本测算精准度。 |
| **协同设计 (BIM Audit)** | 专有智能体自动解析由不同工程设计事务所提交的庞大建筑信息模型（BIM），对数百万个构件进行本体级别的碰撞检测（Clash Detection）。 | 大幅压缩过去由工程师手动审查图纸的周期，减少不同专业间的冲突返工。 |
| **现场施工 (Construction)** | 在波多黎各机场跑道扩建等现场，部署 7x24 小时全天候自主作业的挖掘机与推土机机群，通过桌面端智能体协同调度。 | 将现场蓝领工人和重型机械操作从危险、机械的重体力劳动中解放出来。 |
| **运维阶段 (O&M)** | 如高铁轨道健康检测与隧道实时监控，全天候感知结构病害与异物入侵。 | 规避行车风险、降低运营波动率并沉淀高质量决策上下文。 |
### [11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) 拥抱 70 年数据资产：构建专有智能体与企业护城河
针对智能体规模的讨论，Jason 引用了微软 CEO 萨提亚·纳德拉关于“未来人均拥有 80 到 100 个智能体”的观点。Dimitris 透露，Ferrovial 内部已经拥有成千上万个智能体。他坚信，企业的每一个细分业务领域、部门乃至具体流程，都需要定制具备上下文理解能力的专有智能体，以此复制专业人类的决策机制并实现组织效能的放大。
在被问及这些智能体是自主训练还是直接外购时，Dimitris 坦诚道，他们绝不会“重新发明轮子”，基础模型与开源生态的能力会被广泛引入；但真正能构建长期护城河、处理极端工况的关键模型，全部依托于 Ferrovial 过去 70 年来在工程一线所沉淀的深厚行业数据：
- 企业的历史底蕴绝非落后的“技术包袱（Legacy）”，相反，它是最核心的差异化竞争壁垒。
- 相比于展厅里那些轻巧敏捷的初创公司，传统工程企业拥有经过长期物理现实校验的专有元数据、图像标注与非结构化工程档案。
- 用这些经过现场检验的专有数据微调出的模型，在垂直任务上的表现远超通用的商业化模型。他呼吁所有与物理世界打交道的传统工业企业，务必尽快激活自身的历史数据资产。
### [13:53](https://www.youtube.com/watch?v=gDONu82P5NM&t=833s) “失败不是一种选项”：物理世界的绝对护栏与人机协同
在大型工程基础设施中，错误往往意味着生命危险与数以十亿美元计的经济损失。Jason 询问，当自主智能体越发深入业务流程时，团队对安全的态度发生了何种变化。
Dimitris 坚定地表示："在我们的行业里，失败绝不是一种选项（Failure is not an option）。" 桥梁的设计与建造不允许存在哪怕极微小的试错容忍度。因此，AI 的引入并不是为了冒险，而是为了通过超大规模的变量模拟来为现实除险（De-risk）：
- **绝对严格的物理护栏（Guardrails）与地理围栏（Geo-fencing）**：在无人机械作业区，系统会施加极度严苛的地理围栏，绝不允许在非确定性（Nondeterministic）环境下产生模糊边界。
- **全方位的人机监管机制**：严格遵循回路中（In-the-loop）、回路旁（On-the-loop）的人机协同标准，唯有在安全护栏经过多轮验证且完全闭环时，系统才被允许全自主运作。
- **沉浸式虚拟环境训练**：利用 AI 智能体将复杂的施工场景模拟成类似于电子游戏的沉浸式虚拟空间，对施工人员进行高拟真度的安全培训，将原本依靠随机经验积累的安全防范转化为系统化的预测演练。
### [16:47](https://www.youtube.com/watch?v=gDONu82P5NM&t=1007s) 像管理自动驾驶一样管理工地：波多黎各的无人机械创新
Jason 追问，当 Ferrovial 提出在波多黎各机场跑道扩建工程中使用自主重型机械时，当地政府是何种反应。
Dimitris 幽默地指出，现在公众对特斯拉展厅中的无人出租车（Robotaxi）和路上的 Waymo 自动驾驶汽车早已习以为常。相比于开放道路上面对行人、未知车辆和动态路线的极端不可控性，封闭的施工现场在安全控制上反而更为确定：
- 施工机群有精确界定的三维开挖体积，作业环境在三维坐标系中被严密映射且全封闭隔离。
- 施工区域配备了全覆盖的摄像头监控，并设定了极度详尽的自动化紧急停机（Automatic Shutdown）机制：只要有人违规跨过隔离线，或者现场能见度降低到阈值以下，整套机械系统会立即全自动刹停。
- 这种技术已在港口和高危矿区得到了充分印证，而建筑工地的规模化推广正成为新的爆发点。尽管欧美和亚洲的监管标准与落地路径各有不同，但专注于建筑自动化的大批高潜力初创企业正在迅速崛起。
### [19:31](https://www.youtube.com/watch?v=gDONu82P5NM&t=1171s) 从原型到规模化：跨国复杂工程的“最大公约数”
在现场问答环节，有听众提问：Ferrovial 承接的都是规模宏大的国家级工程，但如果是一家在全美分散运营着一千个网点的普通零售企业，该如何降维落地这些 AI 经验？
Dimitris 叹了口气，坦言工程建造领域最难啃的骨头在于——每一个施工现场本质上都是一个独特的“原型（Prototype）”，工程行业永远无法像制造业那样去批量生产长系列产品，这里的批次永远只有“一件（A series of one）”。Ferrovial 目前在全球（从悉尼到多伦多再到美国本土）同时有大约 1,500 个正在施工的资产项目。
```mermaid
flowchart TD
accTitle: 跨国盾构机集群智能协同架构
subgraph Sites["全球五个在建地铁项目 (现场数据采集)"]
S1["悉尼地铁"]
S2["伦敦地铁"]
S3["巴黎地铁"]
S4["波尔图地铁"]
S5["多伦多地铁"]
end
subgraph TBM["每台盾构机独立工况"]
D1["10,000 个传感器"]
D2["每小时生成 1 TB 数据"]
end
subgraph AI["云端智能体中枢"]
A1["多智能体分析系统 (Agents)"]
A2["潜在设备故障前置预测"]
end
Sites --> TBM
TBM --> AI
```
解决这种高度定制化矛盾的核心在于：**提取异构项目间的“最大公约数（Common Denominators）”并在局部实现标准化。**
- 比如，公司当前正在悉尼、伦敦、巴黎、波尔图和多伦多五座城市同步施工五个大型地铁工程。
- 现场投入的每台超大型盾构机（Tunnel Boring Machine）就像是一座正在运转的移动工厂，机身搭载了 10,000 个传感器，每小时产生高达 1 TB 的海量工业数据。
- 团队利用统一的 AI 智能体跨越五大项目实时解析这些传感器流，大幅提升对机械故障与地质隐患的前置预测精度。即便每个工地的地理环境千差万别，底层重型工业装备的失效模式与运维逻辑却高度一致，这也是不同规模企业寻求 AI 规模化收益时值得借鉴的路径。
### [20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) 从原型验证到千座工地的规模化扩张
在谈及技术落地的演进路径时，Dimitris Bountolos 强调了从概念验证迈向全面推广的核心逻辑。原型开发固然具有其独特的价值，它能够作为沟通桥梁，帮助最终用户建立起对新技术的信任与信心；然而，企业的终极目标绝不能止步于局部试验，团队真正的“执念”（obsession）在于识别出能够标准化的流程与解决方案，并将其大规模复制推广到多达上千个施工现场。唯有具备跨项目、跨场景的延展能力，数字化与智能化方案才能真正摆脱“盆景式”展示，转化为沉淀在整个基建网络中的核心生产力。
```mermaid
flowchart LR
accTitle: 从原型验证到规模化落地的演进流程
subgraph Lab["验证阶段"]
A["原型开发"] --> B["建立用户信任"]
end
subgraph Scale["推广阶段"]
B --> C["识别可扩展流程"]
C --> D["覆盖1000+施工现场"]
end
```
### [21:21](https://www.youtube.com/watch?v=gDONu82P5NM&t=1281s) 一线赋能：现场工程师为何对实时 AI 数据“上瘾”
现场观众随后提出了一个直击痛点的问题：企业在集团层面利用大数据和 AI 预测未来项目的成功率，但这些能力是否真正下沉到了作业一线？工地的现场总监（superintendent）和负责人们，是否能在日常排期、现场协调以及琐碎事务的管理中直接使用这些工具？
针对这一疑问，Dimitris Bountolos 明确表示，现场团队不仅在广泛使用，而且正日益成为主动“举手”索取实时、精准信息的最主要群体。那些在具体业务中率先引入 AI 工具的“技术先锋”（champions），已经用实绩证明了智能辅助系统在增强现场决策、改善项目把控方面的卓越价值。这种体验甚至被形容为一种积极意义上的“上瘾”——一旦一线人员体验到了实时信息带来的赋能，不再被繁杂枯燥的低价值事务所绑架，而是将全部精力集中在真正创造价值的紧要任务上，整个现场和运营部门就会深度融入其中，从而彻底从被动的技术接受者转变为这项解决方案最热忱的推动者与拥护者。
- **自下而上的主动需求**：工地主管与一线负责人不再是被动接受数字化考核，而是主动要求系统提供实时流数据（live information）与高精度支持。
- **决策增强与时间解绑**：AI 工具的介入并非增加汇报负担，而是通过过滤低价值消耗，让现场人员将精力完全聚焦在关键节点的调度与纠偏上。
- **倡导者效应**：现场团队尝到效率红利后，迅速转变为技术落地的内生驱动力与内部倡导者（advocators）。
## Notable quotes
> [13:53](https://www.youtube.com/watch?v=gDONu82P5NM&t=833s) 在我们的行业里，失败绝不是一种选项（Failure is not an option）。 — Dimitris Bountolos
> [11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) 企业的历史底蕴绝非落后的“技术包袱（Legacy）”，相反，它是最核心的差异化竞争壁垒。 — Dimitris Bountolos
> [19:31](https://www.youtube.com/watch?v=gDONu82P5NM&t=1171s) 每一个施工现场本质上都是一个独特的“原型（Prototype）”，工程行业永远无法像制造业那样去批量生产长系列产品，这里的批次永远只有“一件（A series of one）”。 — Dimitris Bountolos
> [19:31](https://www.youtube.com/watch?v=gDONu82P5NM&t=1171s) 提取异构项目间的“最大公约数（Common Denominators）”并在局部实现标准化。 — Dimitris Bountolos
> [13:53](https://www.youtube.com/watch?v=gDONu82P5NM&t=833s) AI 的引入并不是为了冒险，而是为了通过超大规模的变量模拟来为现实除险（De-risk）。 — Dimitris Bountolos
> [21:21](https://www.youtube.com/watch?v=gDONu82P5NM&t=1281s) 一旦一线人员体验到了实时信息带来的赋能，不再被繁杂枯燥的低价值事务所绑架，而是将全部精力集中在真正创造价值的紧要任务上，整个现场和运营部门就会深度融入其中，从而彻底从被动的技术接受者转变为这项解决方案最热忱的推动者与拥护者。 — Dimitris Bountolos
> [20:59](https://www.youtube.com/watch?v=gDONu82P5NM&t=1259s) 团队真正的“执念”（obsession）在于识别出能够标准化的流程与解决方案，并将其大规模复制推广到多达上千个施工现场。 — Dimitris Bountolos
> [11:06](https://www.youtube.com/watch?v=gDONu82P5NM&t=666s) 未来人均拥有 80 到 100 个智能体 — Satya Nadella
## Mentioned
Jason Abbruzzese, Dimitris Bountolos, Satya Nadella, Ferrovial, NBC News, JFK International Airport, ChatGPT, Tesla Robotaxi, Waymo, Tunnel Boring Machine (TBM), BIM (Building Information Modeling)
## Video description
Jason Abbruzzese — Assistant Managing Editor, NBC News
Dimitris Bountolos — CIIO, Chief Information and Innovation Officer, Ferrovial
The infrastructure industry is at a turning point, as AI begins to reshape how projects are planned, delivered, and maintained–and how next-generation mobility systems are designed.
This session will explore how AI is being applied across the full project lifecycle of major infrastructure initiatives, from predictive planning and demand forecasting to real-time visibility and decision-making. The conversation will explore how AI is moving beyond pilots into day-to-day operations and what it takes to integrate these technologies into complex, legacy-heavy environments.
Attendees will leave with a clear understanding of how AI is being used on the ground today and how live data is improving safety, efficiency, and overall project performance.
[Source](https://www.youtube.com/watch?v=gDONu82P5NM)
Source quotes
[13:53] Dimitris Bountolos: 在我们的行业里，失败绝不是一种选项（Failure is not an option）。
[11:06] Dimitris Bountolos: 企业的历史底蕴绝非落后的“技术包袱（Legacy）”，相反，它是最核心的差异化竞争壁垒。
[19:31] Dimitris Bountolos: 每一个施工现场本质上都是一个独特的“原型（Prototype）”，工程行业永远无法像制造业那样去批量生产长系列产品，这里的批次永远只有“一件（A series of one）”。
[19:31] Dimitris Bountolos: 提取异构项目间的“最大公约数（Common Denominators）”并在局部实现标准化。
[13:53] Dimitris Bountolos: AI 的引入并不是为了冒险，而是为了通过超大规模的变量模拟来为现实除险（De-risk）。
[21:21] Dimitris Bountolos: 一旦一线人员体验到了实时信息带来的赋能，不再被繁杂枯燥的低价值事务所绑架，而是将全部精力集中在真正创造价值的紧要任务上，整个现场和运营部门就会深度融入其中，从而彻底从被动的技术接受者转变为这项解决方案最热忱的推动者与拥护者。
[20:59] Dimitris Bountolos: 团队真正的“执念”（obsession）在于识别出能够标准化的流程与解决方案，并将其大规模复制推广到多达上千个施工现场。
[11:06] Satya Nadella: 未来人均拥有 80 到 100 个智能体
Source quotes
[0:04] Dimitris Bountolos: ChatGPT moment
[4:17] Dimitris Bountolos: The core strategic challenge today is not basic feasibility, but orchestrating and industrializing agent adoption across hundreds of decentralized facilities worldwide.
[7:06] Dimitris Bountolos: swarm
[11:06] Dimitris Bountolos: failure is not an option
[11:06] Dimitris Bountolos: Long corporate history should not be treated as a legacy liability, but as a rich source of differentiated metadata, inspection imagery, and physical-world experience that generic commercial AI models cannot replicate.
[16:47] Dimitris Bountolos: series of one
[20:59] Dimitris Bountolos: obsession is to identify processes that could scale
[20:59] Dimitris Bountolos: raising their hands and demanding live information, precise information
[20:59] Dimitris Bountolos: it's like an addiction
Original transcript
[0:04] I think we have one more panel that I'm pretty
psyched to talk about. Uh I've talked to a lot of
[0:08] people about AI today. Uh a lot of conversations
about very esoteric things. Uh but you guys deal
[0:16] with real world building things. And you know I
think when it comes down to where there's a lot of
[0:22] promise in AI, it is about improving efficiencies.
It is about uh doing things better than before.
[0:28] And what better company than Forovville to talk
about that right now. You guys are doing a product
[0:33] project near and dear to my heart, the JFKA uh
terminal. Um, and you guys are obviously one
[0:39] of the largest infrastructure companies in the
world. Now, that said, we're talking about AI,
[0:44] a a technology here that I think that everybody
here is probably pretty on board with,
[0:48] but there's also plenty of skepticism about
where it's going, how it should be used,
[0:53] where the safety valves. So I I think if we wanted
to start did the start with the video or did
[0:59] uh we were just starting the conversation. They
have not found the video yet but no video right
[1:04] now but if we could talk like you know give me
your favorite example of how you guys have really
[1:10] pulled AI into your organization to do a specific
task. Okay. So you you mentioned it. No I think we
[1:17] we have been exposed during the last years to to
the hype. Right now the hype I think is bringing
[1:22] more tangible examples all then are scaling for
us before the chap GBD moments during the last 10
[1:29] 12 years that we have been intensively using all
kind of AI the goal was to bring that AI into the
[1:37] ground into the operational environment or into
the process that could improve dramatically the
[1:43] way we were impacting the physical world no and
we have explored that in many different ways. is
[1:49] because we are operating um highways, airport,
nuclear, solar, water. Um the applicability in
[1:58] each those verticals is completely different. One
example very simple that we have improved along
[2:03] the the years. you know, we are maintaining the
the highest speed um network, highest speed train
[2:10] network in in Spain and many other countries.
And we have our proprietary trains that are on
[2:16] a daily basis circulating along the network
in order to verify the geometric stability,
[2:23] the potential hazards in the middle of the of the
rail track uh widths that could potentially impact
[2:31] the the speed or create any any any risk. Um we
are dossifying and we are sprrying from from the
[2:38] back of of the train herbicides in order to kill
the wheats. One implementation that we introduced
[2:44] like six years ago was to scan with cameras with
our proprietary large BCM model. Right now it's
[2:51] a small vision model running on the edge that is
continually retro feed and improve scanning the
[2:59] rail track identifying um at 70 80 kilometers per
hour 50 60 miles per hour at certain points any
[3:08] potential hazards. um 1 seconds later dosifying
selectively the the herbicides reducing 80% and
[3:16] dosifying in a much more precise way. And so I
guess one of the obtuse questions I like to ask
[3:22] as a journalist is like why is AI better than that
rather than the systems you had in place before or
[3:27] even humans because right now with generative AI
we are able to improve and accelerate the way the
[3:35] models are performing. Before we're using typical
labeling mechanism, using machine learning,
[3:43] using a specific scripts in order to identify the
frame. Right now we are able basing information
[3:48] that we have created, accelerate the autoleing,
introduce agents that are correcting, introduce
[3:55] generative AI to simulate synthetic environment
because we are circulating um in the middle of the
[4:00] night under tunnels with bad weather conditions.
Those edge cases right now are automatically
[4:07] managed with agents that are understanding
the context and improving the algorithm. No,
[4:12] that was impossible to be done before with the
typical OCRs. No, something you just mentioned
[4:17] I think is really interesting. You said you guys
have been using AI for more than six years. Yeah,
[4:23] you know I think you know for a decent number of
people maybe not the people who attend this kind
[4:28] of conference or in this room they think of AI as
maybe a three or four years ago chat GPT came out
[4:33] and everything changed but talk to me about what
it was like using AI in machine learning maybe
[4:38] prels and how that's different now in the current
moment. So had changed dramatically. Obviously
[4:46] the boundary conditions, the type of use cases
are amplified the spectrum of applicability. But
[4:51] as you said, no more than 15 years in many cases
we were using machine learning, deep learning in
[4:57] order to analyze vast amount of data. At that
moment structured data to to forecast and to
[5:05] predict. In our business typically we are running
the whole spectrum of the infrastructure design,
[5:10] financing, building and operating infrastructure.
And when we are betting for a long-term asset
[5:16] that we need to operate for 40 years putting
the money up front, we need to be as secure
[5:21] as possible of all the variables. So we were
using machine learning to predict to forecast
[5:26] and to improve the projection of many of the
variables. At that moment we were using typical
[5:32] um data scientist oriented activities models that
we were trying to confront with all the available
[5:39] data and we had extrapolate that to a much more
nonstructured type of data that we are able right
[5:49] now with the capability of the agents to triage in
real time using the agents as a proxy of experts
[5:57] and using those as a procedure of expert that
we are deploying on the field. Every valve,
[6:03] every meter, every element that could be
observed in the middle of operation right now,
[6:10] it's able to be managed in real time in an
isolated environment. In many occasions even
[6:16] using very poor connectivity and solar panel and
in an edge environment we are able to manage that
[6:25] infrastructure with a supervision that before was
impossible to be managed unless you have a human
[6:31] being in that place. No. And we were managing
remote operation, tele operation. And right now
[6:37] we are observing the performance of many of our
facilities with a type of AI that was impossible
[6:46] to be imagined five six years ago when [ __ ] and
many of the transformer started. We saw the light
[6:54] and right now it's a reality and the problem right
now is how to orchestrate and to industrialize
[7:00] the adoption of that in all of our facilities. M
and so the the title of this is you know applying
[7:06] AI across the construction life cycle. You've
given us some good examples of where AI is good
[7:11] uh to use but where in the life cycle is it
maybe more challenging to implement AI. Yeah
[7:17] it's challenging for us because at the end there
is a compounding effect. No we are not maintaining
[7:24] the level of competitiveness in each of those five
processes starting from the ideation and ending
[7:30] in the operation. The ideation right now has been
transformed completely. The understanding of when
[7:37] we need to deploy one of the infrastructure that
we are able to operate right now is is injected by
[7:46] the analysis of vast amount of data in real time
by the allocation of agents that are scanning the
[7:54] offer on demand that are analyzing the potential
impact of many variables that are um trying to to
[8:02] design thousand of scenario in order to identify
for example where is the best path and route that
[8:11] we should follow in order to put that transmission
line in the middle of the Amazonas for example
[8:17] um that was impossible to be done 5 years
ago know and that is in the ideation stage
[8:24] when we are running the analysis of what would
be the projection of the cost of the materials
[8:29] in order to identify the most probable bill of
quantities in environments where we are putting
[8:34] billions up front for the construction. Right now
we are exploring the potential evolution of the
[8:41] cost for each of the item bringing our productary
agents to analyze the markets to analyze the
[8:48] sensibility and to improve the accuracy in that
stage. You know during the construction there
[8:53] are hundred of examples where we are augmenting
the capabilities of our employees in back office
[9:00] activities in activities connected to the to
the frontline employees to the blue colors even
[9:07] progressively replacing activities that typically
were connected with the tand of a heavy machine
[9:13] and a driver. Right now for example in in Puerto
Rico we are actively working in in the airport.
[9:20] We're expanding the runway and we are using 24/7
autonomous heavy machines, excavators, bulldoers
[9:27] that are operating according to our plan are
running a software and you need to understand that
[9:34] like the proxy of what we are doing with agents
that are operating in our desktop. Those agents
[9:41] are driving conceptually each of the machine and
we are right now exploring what's the best way to
[9:47] orchestrate not one machine or two machines dozens
of machine that are acting like a swarm in order
[9:54] to operate under a very very strict one rails
in this case what we have defined or what our
[10:01] colleague has designed no we are doing the same
for example in the in the audit of our plans or
[10:07] our three designs no we have created proprietary
agents that are analyzing beam which is the
[10:14] terminology that we are using in our sector.
Million of elements with their own ontology
[10:20] scanning if there is any clash is there is any
collision in the different designs that we are
[10:25] receiving for the different engineering firms. Um
this is something that before was massively doing
[10:31] with engineers and right now we are progressively
introducing agents that are scanning all those
[10:35] plans and identifying class communicating with
the engineer frames um and so on. If we progress
[10:42] until the maintenance and operation like the
case of the train that I was mentioning before,
[10:47] each of the use cases is improving the speed,
assuring the risk, reducing the volatility,
[10:54] gaining efficiency, bringing us more context in
order to improve the quality of the decision. So,
[10:59] it's a quite wide spectrum of benefit, not
just efficiency. Sure. How many agents how many
[11:06] different agents do you guys have? Wow, we have
thousand of of of aliens. Their ratios right now
[11:13] uh Satin Adela the other day gave a speech and
made a referral about that they have in the range
[11:20] of 80 100 agents per person. 20 million is what he
said. We are not in that stage but we believe that
[11:29] the ratio connected to any domain any department
any particular activity that is proprietary of
[11:37] your business. We are fully convinced that
will require is requiring nowadays to have
[11:43] a proprietary or an specific agent capable to
reproduce the context of the human decisions
[11:49] and augment that department or that person of
that team in order to accelerate the decision.
[11:54] And that brings the importance of the context and
how to train those agents in order to replicate.
[12:01] I was going to ask how much of the agent training
is done in-house versus are you buying from from
[12:05] other companies? So we are consuming as probably
most of you atheok capabilities. We are not
[12:15] reinventing the wheels obviously and we are taking
the best of the open ecosystem in those scenarios.
[12:22] those that are really making the dent and
bringing the edge cases and connected to
[12:29] what we believe could be a long-term competitive
advantage are fully trained and developed with
[12:35] our proprietary data which is something very
interesting for companies like ours that has
[12:40] been over the field during the last 70 years no
um age is not a legacy we are probably not as fast
[12:51] and seamless reacting compared with company that
are like what we have seen in the exhibition no
[12:57] startup but we have a vast propriatory data that
is not just a legacy it's a differentiated factor
[13:03] that we are activating in the shape of metadata
labeling pictures more and more not structured
[13:11] data and what we have realized that many of
our models our agents and our algorithms are
[13:17] performing much better than those that are trying
to be commercialized because we have something
[13:24] that is very difficult to be replicated which is
the the source that has fine-tuned those model
[13:31] um and I think that's something that I encourage
to many of those companies that are connected
[13:37] with the industrial interaction of the physical
world you have experience data and historical
[13:45] legacy activate as soon as possible and convert
in your differentiated factor. I mean working in
[13:53] infrastructure obviously uh you know this is high
leverage stuff uh you know lots of money lives uh
[14:01] you know well-being um it can be a target uh
you know certainly as automation kind of gets
[14:07] introduced I think people get a little nervous
about safety talk to me about how you guys think
[14:11] and and and have changed your viewpoint on safety
as you introduce agents to more of the processes
[14:19] we need to understand that for as failure is is
not an option. That seems to be only valid in in
[14:25] the space industry which is not the case. No,
nobody designed and built a bridge considering
[14:31] that there is a small margin of of failure. No.
So what we are doing is to the risk that potential
[14:38] risk inheriting risk that is always connected with
any physical activity simulating and introducing
[14:46] more variables that what we were doing before
thanks to the AI. Yeah. And trying to guarantee
[14:52] at the same time with very strict policies that
the one rails that we are introducing in those
[14:57] uh applica applications or applicabilities
that are connected with the physical world
[15:02] are strict. We are not giving obviously options
to run nondeterministic environment where we
[15:10] are in very clear boundaries. You know, we
need to define a your fencing area in order
[15:15] to not allow an autonomous machine to operate
outside an environment that is not a very vague
[15:24] boundary. It's a very very strict boundary
and always supervised by our employees and
[15:30] always maintaining in the three ways. you know
on the loop in the loop and those that are fully
[15:35] autonomous is because we are fully convinced that
the gu rails are strictly implemented and we have
[15:42] verified continuously h that that's the the case
no but um to the risk and to improve the always
[15:52] delicate safety related condition that in any
construction or any industrial facility eventually
[15:59] happen uh because the human intervention and
would represent the behavior of thousand of
[16:04] employees in open environment. We are trying to
reduce that thanks to simulation, the analysis,
[16:11] thanks to training that we are improving with the
creation of immersive environment, thanks to agent
[16:18] that are treating those immersive environments
like a video game and we are improving those
[16:24] capabilities to simulate the training. So we have
seen that that risk has been just the opposite
[16:32] like a a benefit to reduce dramatically what was
random exposure to to a risk. No great we got
[16:40] about five minutes left. I'm going to ask one more
question but if people have the audience have any
[16:43] questions start thinking of them now. Last thing I
wanted to ask was you kind of mentioned the Puerto
[16:47] Rico example is really cool. This thinking about
automated uh kind of heavy machinery just doing
[16:53] this. Uh how did the conversation go with the
Puerto Rican government when you guys were like
[16:58] we want to bring in some automated heavy machinery
to do this? I think everybody is exposed. No,
[17:05] we have been in front of the of the robo taxi
that [snorts] in the exhibition center has been
[17:11] brought by our colleagues of of Tesla and
we have normalized in our daily life. No,
[17:17] many of you probably have been in a Whimo car
or in another activities. No. And at the end
[17:22] is essentially the same but in a much more open
environment. No, let a car to drive in an open
[17:28] road and circulate according to a route that you
are defining on the spot. have similar boundary
[17:34] conditions but in a much more open environment
compared what we are doing. No, we are defining
[17:41] the volume that we need to excavate. We are
mapping a three-dimensional environment in a
[17:46] confined area. We are isolating the environment.
We have cameras. We have an automatic shutdown in
[17:54] case any of the boundary condition is broken.
If there is someone that is crossing the line,
[17:59] if there is low visibility. So the list of
the elements that could automatically stop
[18:04] the process is so defined that when you compare
what we are letting uh to be h happen in our roads
[18:13] and what we are doing is is many factors safer.
No this is happening in the ports it's happening
[18:21] in the mines. There are other environments that
traditionally were intensively in that kind of
[18:26] activities and very risky by by the way have
been implementing progressively to do it in in
[18:33] the construction environment. It's something that
we have seen um increase it in the last months and
[18:41] the amount of startups the ecosystem and the
applications that we are seeing in particular
[18:48] coming from Asia as you know the regulation and
the and the standards probably are different or
[18:55] apply along the process of h learning how to
govern. Um this is something that is happening
[19:03] really really rapidly in in in Europe and and
in particular in US and we have seen the last
[19:08] uh months interesting startups raising um quite
impressive value proposition in that particular
[19:16] domain. Yeah. Uh questions. Anybody out there
got anything they'd want to float? question here.
[19:31] So you're you're talking about like huge scale
projects. How would you parse this down maybe
[19:37] for like a retailer in the United States where
might have like a thousand projects across the
[19:42] country operating all at the same time? [sighs]
As you know um we have right now not thinking
[19:50] in the asset that we operate no asset that are
under construction in the range of 1,500 all over
[19:56] the world from from Sydney to to Toronto and
many of them in in US no and to identify the
[20:02] common denominator in those standard processes
where you can scale the solution acknowledging
[20:08] that each construction site is a prototype
itself which is the difficulty. You know,
[20:12] we are not a an industry that could improve with
long series. Our series is a series of one. Um,
[20:20] but we are improving the the scenaries
among common areas. For example,
[20:26] we are running right now five subways, the
construction of five subways. Sydney, London,
[20:32] Paris, Oorto, Toronto. and the tunnel boring
machines that the end each of them is like
[20:39] a factory itself in motion generating one
terabytes per hour with 10,000 sensor. We
[20:46] are connecting all of them with agents in order
to analyze the data and improve the forecast of
[20:52] eventual potential failures. We have identified
across many of our divisions common denominators
[20:59] where to scale the solution that otherwise
will be prototypes. No, the prototypes are
[21:03] interesting in order to transfer the confident
to the final user. But the obsession is to
[21:09] identify processor that could scale hopefully
to up to the 1,000 construction sites. question.
[21:21] Uh you sorry testing testing uh you mentioned
that you're you know you're using at a a corporate
[21:27] level for an analyzing all your data to project
you know the success of future projects. Do you
[21:32] also give it to like the field team to the the
people in the field your superintendent or the
[21:37] people in charge of the site to use on like
a everyday basis for coordination or you know
[21:42] uh just scheduling and and all the little
bits and They they are the ones that more
[21:47] and more are raising their hands
and demanding life information,
[21:51] precise information. Those that are the champions
in particular applying AI related activities are
[21:57] the one that has demonstrated the value of that
support augmenting the decisions improving the
[22:03] steering and it's like an addiction. Once you are
enhanced by that kind of information that is not
[22:10] consuming your time and you put your emphasis
in those tasks that are uh bringing real value,
[22:17] the field and the operational areas are
really who can engage and they are the
[22:22] ones promoting and being the advocators of
the solution. Yeah, it's a good question.
