CoolSys Deploys AI to Automate Dispatch and Billing Across 3,000 Techs
Ai4
Intelligent systems deliver value in physical infrastructure by removing administrative friction and dispatch bottlenecks rather than attempting to replace manual trade labor.
## Background
Published on YouTube by the Ai4 channel on August 15, 2026, this presentation features Danny Rodriguez, Chief Information Officer at CoolSys. The session outlines how intelligent systems can generate tangible operational value in physical, asset-intensive environments where downtime is costly, data is fragmented, and service providers do not own the equipment they maintain.
## TL;DR
Rodriguez details how CoolSys coordinates 3,000 field technicians servicing mission-critical HVAC and refrigeration infrastructure for massive retail clients like Target and Walmart. Because CoolSys lacks continuous machine telemetry and operates on legacy systems, the company deploys a structured 'Sense-Decide-Act-Learn' framework with human-in-the-loop checkpoints to tackle messy field documentation, smart truck inventory, automated dispatch routing, and back-office billing reconciliation. The presentation emphasizes that AI cannot replace hands-on physical labor, but instead serves as a force multiplier to eliminate crushing administrative overhead and optimize decision-making at the operational edge.
## Atmosphere & style
The presentation is grounded, candid, and refreshingly free of Silicon Valley hype, rooted firmly in the unglamorous realities of commercial facilities and 130°F rooftop repairs. Rodriguez speaks with pragmatic authority and self-deprecating humor, drawing chuckles over Excel workbooks, accidental photos of technicians' shoes, and the essential role of corporate lawyers in keeping teams out of trouble. It feels like an authentic field briefing that bridges high-level enterprise architecture with deep, genuine respect for the physical tradecraft of frontline technicians.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=XJTvhssO7iM&t=4s) In physical asset servicing, AI creates value by optimizing workflows before and after on-site arrival rather than attempting to automate hands-on mechanical repairs.
- [00:04](https://www.youtube.com/watch?v=XJTvhssO7iM&t=4s) Because third-party service providers do not own client assets, AI systems must derive operational context from historical service records and sporadic alerts rather than continuous real-time telemetry.
- [02:27](https://www.youtube.com/watch?v=XJTvhssO7iM&t=147s) Edge-level AI dispatching must dynamically balance technician skill profiles, geographical locations, traffic, and strict two-hour emergency response SLAs.
- [02:27](https://www.youtube.com/watch?v=XJTvhssO7iM&t=147s) HVAC failures in retail environments carry extreme financial stakes, where indoor ambient temperatures rising above 75°F–80°F force complete store evacuations and shutdowns.
- [04:35](https://www.youtube.com/watch?v=XJTvhssO7iM&t=275s) Frontline AI workflows clean and structure highly fragmented field inputs, translating multilingual notes, terse summaries, and accidental photos into professional enterprise records.
- [04:35](https://www.youtube.com/watch?v=XJTvhssO7iM&t=275s) Automating back-office reconciliation relieves technicians from administrative burdens while ensuring strict compliance with client contract margin caps and verified on-site labor logs.
- [07:49](https://www.youtube.com/watch?v=XJTvhssO7iM&t=469s) Predictive truck stocking models improve first-time fix rates by dynamically shifting parts inventory between winter preventative maintenance and summer emergency breakdown cycles.
- [13:32](https://www.youtube.com/watch?v=XJTvhssO7iM&t=812s) CoolSys structures its automation lifecycle across a Sense-Decide-Act-Learn architecture that integrates automated suggestions with mandatory human-in-the-loop validation.
- [13:32](https://www.youtube.com/watch?v=XJTvhssO7iM&t=812s) Inference models analyze technician work summaries against historical repair data to flag unbilled materials and prevent missed revenue capture.
- [16:48](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1008s) Scaling operational AI safely requires strict cost boundaries, data privacy guardrails, and ongoing legal alignment to comply with client contract restrictions.
- [17:51](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1071s) AI strategy must begin with verified operational bottlenecks and frontline friction rather than implementing technology for its own sake.
- [17:51](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1071s) AI serves as a force multiplier that amplifies skilled human technicians by eliminating clerical overhead so they can focus on mechanical craftsmanship.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=XJTvhssO7iM&t=4s) Servicing Critical Infrastructure Without Owning the Asset
CoolSys operates in the physical world fixing mission-critical HVAC and refrigeration infrastructure for massive enterprise clients like Target, Walmart, grocery stores, and data centers. The stakes in these environments are substantial: if cooling fails, stores risk spoiling hundreds of thousands of dollars in perishable food or suffering severe revenue disruptions. CoolSys employs approximately 3,000 field technicians across the United States who frequently service machinery that is 30 to 40 years old. Because "the robots haven't caught up yet" to handle physical repairs, the immediate operational priority is optimizing every step before and after a technician arrives on site rather than attempting to replace human labor.
A core hurdle in applying AI to this sector is the fundamental lack of continuous telemetry. Vendors routinely ask for deep real-time machine telemetry data, but CoolSys does not own the assets—the retail clients do. At best, CoolSys receives sporadic alert pings or vague service desk tickets indicating "water dripping" or "not cooling." To prevent technicians from walking into jobs blind, the company synthesizes historical records—tracking what machinery exists on site, how often it has failed, and what parts were previously replaced—to filter and deliver the most relevant historical context directly to the field.
- CoolSys manages critical cooling infrastructure for retail giants, grocery chains, and data centers.
- The company deploys roughly 3,000 field technicians across the United States.
- Equipment serviced in the field spans modern units to 30- to 40-year-old legacy systems.
- CoolSys does not own client assets, meaning AI must function on limited event alerts and past service histories rather than continuous telemetry.
---
[02:27](https://www.youtube.com/watch?v=XJTvhssO7iM&t=147s) Operational Decisions at the Edge: Routing, SLAs, and Extreme Heat
While generative AI delivers immediate utility for desk-bound knowledge workers, operational AI for physical field execution presents distinct challenges. In field services, value stems from making better decisions at the edge rather than just spitting out better text answers. Intelligent scheduling and dispatch routing represent major focus areas. Efficiently assigning a work order requires balancing multiple dynamic constraints: matching technician experience levels to specific customer equipment, accounting for where technicians live across a city, and minimizing non-productive windshield time behind the wheel.
These dispatch decisions are governed by strict emergency Service Level Agreements (SLAs) requiring technicians on site within two hours of a distress call. CoolSys ingests incoming service requests from customer emails, phone calls, and automated Computerized Maintenance Management Systems (CMMS) like Corrigo. Peak demand coincides with intense summer heat waves. While blistering 115°F weather in places like Phoenix is dangerous to walk in, it drives heavy business as systems fail under strain. In high-stakes commercial environments like Target, if indoor temperatures reach 75°F to 80°F, corporate protocols require the store to evacuate customers and close entirely, causing massive revenue loss every hour the HVAC remains down.
- Emergency SLAs often demand technician arrival within a 2-hour window.
- AI-driven routing integrates technician skill level, equipment familiarity, geographical location, and traffic.
- Work orders flow into dispatch via automated CMMS platforms (such as Corrigo), email queues, and inbound calls.
- Indoor ambient store temperatures climbing above 75°F–80°F can force mandatory retail shutdowns.
---
[04:35](https://www.youtube.com/watch?v=XJTvhssO7iM&t=275s) The Realities of Physical Work: Messy Documentation and Back-Office Billing
Physical repair work operates under uncompromising physical laws where "there is no undo button." Context gathered from the field is heavily fragmented and notoriously messy. Technicians are hired to fix complex machinery, not to write prose; documentation ranges from non-native English speakers using Google Translate to terse, three-word notes. When asked to submit mandatory proof photos of repairs, technicians frequently upload images of the sky, their own shoes, or irrelevant surroundings. AI workflows inside mobile apps are used to clean, filter, and summarize this raw intake into structured service packages that commercial clients require.
Back-office reconciliation is equally demanding. Enterprise clients demand full transparency before paying, requiring exact time-on-site logs generated from automated Corrigo check-in/check-out timestamps. Many customer contracts cap material markups at 15% and mandate proof of purchase invoices, travel tracking for off-site part runs, and strict rules regarding which material classes are billable. The physical strain on technicians is severe—evidenced by four-hour troubleshooting sessions in 130°F rooftop heat in Dallas during August—making it critical that AI removes administrative friction from field workers and automates back-office billing compilation.
- Field inputs suffer from incomplete technician logs, language barriers, and accidental photos of shoes or skies.
- Client contracts regularly enforce strict margin caps (e.g., maximum 15% material markup) requiring documented purchase invoices.
- System integration with Corrigo automates arrival and departure check-ins to verify on-site labor hours.
- Rooftop temperatures on commercial buildings can reach 130°F during peak summer heat waves.
---
[07:49](https://www.youtube.com/watch?v=XJTvhssO7iM&t=469s) Branch Operations, Smart Truck Stocking, and Compliance Tracking
CoolSys coordinates its operations across multiple core pillars: field support, branch operations, procurement, and regulatory compliance. At the branch level, service managers handle dozens of daily service tickets across fleets of roughly 30 technicians who each complete 1.5 to 2 jobs per day. Branch dispatchers use AI interfaces to validate technician checkout notes, inspect photos before closing out tickets, and verify contract billing rules. Mobile tools also empower technicians to view historical repair summaries by asset barcode or contact the specific technician who serviced the unit six months earlier.
```
+-----------------------------------------------------------------------+
| CoolSys AI Lifecycle Engine |
+-------------------+-------------------+---------------+---------------+
| SENSE | DECIDE | ACT | LEARN |
| Raw Notes, Ticket | Recommend Parts, | Dispatch Tech,| Re-train Model|
| Ingestion, Alerts | Contract Billing, | Notify Client,| Validate Missed|
| & Asset History | Human Validation | Auto-Billing | Material Items|
+-------------------+-------------------+---------------+---------------+
```
Parts logistics and first-time fix rates are paramount. CoolSys uses AI models to analyze past job histories and dynamically optimize truck inventory, balancing the distinct demands of winter maintenance against summer breakdown surges. When a part is missing on site, mobile logic maps the nearest supply house holding inventory and tracks technician travel times. Additionally, AI tracks strict environmental regulations covering around ten distinct refrigerant gases that require meticulous capture and reporting to government agencies. To alleviate customer service desks from constant status inquiries, CoolSys is developing conversational agents that pull live field updates directly from work order logs.
- Branch managers oversee roughly 30 technicians averaging 1.5 to 2 completed work orders per day.
- Truck stock optimization balances seasonal shifts between winter maintenance and summer breakdown surges.
- Asset tracking relies on barcode tagging and historical service logging per machine.
- Environmental compliance engines track roughly 10 distinct regulated refrigerant gases to meet legal reporting mandates.
---
[13:32](https://www.youtube.com/watch?v=XJTvhssO7iM&t=812s) Data Foundations, Gamified Quality, and the Sense-Decide-Act-Learn Framework
Deploying effective AI in legacy operational environments requires accepting messy data while building disciplined data pipelines. Working alongside their implementation partner, Kaizen, CoolSys navigated disparate data pools, ranging from construction pricing estimation models housed in Microsoft Excel—jokingly referred to as "the world's greatest ERP"—to disjointed material purchase orders. To improve baseline data hygiene across regional operations, leadership made quality metrics visible on internal dashboards; operations managers who saw peers receiving superior automated reporting quickly adopted standardized logging processes to unlock the same operational tools.
CoolSys structures its automation strategy around an operational loop: **Sense** (ingesting tickets, asset barcodes, and technician notes), **Decide** (determining appropriate next steps, part allocations, and billing logic), **Act** (orchestrating notifications, status updates, and mobile workflows), and **Learn** (retaining continuous feedback loops). The system strictly maintains a "human in the loop" to validate AI recommendations before automating end-to-end steps. A newly launched model analyzes technician notes against work descriptions and historical parts data to flag unbilled materials or consumables that technicians forgot to log, serving automated suggestions to branch managers to ensure complete revenue capture.
- Partnered with Kaizen to bridge legacy spreadsheets, purchase orders, and core platforms into governed data paths.
- Leveraged dashboard transparency and peer competition to drive internal adoption of strict data entry processes.
- Operates under a structured **Sense $\rightarrow$ Decide $\rightarrow$ Act $\rightarrow$ Learn** architecture with human-in-the-loop checkpoints.
- Deployed a newly developed materials recommendation engine that infers missing job materials from technician work notes.
[16:48](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1008s) Iterative AI Rollouts and Crucial Operational Guardrails
Reflecting on their recent deployment from the prior week, Danny Rodriguez emphasizes that CoolSys is currently deep in the feedback and learning loop, continually tweaking and tuning the system. He acknowledges the messy reality of AI adoption, noting that while not every AI initiative turns out to be an immediate home run, this specific project clearly "has legs." Building sustainable momentum requires enforcing strict guardrails rather than letting models run unchecked. Human accountability, ironclad data security, and rigid cost controls are essential components of that boundary layer. Without proactive boundaries, AI compute and operational expenses can escalate rapidly out of control, making a disciplined blend of automated guardrails, active human interaction, and ongoing adjustments vital.
Beyond technical and cost boundaries, Rodriguez highlights the critical role of corporate legal and compliance teams. Half-jokingly referring to legal counsel as the people "that keep you out of jail" and keep the company out of trouble, he points out the practical benefit of close collaboration: when legal signs off, they share ownership of the risk. On a serious note, field service agreements and customer contracts often carry specific limitations regarding how client information can be handled. CoolSys remains vigilant around data compliance and contractual mandates, ensuring that customer proprietary data or sensitive work order details are not inadvertently swept into models against a client's wishes.
- **Iterative Refinement:** AI implementation requires an ongoing feedback loop of testing, learning, and fine-tuning rather than expecting an instant, flawless rollout.
- **Cost Controls & Guardrails:** AI workloads risk exponential budget overruns unless strict financial and operational boundaries are maintained alongside human oversight.
- **Contractual & Legal Compliance:** Close partnership with legal teams protects customer trust by ensuring AI processing complies with specific client contracts and data privacy constraints.
---
[17:51](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1071s) Starting with Constraints to Amplify the Frontline
Rodriguez delivers a fundamental rule for AI strategy: organizations must always start with business constraints rather than jumping at new technologies. Starting with the technology itself virtually guarantees failure. Instead, leaders must examine the actual business operations, identify deep-seated operational bottlenecks, and locate the massive manual workloads draining staff capacity. Rather than attempting a sweeping transformation across the entire enterprise at once, CoolSys targets clearly defined segments of friction and designs automation specifically to resolve them. This strategy depends entirely on going out to the field, listening to technicians, and pinpointing their most painful, day-to-day friction points first.
Ultimately, Rodriguez stresses that artificial intelligence in physical operations is never about replacing frontline workers, but about amplifying their capabilities. Drawing a direct parallel between software engineering and physical trades, he observes that "AI will make a bad developer worse, it'll make a great developer even better, and that's the same for a technician." In the physical world, AI cannot turn the wrench or swap out a compressor, but it can strip away the crushing administrative overhead that bogs workers down. By automating reporting, data entry, and procedural busywork, technicians can direct their undivided attention to the equipment in front of them, dramatically boosting frontline productivity. Wrapping up the presentation with a minute remaining, Rodriguez opens the floor to the audience for questions before thanking the room and giving attendees their time back.
- **Problem-First Philosophy:** Successful deployments target specific, verified field pain points and manual inefficiencies rather than searching for problems to match a technological solution.
- **Skill Amplification Over Replacement:** AI acts as a force multiplier that sharpens skilled workers and magnifies competence, rather than substituting for physical hands-on craftsmanship.
- **Administrative Relief:** Relieving field technicians of logistical and clerical burdens allows them to focus purely on core trade execution and customer service.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=XJTvhssO7iM&t=4s) the robots haven't caught up yet — Danny Rodriguez
> [04:35](https://www.youtube.com/watch?v=XJTvhssO7iM&t=275s) there is no undo button — Danny Rodriguez
> [13:32](https://www.youtube.com/watch?v=XJTvhssO7iM&t=812s) the world's greatest ERP — Danny Rodriguez
> [13:32](https://www.youtube.com/watch?v=XJTvhssO7iM&t=812s) Sense, Decide, Act, Learn — Danny Rodriguez
> [16:48](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1008s) has legs — Danny Rodriguez
> [16:48](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1008s) that keep you out of jail — Danny Rodriguez
> [17:51](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1071s) AI will make a bad developer worse, it'll make a great developer even better, and that's the same for a technician. — Danny Rodriguez
## Mentioned
Danny Rodriguez, CoolSys, Target, Walmart, Corrigo, Microsoft Excel, Kaizen
## Video description
Danny Rodriguez - Chief Information Officer, CoolSys
AI’s next wave will be defined not by better answers, but by better decisions and execution in the places where physical work actually happens, including trucks, branches and stores, facilities, and equipment rooms. This session explores how intelligent systems can create measurable value across field service, branch operations, procurement, and other asset-intensive environments, where downtime is real, context is messy, and trust is essential.
Attendees will learn how to strategically position where AI and other automations tools can provide value, especially when dealing with legacy systems, disconnected processes, incomplete data and lack of ownership of the equipment being serviced. Attendees will leave with a practical framework for starting small, proving value, and expanding AI safely while amplifying skilled teams.
CoolSys is deploying operational artificial intelligence to optimize field service logistics and back-office administration across legacy infrastructure, aiming to amplify its 3,000 technicians rather than replace physical labor.
Servicing Legacy Assets Without Direct Telemetry (0:04)
CoolSys maintains mission-critical refrigeration and HVAC infrastructure for enterprise customers such as Target, Walmart, and data centers. The company services machinery that is often 30 to 40 years old, under conditions where cooling failures risk hundreds of thousands of dollars in spoiled perishables or mandatory retail shutdowns when ambient temperatures exceed 75°F to 80°F. Because the retail clients own the physical machinery, CoolSys cannot access continuous machine telemetry. Instead, the company ingests sporadic alert pings, emails, and automated tickets from systems like Corrigo.
To compensate for the lack of real-time sensor streams, CoolSys uses an AI framework to aggregate historical service records, past part replacements, and asset barcode logs. This historical context is supplied to technicians before they arrive on site, preventing workers from diagnosing complex legacy failures without background data.
Dynamic Dispatch Under Strict Operating Constraints (2:27)
Field service operations require intelligent decisions at the edge rather than simple text generation. Scheduling algorithms must continuously balance technician skill levels against customer equipment types, worker home locations, and urban traffic patterns to reduce unbilled travel time. These dispatches operate under severe emergency service level agreements (SLAs) requiring on-site arrival within two hours of an alert.
Operational demand peaks during extreme heat waves, such as 115°F ambient conditions in Phoenix and 130°F rooftop working environments in Dallas. AI routing aims to ensure the right technician reaches the site with minimal transit delay while equipment strains under peak summer loads.
Automating Messy Documentation and Compliance (4:35)
Technician reporting from the field is frequently incomplete or fragmented, complicated by language barriers, brief notes, or inadvertent photos of shoes and the sky instead of repaired components. CoolSys uses mobile AI applications to filter, clean, and convert these raw field submissions into structured service packages acceptable to corporate clients.
Back-office reconciliation also demands granular tracking: client contracts routinely cap material markups at 15%, mandate purchase receipts, and audit labor hours against Corrigo automated check-in and check-out logs. In addition, branch managers track environmental compliance across roughly 10 regulated refrigerant gases. Machine learning models analyze job notes against billing descriptions to flag unbilled materials and ensure complete revenue capture.
The Sense-Decide-Act-Learn Architecture and Guardrails (13:32)
Developed with implementation partner Kaizen, CoolSys coordinates its operations through a structured loop: Sense (ingesting raw tickets, barcodes, and work notes), Decide (recommending parts and checking billing rules), Act (orchestrating dispatches and invoicing), and Learn (retraining models on verified outcomes). The system maintains human-in-the-loop checkpoints, where branch managers oversee fleets of about 30 technicians—averaging 1.5 to 2 jobs daily—and validate automated suggestions before final execution.
According to Danny Rodriguez, deploying these tools requires strict guardrails, cost controls on AI compute, and close coordination with corporate legal teams to safeguard client contractual data. Rodriguez notes that rather than pursuing broad enterprise automation, the strategy focuses on specific manual bottlenecks, using AI to strip administrative paperwork so technicians can focus entirely on physical repairs.
CoolSys views operational AI as a capability multiplier that relieves frontline trade workers from clerical and logistical overhead. The ongoing challenge centers on maintaining disciplined data pipelines from messy field inputs while navigating strict client contracts and cost controls.
## Background
本视频由 Ai4 于 2026 年 8 月 15 日在 YouTube 平台发布，记录了 CoolSys 首席信息官（CIO）Danny Rodriguez 的主题演讲。内容主要探讨在缺乏设备所有权、缺少连续遥测数据及系统流程割裂的物理工作环境中，智能系统如何在外勤服务、分支运营与供应链等重资产场景中创造可衡量的实际业务价值。
## TL;DR
CoolSys 依托约 3,000 名外勤技工负责商超与数据中心的关键暖通制冷运维，但在实操中面临缺乏连续遥测数据、现场填报质量粗糙及严苛 SLA 考核等多重挑战。公司利用 AI 挖掘历史沉淀数据，实现极端天气下的智能派单调度、现场笔记与照片的自动化合规清洗，并动态优化随车备件库存以提升首次修复率。整体架构采用“感知-决策-执行-学习”与人机协同（Human-in-the-Loop）闭环推进。演讲强调技术落地必须以一线业务痛点为先，AI 的核心使命是消除技工的繁琐行政负担，而非替代实地作业的人工双手。
## Atmosphere & style
演讲风格务实干练且极富现场画面感，彻底摒弃了空泛的科技概念炒作，以直面 130 华氏度酷暑屋顶和技工误传工鞋照片等真实蓝领细节开场。Danny Rodriguez 语言幽默生动且充满对一线员工的同理心，在清醒剖析物理世界“没有撤销键”的残酷约束时展现出资深管理者的沉稳与敏锐。整场分享节奏紧凑、案例翔实，带给听众一种深入工业级现场攻坚克难的真实感与工程落地启发。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=XJTvhssO7iM&t=4s) 物理世界关键基础设施（如商超与数据中心暖通制冷）停摆代价极其高昂，AI 的核心使命在于赋能全美 3,000 名一线技工并消除全流程低效环节，而非替代人工实地作业。
- [01:20](https://www.youtube.com/watch?v=XJTvhssO7iM&t=80s) 在缺乏设备资产所有权与持续遥测数据（Telemetry）的现实下，AI 的核心价值是通过沉淀历史维修记录与有限报警信号，在信息残缺环境下为技工提供边缘决策支持。
- [02:47](https://www.youtube.com/watch?v=XJTvhssO7iM&t=167s) 面对夏季极端高温带来的故障激增与紧急 2 小时 SLA 要求，企业通过集成客户报修与 CMMS 数据，运用 AI 算法优化就近派工调度以规避商场高温闭店损失。
- [03:45](https://www.youtube.com/watch?v=XJTvhssO7iM&t=225s) 物理工作没有撤销键且现场数据质量极差（如简短笔记、模糊照片），AI 负责清洗低质噪声并自动匹配严苛的合同报销与加价率规则，生成合规的对账计费包。
- [07:49](https://www.youtube.com/watch?v=XJTvhssO7iM&t=469s) 分支机构通过移动端出场审核把关服务质量，并利用 AI 结合季节性特征（预防性维护 vs 突发抢修）动态预测和调配抢修车的随车备件库存。
- [10:03](https://www.youtube.com/watch?v=XJTvhssO7iM&t=603s) 系统通过智能规划就近备件采购路径并自动生成多类制冷剂合规报告，全力提高暖通制冷服务的首次修复率（First-Time Fix Rate）。
- [11:29](https://www.youtube.com/watch?v=XJTvhssO7iM&t=689s) 构建实用数据层（Practical Data Layer）而非盲目追求完美数据，通过自动关联核对 Excel 报价与采购订单（PO）理顺历史遗留数据并赋能客户查询 Agent。
- [14:49](https://www.youtube.com/watch?v=XJTvhssO7iM&t=889s) 建立“感知-决策-执行-学习”闭环并推行人机协同（Human-in-the-Loop），利用 AI 结合非结构化笔记主动推断并补全技工漏报的耗材与备件。
- [17:05](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1025s) AI 落地应始终坚持业务约束先行并建立成本与合规护栏，聚焦一线最耗时、最痛的业务瓶颈，从而放大技工与专家的专业价值。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=XJTvhssO7iM&t=4s) 物理世界的关键基础设施与现场技术人员的现实挑战
作为 CoolSys 的首席信息官（CIO），Danny Rodriguez 介绍了公司的核心使命：维护和抢修超市便利店、数据中心、Target 以及 Walmart 等大型零售场所的关键基础设施，特别是暖通空调（HVAC）与商业制冷系统。保障这些系统的运转至关重要，一旦设备停摆，不仅商户无法开门营业，还可能瞬间造成价值数十万美元的冷冻冷藏食品变质腐烂。在实际作业中，现场技术人员（Technician）面对的往往是服役了 30 年甚至 40 年的老旧机组，现场环境极其复杂，故障千奇百怪，能修则修，必须更换时则坚决替换。
谈及人工智能（AI）在这一领域的应用现状，Danny 打趣道，当前的现实是“机器人技术还没赶上来，我们还在苦苦等待能够实地干活的机器人”，因此 CoolSys 遍布全美约 3,000 名现场技术工人依然是中流砥柱。引入 AI 的根本目的绝不是为了取代人工，而是聚焦于技术员抵达现场前和离开后的所有全流程支持，消除一切低效环节，为技工赋能。在蓝领技工行业，员工流失率（Turnover）向来是一场“绝对的噩梦”，通过智能化流程让技工工作更顺畅、更早收工回家，是企业提升留存率与运营效率的核心驱动力。
- 核心业务：大型商超、数据中心等关键制冷与暖通系统运维。
- 技工规模：全美约 3,000 名外勤服务技术人员。
- 核心痛点：设备老旧（30-40年）、环境极端、技工流失率极高。
---
### [01:20](https://www.youtube.com/watch?v=XJTvhssO7iM&t=80s) 没有数据遥测与资产所有权：如何在信息残缺下做决策
在科技界关于 AI 的普遍讨论中，各类供应商最常提出的要求往往是“请提供设备的遥测数据（Telemetry），剩下的交给我们”。然而在真实的商业服务场景中，CoolSys 根本无法提供这类持续流式传输的运行数据，因为公司并不是这些资产的所有者——设备的产权属于沃尔玛、Target 或星巴克。即使偶尔能获取部分信号，通常也仅仅是简单的故障警报，而非连续的监控数据流。
当缺乏连续数据源时，技术人员接到的往往只是极其含糊的服务工单，上面可能只写着“发现有水在往下滴”或者“系统不制冷/不工作了”。技术人员不能两眼一抹黑地直接奔赴现场摸索。CoolSys 通过梳理历史沉淀数据进行破局：过往在此网点修过多少次、更换过哪些零部件、历史记录显示哪些模块经常损坏。AI 在这里的价值不是像知识工作者工具那样生成优美文字，而是在信息极度碎片化、不对称的运营一线，对所有有限信号进行精炼和提炼，辅助现场做出更优决策。
```mermaid
flowchart TD
accTitle: 残缺信息下的现场决策与历史数据补全流程
subgraph S1["输入端：碎片化信息"]
A["第三方业主资产<br/>(Walmart / Target / Starbucks)"] --> B["模糊工单/报警<br/>('不制冷' / '在滴水')"]
end
subgraph S2["系统端：历史数据提炼"]
B --> C["资产历史档案库"]
C --> D["过往维修记录 / 替换部件明细 / 经手技工"]
end
subgraph S3["边缘端：决策赋能"]
D --> E["AI 关键信息摘要与上下文"]
E --> F["技工移动端决策支持<br/>(避免盲目进场)"]
end
```
---
### [02:47](https://www.youtube.com/watch?v=XJTvhssO7iM&t=167s) 边缘调度、紧急 SLA 与极端天气的运营挑战
现场运营 AI 的首要难题并非提供“更好的答案”，而是在边缘端实现“更优的决策与执行”。派单调度看似只是简单的“找个人去 Target 门店”，但现实要复杂得多：每位技术员的技术专长与经验等级各不相同、不同客户场地的机型设备差异巨大、技工各自居住在城市的不同角落，且技工只要手握方向盘堵在路上，就意味着没有在现场解决问题，这是巨大的工时浪费。更严苛的是，许多关键客户签署了紧急服务级别协议（SLA），强制要求接到报修电话后 2 小时内必须赶到现场。
每到夏季，空调制冷负荷飙升，设备故障率剧增，是业务最繁忙的高峰期。Danny 调侃道，诸如凤凰城（Phoenix）115 华氏度（约 46.1 摄氏度）的极端高温虽然让人根本不敢走到室外，但对公司的营业额来说却是“极好的生意”。在这类高温天气下，时效性关乎客户的核心营收——以 Target 等大型卖场为例，商场内部气温一旦突破 75 至 80 华氏度（约 23.9-26.7 摄氏度），法律或运营规范就会强制要求闭店疏散顾客，每耽误一小时都是巨额的营业损失。CoolSys 通过将客户来电、邮件工单以及 Corrigo 等第三方计算机化维护管理系统（CMMS）的数据全量集成，利用多套 AI 算法实现智能调度与路径优化，精准判定谁该去、何时去、去哪里，以及哪些任务必须当天抢修、哪些可以推迟到次日。
- 紧急 SLA 要求：报修后 2 小时内必须抵达现场。
- 关店阈值：商场室温达到 75-80°F 时必须清场闭店，造成重大营业中断损失。
- 天气影响：夏季 115°F 高温直接带来业务量激增，极度考验快速响应与就近派工能力。
---
### [03:45](https://www.youtube.com/watch?v=XJTvhssO7iM&t=225s) 物理工作没有“撤销键”：杂乱的现场数据清洗与计费包合规
物理世界的工作与软件世界存在本质区别——这里没有“Ctrl+Z”（撤销按钮），一旦设备被弄坏，就必须实地修好。现场采集的数据质量往往极其糟糕：技工的核心诉求是“动手修好机器”，对填报文字资料抱有极强抵触情绪；部分技工英语非母语，甚至需要依赖 Google 翻译撰写维修笔记；许多人被要求写工单笔记时，往往只草草输入两三个单词；公司要求现场至少拍摄两张维修照片以便向客户证明，技术员传上来的照片里却充斥着自己的工鞋、天空或无关杂物。系统必须依靠 AI 工作流从这些充满噪声的低质数据中提取出关键有效信息。
技术人员完成抢修离开现场后，繁重的任务直接转移到了后台计费与客服团队。大型商超客户对于开支审查极为严苛，绝不接受“维修了 6 小时、耗材 5,000 美元”这种概括性账单，甚至合同明确规定“材料加价率（Markup）不得超过 15%”，并要求审查零部件的原始进货发票、技工离场去取零件的时长以及往返时间。若合同中规定某些耗材类型不予报销，计费团队还得在合规范围内寻找可替代的相近报销科目。Danny 分享了自己入职初期在 8 月份陪同技工上门的真实经历：在得克萨斯州达拉斯市的一家 Target 门店屋顶，室外气温 115 华氏度，吸热的金属屋顶温度高达 130 华氏度（约 54.4 摄氏度），他在上面跟着技工排查故障整整站了 4 个小时。“技工的日子过得非常艰辛，”因此 CoolSys 极力通过移动端 AI 工具替他们分担文书负担，把合同逻辑自动嵌入后台工作流，生成严密无缝的对账计费包（Billing Package）。
- 数据噪声：技工填报仅写寥寥数语、照片误拍鞋子或天空、语言依赖翻译软件。
- 审查红线：客户对材料加价率设置 15% 的硬性上限，要求提供原始发票、进出场打卡与耗材明细。
- 极端作业环境：达拉斯夏季屋顶温度超 130°F，技术人员需顶着酷暑持续排查数小时。
---
### [07:49](https://www.youtube.com/watch?v=XJTvhssO7iM&t=469s) 分支机构运营、技术员协同与零部件备件的智能管理
现场支持体系不仅要实现“正确的人、在正确的时间、去往正确的地点”，更要将繁琐信息浓缩成简明扼要的移动端摘要呈现给一线。CoolSys 开发了维修历史速览功能，标明此前由哪位技工、在什么日期修过相同部件，并开通了一线技工跨班次直接沟通的渠道，让技工能直接向半年前修过该机器的同事请教经验：“嘿，我看到你 6 个月前修过这台机器，当时是什么情况？你是怎么处理的？”
在分支机构（Branch Operations）层面，管理者需要平衡技工追求高度自主的习惯与严格的业务纪律。例如一个分支拥有 30 名技工，每天承接约 50 张工单（平均每人每天跑 1.5 到 2 个工单）。在技工离开现场接下一单之前，系统与分支调度人员必须进行出场审核（Checkout Procedure）：核实与谁进行了交接、维修记录是否完整、上传的照片是否合规（绝不能是自拍或拍到鞋子）。后台利用 AI 解析复杂的客户合同逻辑，将其无缝集成到分支操作界面，辅助运营人员极速完成工单结案，并顺畅交接给计费团队。
在采购与备件供应链方面，针对遍布全美的大量老旧设备，零部件短缺是一项永恒挑战。每辆抢修车携带的备件空间有限，CoolSys 运用 AI 结合历史周期规律对车载备件库存进行动态优化。这种预测并非线性——例如 12 月份主要是以常规预防性维护（Maintenance）为主，而 8 月份则是以突发紧急故障抢修（Breakdowns）为主，不同季节的耗材消耗模式截然不同，算法必须根据业务季节性准确预测并推荐车辆应装载的物资。
- 网点负载：典型分支约 30 名技工，日均处理 50 张服务工单，人均日完成 1.5 至 2 项任务。
- 质量把关：移动端出场审核前置，拦截不合格填报与无效照片，合格后方可放行下一单。
- 车载库存动态优化：基于季节性特征（12月预防性维护 vs 8月突发抢修）智能调配车装零配件。
### [10:03](https://www.youtube.com/watch?v=XJTvhssO7iM&t=603s) 备件物流优化、合规性要求与首次修复率
在现场暖通空调与制冷（HVAC/R）服务中，“首次修复率”（first-time fix rate）至关重要。当技术人员在现场发现缺少某个零件时，系统内置的逻辑会自动判断并推荐最近且有库存或值得联系的零件供应点（parts house），并给出最佳导航路线。系统对技术人员离开和返回现场的时间全程追踪，借此分析驾驶耗时、采购零件类型，并评估是否应直接在服务车上常备这些备件。对于商业客户而言，若空调系统故障且无法在第一天修复，后果非常严重——尤其在炎热的南方地区，室内温度每上升一度都可能逼近停业阈值。
此外，资产记录与行业合规也是核心环节。团队为维护的每台设备贴上条码（bar tag）以沉淀历史数据。由于制冷行业高度受限，涉及约 10 种不同类型的制冷剂气体，法规严禁私自排放，必须全量回收记录。系统正在自动抽取这些数据生成报告，直接对接监管机构的环境合规审查。
```mermaid
flowchart TD
accTitle: 现场服务与合规工作流
subgraph S1["现场需求与调度"]
A["技术人员发现缺件"] --> B["系统智能匹配就近备件库"]
B --> C["路线规划与行程追踪"]
end
subgraph S2["数据沉淀与履约"]
C --> D["资产条形码扫描与建档"]
D --> E["制冷剂气体回收与合规记录"]
end
subgraph S3["输出与优化"]
E --> F["监管机构合规报告"]
E --> G["服务车常备库存优化分析"]
end
```
- 现场资产通过条码（bar tag）建立完整生命周期与历史维修记录。
- 依据设备使用年限不同，涉及多达 10 种制冷剂气体，必须严格执行回收与合规申报。
- 综合分析行车耗时与零件采购频次，动态调整随车备件库，全力提升首次修复率。
---
### [11:29](https://www.youtube.com/watch?v=XJTvhssO7iM&t=689s) 客户状态实时同步与“实用数据层”构建
客户沟通一直是传统现场服务的痛点。商业客户的区域负责人经常因联系不上巡店中的店长而直接致电呼叫中心质询工单进度。CoolSys 正致力于开发面向客户的智能代理（Agent），允许客户随时查询工单状态，并实时提取技术人员已录入的最新动态（例如技术人员是否已在途中、是否预计在 15 分钟内到达等）。
在底层支撑方面，演讲者强调了构建“实用数据层”（practical data layer）的紧迫性：“如果数据质量很差，AI 只会以极快的速度做出糟糕的决策”。CoolSys 并不追求完美无瑕的数据，而是基于结构化与非结构化数据的现状展开工程。合作伙伴 Kaizen 深入梳理了系统中的数据缺失与数据冗余问题。在建筑与工程计价中，大量数据存在于被称为“世界最伟大 ERP”的 Excel 表格里；系统通过将这些估价数据与进场的材料采购订单（PO）自动关联校验，构建出标准化的数据治理路径（governed pathways），保障数据流转的规范性。
- 打造客户查询 Agent，直接调用技术员实时填报的现场进度，缓解客户因高温停业风险带来的焦虑。
- 与 Kaizen 合作理顺历史遗留数据，将 Excel 预估报价与实际采购 PO 进行交叉比对与自动归集。
- 通过内部运营仪表盘推行“质量可视化”（make quality visible）：利用团队间的攀比心理推动流程规范化落地，当管理人员看到同行获取的深度报表后，便会主动按规范录入数据。
---
### [14:49](https://www.youtube.com/watch?v=XJTvhssO7iM&t=889s) 感知、决策、执行、学习与人机协同闭环
CoolSys 建立了“感知（Sense）- 决策（Decide）- 执行（Act）- 学习（Learn）”的智能化闭环体系。系统首先综合资产信息、工单详情与历史备注进行环境感知；随后推断下一步操作（如匹配合适备件、核对账单与合同条款）；进而触发自动化工作流通知相关责任人或更新系统；最后通过反馈回路不断调优模型。
```mermaid
flowchart LR
accTitle: AI 决策与人机协同闭环
Sense["1. 感知 (Sense)<br/>工单/资产/备注/数据"] --> Decide["2. 决策 (Decide)<br/>推荐步骤/备件/账单匹配"]
Decide --> HITL{"人机协同<br/>(Human-in-the-Loop)"}
HITL --> Act["3. 执行 (Act)<br/>触发工作流/状态更新/通知"]
Act --> Learn["4. 学习 (Learn)<br/>收集反馈/模型微调迭代"]
Learn -.-> Sense
```
演讲者坚定倡导“人机协同”（Human-in-the-Loop）：AI 与 RPA 负责前端分析，由人类专家在关键节点把关验证；随着系统准确率提升，再逐步移除部分人工检查点。系统近期上线了“耗材与备件智能推荐”功能——技术人员在完成维修后往往会漏填使用的零配件或耗材，AI 通过综合技术员工作笔记、工单描述、客户报修诉求及历史相似案例，主动向分公司或服务经理提示可能遗漏的材料项。尽管该模型刚刚上线一周且仍处于调优阶段，但已展现出明确的业务价值。
- 针对技术员容易漏报耗材的问题，利用 AI 解析非结构化笔记并推断可能使用的零件。
- 引入阶段性人工校验，在模型成熟后逐步释放自主权，实现受控的自动化推进。
- 保持敏捷迭代，不苛求每个 AI 项目都一击即中，重点在于持续收集反馈并优化模型。
---
### [17:05](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1025s) 安全护栏、成本控制与业务约束优先原则
在规模化应用 AI 时，必须建立严格的安全与成本护栏。AI 若缺乏约束会导致计算和调用成本失控，因此必须设置明确的开销监控。在法务与合规层面，必须与法律团队紧密协同，严格审查与客户签订的服务合同，确保数据使用边界符合法律法规与客户协议约定。
技术落地的核心指导原则是“从业务约束出发，而非从技术本身出发”。若一开始就执着于某种特定技术，项目注定失败。企业应当深入一线调研技术人员的最大痛点，识别流程中高耗时的手工环节并进行精准自动化改造。演讲者强调，“AI 不会取代工人的双手，而是放大他们的能力”——正如 AI 会让拙劣的开发者写出更差的代码、让优秀的开发者如虎添翼一样，AI 的终极目标是将技术人员从繁琐的行政事务中解脱出来，让他们全力专注于核心的现场维修作业。
- **成本与合规护栏**：严格把控 AI 运行成本，结合法务团队审查客户合同与数据隐私边界。
- **业务约束先行**：不盲目追求前沿技术，优先针对现场一线最耗时、最痛的业务瓶颈进行局部攻坚。
- **放大人员价值**：AI 的定位是协助而非替代现场操作，通过消除行政负担全面提升一线员工的单位效能。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=XJTvhssO7iM&t=4s) 机器人技术还没赶上来，我们还在苦苦等待能够实地干活的机器人 — Danny Rodriguez
> [01:20](https://www.youtube.com/watch?v=XJTvhssO7iM&t=80s) 请提供设备的遥测数据（Telemetry），剩下的交给我们 — Danny Rodriguez
> [01:20](https://www.youtube.com/watch?v=XJTvhssO7iM&t=80s) 发现有水在往下滴
> [01:20](https://www.youtube.com/watch?v=XJTvhssO7iM&t=80s) 系统不制冷/不工作了
> [02:47](https://www.youtube.com/watch?v=XJTvhssO7iM&t=167s) 极好的生意 — Danny Rodriguez
> [03:45](https://www.youtube.com/watch?v=XJTvhssO7iM&t=225s) 技工的日子过得非常艰辛 — Danny Rodriguez
> [07:49](https://www.youtube.com/watch?v=XJTvhssO7iM&t=469s) 嘿，我看到你 6 个月前修过这台机器，当时是什么情况？你是怎么处理的？
> [11:29](https://www.youtube.com/watch?v=XJTvhssO7iM&t=689s) 如果数据质量很差，AI 只会以极快的速度做出糟糕的决策 — Danny Rodriguez
> [11:29](https://www.youtube.com/watch?v=XJTvhssO7iM&t=689s) 世界最伟大 ERP — Danny Rodriguez
> [17:05](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1025s) 从业务约束出发，而非从技术本身出发 — Danny Rodriguez
> [17:05](https://www.youtube.com/watch?v=XJTvhssO7iM&t=1025s) AI 不会取代工人的双手，而是放大他们的能力 — Danny Rodriguez
## Mentioned
Danny Rodriguez, CoolSys, Target, Walmart, Starbucks, Corrigo, Google 翻译, Kaizen, Excel
## Video description
Danny Rodriguez - Chief Information Officer, CoolSys
AI’s next wave will be defined not by better answers, but by better decisions and execution in the places where physical work actually happens, including trucks, branches and stores, facilities, and equipment rooms. This session explores how intelligent systems can create measurable value across field service, branch operations, procurement, and other asset-intensive environments, where downtime is real, context is messy, and trust is essential.
Attendees will learn how to strategically position where AI and other automations tools can provide value, especially when dealing with legacy systems, disconnected processes, incomplete data and lack of ownership of the equipment being serviced. Attendees will leave with a practical framework for starting small, proving value, and expanding AI safely while amplifying skilled teams.
[Source](https://www.youtube.com/watch?v=XJTvhssO7iM)
Source quotes
[0:04] Danny Rodriguez: 机器人技术还没赶上来，我们还在苦苦等待能够实地干活的机器人
[1:20] Danny Rodriguez: 请提供设备的遥测数据（Telemetry），剩下的交给我们
[1:20] 发现有水在往下滴
[1:20] 系统不制冷/不工作了
[2:47] Danny Rodriguez: 极好的生意
[3:45] Danny Rodriguez: 技工的日子过得非常艰辛
[7:49] 嘿，我看到你 6 个月前修过这台机器，当时是什么情况？你是怎么处理的？
[11:29] Danny Rodriguez: 如果数据质量很差，AI 只会以极快的速度做出糟糕的决策
[11:29] Danny Rodriguez: 世界最伟大 ERP
[17:05] Danny Rodriguez: 从业务约束出发，而非从技术本身出发
[17:05] Danny Rodriguez: AI 不会取代工人的双手，而是放大他们的能力
Source quotes
[0:04] Danny Rodriguez: the robots haven't caught up yet
[4:35] Danny Rodriguez: there is no undo button
[13:32] Danny Rodriguez: the world's greatest ERP
[13:32] Danny Rodriguez: Sense, Decide, Act, Learn
[16:48] Danny Rodriguez: has legs
[16:48] Danny Rodriguez: that keep you out of jail
[17:51] Danny Rodriguez: AI will make a bad developer worse, it'll make a great developer even better, and that's the same for a technician.
Original transcript
[0:04] I work for CoolSys. I'm the CIO there and we do we
fix critical infrastructure, HVAC, refrigeration,
[0:12] so you know, your your grocery stores, data
centers, Targets, Walmarts. We make sure
[0:18] they get to keep the doors open and don't ruin
hundreds of thousands of dollars worth of food.
[0:23] So, we have techs. Techs get to show up and a
lot of times we're servicing equipment that's 30,
[0:28] 40 years old. Everything under the sun we need to
show up and and take care of repair if possible,
[0:33] replace if needed. So, AI is a little tricky
because the robots haven't caught up yet,
[0:38] so we're still waiting for those. So, in the
meantime, we have about 3,000 techs that are
[0:42] kind of out and about around the US. And we have
to make our processes as efficient as possible.
[0:48] And in doing that, it's everything leading up
and everything after to enable the tech to be
[0:53] as efficient as possible and really solve
their the whatever problem they encounter.
[0:59] So, um we're not replacing the tech, we're just
trying to make them more efficient. We're trying
[1:03] to make the entire process more efficient. It's
more efficient for us. Techs is going to have a
[1:07] better day. They're going to get out earlier and
turnover in our industry is an absolute nightmare.
[1:13] Um so, anything to make the techs' life easier
is definitely something that we try and push.
[1:20] So, we don't own the asset. Um I know in a lot
of these AI sessions and over the years when,
[1:25] you know, I speak to different vendors, it's
always like, well, give us our telemetry and like,
[1:30] we don't have it, right? And they're
like, what do you mean? It's like, no,
[1:32] we don't own the asset. Walmart owns the asset.
Target owns the asset. Starbucks owns the asset.
[1:37] We just have to show up, figure it out. Now, there
are cases where we do get some telemetry data
[1:42] and it's usually alerts, so it's not the constant
stream of data. So, when you don't have the data,
[1:47] how do you know what you're going to fix? You
don't. We get a service ticket that just says,
[1:53] hey, we see water dripping or we get a ticket
that says it's not cooling or it's not working.
[1:59] And the tech has to go out there and figure
it out. So, we're definitely not trying to
[2:04] replace the tech. It's more how do we use all the
limited signals to get them as much information
[2:10] going into it so they're not going in blind. And
a lot of that is historic. What have we done? The
[2:15] assets on site, we do keep track of them. How many
times have we fixed it? What have we replaced? So,
[2:20] how do we take all that data and really, you know,
stream it down to give them what's relevant. So,
[2:27] AI is going to be operational. We know for
knowledge workers, um it it does an amazing
[2:32] amount of things for for knowledge workers. But
operational is probably still where there's a
[2:36] bit challenge, right? So, whenever I'm looking at
use cases, it's again, give me all the telemetry,
[2:42] we'll figure it out, but when we really break
it down, it's not just the better answers,
[2:47] but better decisions at the edge. So, how
do we route our techs most efficiently?
[2:51] And it sounds easy. This is hey, just send
it to the target, but the reality is there's
[2:56] different levels, there's different experiences,
uh different equipment at different um customers.
[3:02] They live in different places in the city, and
anytime they're behind the wheel, they're kind
[3:06] of wasting time not solving a problem. We have
a lot of emergency SLAs that say we have to be
[3:11] there within like 2 hours of getting a call.
So, we use a lot of different AI solutions for
[3:17] scheduling and routing. It's a huge area
where we where we spend a lot of time.
[3:23] Better execution in the workflow. So, leading up
to it, there's a lot of things that have to occur.
[3:27] So, customer calls, we have to know who they are.
Customer sends a uh email with uh service desk um
[3:34] sometimes it's all automated with Corrigo and
some of these other CMS platforms, so we ingest
[3:39] all of that to figure out who do we send, where
do we send them to, when do they have to go?
[3:45] Summer's obviously our busiest time. Um we want
this kind of Phoenix weather. 115° is great for
[3:51] us because things break down and things need
to be serviced, so it's great for business.
[3:55] I don't recommend walking outside right
now, but it's great for our company.
[3:58] So, all that efficiency that leads up to who do we
send, when do we send them, where do we send them,
[4:04] and then how fast do they get there, or do they
even have to go today can wait till tomorrow.
[4:09] To put it in perspective, some of our
big stores use Target as an example.
[4:13] When the temperature in the store gets up over
a certain temperature, and I think it's 75-80°,
[4:18] they have to shut the store down. They have to
send everybody out, and that's all lost revenue.
[4:22] So, they don't really appreciate when they have
to do that, so we have to get there on time.
[4:26] So, schedule, notify, quote, document.
How do we do that efficiently?
[4:35] The physical work has different rules than
AI. There is no undo button. If we break it,
[4:39] we have to go fix it, right? Um the context
is very fragmented. We talked about all the
[4:44] different places we can get data from, and data is
not clean. I mean, you take a technician's note,
[4:49] some technicians barely speak English. Um
some technicians are using Google Translate
[4:54] to document their notes. Um some technicians,
great English, but guess what? Technicians do
[4:59] not like to document. They want to go
fix stuff, right? So, when you say,
[5:03] "Oh, and you have to document your notes." They're
like, "Oh, great." And you might get three words.
[5:07] Um you have to take pictures, right? We say,
"Hey, minimum two pictures, right? Of of what
[5:11] you're fixing, cuz our clients want to see this."
We get pictures of their shoes, we get pictures of
[5:15] the sky, and everything in between, right? So,
how do you take all all that context, filter
[5:21] out the noise, and still make it relevant to not
only what you're sending your your end customer,
[5:26] but also again, how do you enable that
tech? So, even after they do all the work,
[5:32] we still get more information, like I was
mentioning on the tech notes and pictures
[5:37] and what was fixed and asset inventory
information. Now, we have to take all that and
[5:42] get it to the customer, cuz the customer doesn't
want to just take our word that says, "Hey, we
[5:46] were there for 6 hours and by the way, it's it's
$5,000 in material." They actually want detail.
[5:53] So, we take all that out of the workflow that
we've kind of built with, you know, AI workflow
[5:57] and automation for the tech on the field in their
kind of mobile app. And now we have to synthesize
[6:02] it for our customers to say, in some cases, we
have automated messaging that goes with them via
[6:07] Corrigo. So, the moment our technician gets there
and checks in, they know our technician's there.
[6:13] Our technician checks out, they know them. So,
the customer can now calculate how many hours
[6:17] they were on site. Billing, tons of fun in this
space. Customers will say, "You can't You can't
[6:22] charge me more than 15% markup on materials."
So, now you have to show them your material cost.
[6:29] As part of that package you're sending, they know
if your They know how long your tech's been there.
[6:33] Uh did they have to leave to go get a part? Did
they come back? Show me the invoice on the part.
[6:38] So, getting that billing package together,
again, combination of what did the tech do?
[6:43] What were the notes? How long were they there?
What materials were they used? And then it
[6:47] even gets more fun when the contract says, "We
won't pay for this type of material." Right? So,
[6:51] then you get a little creative and you say,
"Well, is there a like material I can charge for?"
[6:55] So, the tech just getting there and doing the
job, I'm not going to say that's the easy part
[7:01] because they have a They have a a rough life.
I I shadowed a tech in August last year or when
[7:05] I started 2 years ago. And the tech had me on
the roof of Target in Dallas in August. Anybody
[7:12] that's from Texas knows that about 115, probably
about 130 30 on that roof and I was up there for
[7:16] 4 hours with them while we were troubleshooting.
So, they have a rough life. So, how do we make
[7:21] their life easier getting there? How do we
make their life easier while they're there?
[7:24] And then, I'm not going to say shift the work,
but the work really does shift to the back office,
[7:28] to our billing teams, to our customer service
reps, to anybody that's now interacting with
[7:33] not the person on site at the customer, but
they're interacting either at a corporate level
[7:37] or at a regional level, and making sure that it's
all communicated what we did and how we did it,
[7:42] right? So again, a lot of disparate data all
having to be generated and pulled together.
[7:49] Um so, some of examples on kind of where we're
showing off. I talked about field support and
[7:56] the ability to schedule, send the right person,
make it efficient. Uh right time, right place,
[8:02] right person is very key. Uh the documentation
that they're going to be looking at, it's got to
[8:06] be on their mobile, it's got to be in they got
to be able to digest it in a very easy manner.
[8:11] We do a lot of summary notes for them. We do a
lot of summary history on this part was fixed on
[8:17] this date and it was by this tech, right? We also
enable them to contact other techs. So it's like,
[8:22] "Hey, I saw you were here 6 months ago. What
happened? What did you do?" So all these
[8:27] things again make their life easier. Our branch
operations. So techs love to be independent,
[8:33] they don't really want to be told what to do or
when to do it. That obviously doesn't work for us.
[8:37] So our branch has to have the ability to A
dispatch, B we talked about getting them there.
[8:44] Their checkout procedures. What did you do? Who
did you talk to? Make sure all the documentation,
[8:48] make sure it's not a picture of their shoes. All
that gets looked at before they get released to
[8:52] the next job. The branch also has to see it's
like, "Okay, I had 50 tickets come in today,
[8:57] 50 service tickets cuz I have 30 techs. They each
did one and a half, maybe two jobs that day."
[9:02] I have to close them out. I have to put all
the notes in. I have to check. So we build
[9:07] their their screens and their kind of working
environment. We take the contracts and we're,
[9:11] you know, basically taking information out of
the contract and building in all of that logic
[9:16] and using AI to help them to say to make how
do we make that the easiest thing for them to
[9:21] close out the work order. All the financials,
right? Um then that goes to the billing team.
[9:26] Procurement and parts. You can imagine uh
when we're looking at 30-year-old uh equipment
[9:32] all over the country. In some cases it's new,
in some cases it's old. Parts are a challenge.
[9:38] Um we stock our trucks with X number of parts that
we try and optimize using AI based on history.
[9:45] The history is a little tricky because what
we're doing in December with maintenance is
[9:48] very different than what we're doing
in August with with breakdowns. So,
[9:52] we are using a lot of AI to analyze
this is what should be on the truck.
[9:56] All right. Um in addition, we try and help the
tech. So, if they're on site, something's broken,
[10:03] they're looking for hey, I need this part.
We have a lot of logic built that says, "Oh,
[10:07] by the way, this is the closest parts house
to you that has the part or that you should
[10:11] call cuz if not all parts houses are going to
share inventory with us." And more importantly,
[10:16] here's how you get there. We have a track, so
we know now when the tech leaves comes back.
[10:20] It's all being tracked, so we can do analytics
on how much time our tech is spending driving,
[10:26] what are the parts they're buying, and do we need
to have more of those on a truck to make them more
[10:30] efficient. So, in our world, first-time fix is
is huge, right? No, I mean, think about it. If
[10:36] you were home right now and you live anywhere in
the South, if the technician says, "Hey, I got to
[10:40] go to part. I'll be back tomorrow the next day."
and you have no AC, right? So, it's no different
[10:45] with our commercial customers. They don't like
it when we can't fix stuff in the first day.
[10:50] And that kind of goes to our customer experience,
right? So, um all the assets need to be
[10:56] documented. So, we bar tag all the assets we
work on. We have history on what we've done.
[11:01] Um compliance. Um we're actually a very
highly regulated industry with all the gases
[11:06] and different types of gases. I think there's
like 10 different types of gases depending on
[11:10] how old the system is and you have
to capture it. It can't be released.
[11:13] So, we have a lot of regulatory, environmental
compliance um things that we have to track. So,
[11:18] as we're gathering all that information,
we're now kind of extracting and building
[11:22] a lot of things out to to report, right? To
report to the to the regulatory agencies.
[11:29] Um and lastly, customer updates. We're
not the best at this. This is actually
[11:34] something we're working on now. A lot of times
we'll have customers calling in like, "Hey,
[11:38] what's the status?" Cuz you have someone from the
region calling in about about a store, right? And
[11:43] they can't get a hold of their own store manager,
and the store manager's probably walking around
[11:46] somewhere. And they don't even know what's going
on. So, we get a lot of calls in our into our call
[11:50] center. "What's going on with this?" So, part of
what we're working on now is actually this kind
[11:55] of fourth pillar. So, that you know, can we build
an agent that our can just ask what's the status
[12:03] with this ticket, what's the status with
the service call, and it can pull live,
[12:07] you know, live information from whatever
the tech has documented to that point.
[12:11] Um, it's not locked in yet, so that's where it
gets tricky. But, how do we how do we make it
[12:15] easier for our customers to get updates
on is someone there? Are they in route?
[12:21] Are they going to be there in 15 minutes? You
know, that type of information that customers
[12:24] get really antsy about when, again, the store's
80° and you're you're 3° from shutting down.
[12:32] So, a practical data layer, um, I'm sure
everybody's heard it. Data, if you have bad data,
[12:39] you just with AI you're just going to get a
really bad decision really fast. So, um, our
[12:44] data I'm not going to say is is pristine, but uh,
based on that structured and unstructured data,
[12:50] that's the best we got. So, that's what we've
got to work with. And it gets really creative,
[12:55] right? Um, with our partner Kaizen, they've
learned a lot about our data that about our our
[13:02] lack of it in some cases and excess data in the
other. And then, how do you tie it, and how do you
[13:07] map it? And in some cases we're doing, you know,
pricing in in Excel, which is the world's greatest
[13:12] ERP. And, um, how do you take that for a pricing
estimation in construction, and then tie it to
[13:19] all the material POs as they're coming so that you
can actually tie and say, "Yeah, that was a legit
[13:23] purchase made for this job, right?" So, data is is
absolutely critical. As much as you can clean up,
[13:30] as much as you can tie, that's just going to make
everything a lot easier. Um govern pathways. So,
[13:37] we were a little bit of a hot mess as to how
the data was coming in. We had to build a lot
[13:40] of those pathways and just that structure that
says, "This is how we expect to see the data."
[13:46] Here are some things we built in the process. So,
all that now is flowing a lot cleaner. We still
[13:50] have a history that we deal with, but over time,
you know, that's just going to get a lot better.
[13:55] And then make quality visible. So, um what
we started doing is just with, you know,
[14:01] dashboards and things like that, when our own
internal operation folks, they will see something
[14:06] that their peer has, and they will ask, "Why can't
I get that?" I'm like, "Well, cuz you guys don't
[14:10] do the following." So, if you kind of follow the
process, um you'd be able to get that, too, right?
[14:16] And just pinning them one against each other, it's
amazing how much you're like, "I want that, too."
[14:21] Right? And when they can see it, now it's, "Well,
if you want that, this is what you have to do."
[14:26] And you've told them that 50 times before, but
they didn't want to change their processes,
[14:30] and sometimes that's a challenge. Once they see
what they can get, now all of a sudden, "Oh,
[14:35] wait, I'm willing to do that because I see the
value." Right? So, that's kind of what that making
[14:40] that quality visible, um showing the value, and
then people usually start adopting pretty quick.
[14:45] [sighs and gasps]
[14:49] This is kind of how we look at it, right? Sense,
figure out what's going on, um look at all the
[14:54] data, the assets, the job, the notes, um decide,
"What should we do?" Based on everything we know,
[15:02] this is a reasonable next step, this is a part we
should have, this is the type of bill we should
[15:07] send out, this is the contract um tying to the
actual material. So, there's a lot of different
[15:15] spaces before it gets to the tech and after it
gets to the tech that you can make that decision.
[15:20] Now, I'm a big fan of keep a human in the
loop. So, we have a lot of our examples that
[15:26] have different ones that we can kind of go through
that. AI does a lot. Sometimes it's automation,
[15:30] sometimes it's RPA, but at the end of the
day there's a human that's going to validate
[15:34] at different points. Now, once we get more
comfortable in that, sometimes we'll pull certain,
[15:39] you know, human checkpoints out of that process
and let the AI kind of run a little bit with it.
[15:45] Um act, you know, triggering that workflow, uh
notify someone, take action, update a field,
[15:51] update a person, update a customer. There's
quite a bit that we're kind of orchestrating
[15:57] throughout this process. And then learn. So, we're
constantly doing a, you know, feedback loops. So,
[16:03] what's working, what's not? We actually just went
live with a materials kind of recommendation.
[16:09] Techs do a great job fixing stuff. Sometimes they
forget to actually say, "Oh, I used this part." Or
[16:15] um I used this consumable. So, when the
branch manager gets it or the service manager,
[16:20] now we're actually putting AI to say, "Analyze
their notes, analyze the tech notes, analyze the
[16:24] job description, analyze the customer complaint."
Look at the materials in the job, look at history,
[16:30] find similar scenarios, and say, "Hey, you may
have forgotten this based on what we can infer
[16:35] from all that data." Right? So, now the branch
manager is not having to read through all that,
[16:40] they're getting a recommendation.
Admittedly, it's so-so right now.
[16:44] We're kind of going back and redoing that model
a little bit. We literally just launched this
[16:48] last week. Um so, we're just in that feedback
um and learning and kind of going back and
[16:54] doing a lot of the tweaking. So, um it's
iterative. I think we've all dealt with
[16:58] that. Not every AI project's a home run,
but this one looks like it has legs.
[17:05] Uh the guardrails. Um I think we've touched on
most of these. Human accountability, touch that,
[17:12] security, cost of controls. At the end of the day,
AI can get very expensive if you let it run out of
[17:17] control. So, I'm I'm big on put the guardrails
in place, uh leverage that human interaction,
[17:23] and adjust as you go. And then, legal and
compliance review? Um lawyer's your friend,
[17:29] right? They're the ones that keep you out of jail,
they're the ones that keep you out of trouble,
[17:32] cuz then you can always blame legal. Hey, legal
looked at it and said it was okay. That's kind of
[17:36] how me and legal work very close together, but
um we got to be careful with our contracts with
[17:40] our customers. Um customers may or may not
want certain things being included, so we are
[17:45] very cautious to make sure we're maintaining
data compliance and all that other fun stuff.
[17:51] Um start with the constraints, not the
technologies. Um if you're starting with
[17:56] technology, you already failed, right? Look at
the business, look at the challenges, look at
[18:00] the large manual efforts that everybody's going
through, and that's where you start attacking
[18:05] it. Don't don't attack everything at once, at
least we're not trying that. We're picking very
[18:10] defined segments of inefficiency in our process,
and saying, "How can we bring automation to it?"
[18:15] Work the field to understand their biggest
pain points, and that's what we attacked first.
[18:21] Um AI will not replace the hands, it will just
amplify them. AI will make a bad developer worse,
[18:27] it'll make a great developer even better,
and that's the same for a a technician.
[18:32] The more we help them focus on the job at hand,
so they're not worrying about the administrative,
[18:36] the more efficient they can be in that process.
[18:41] And we've got a whole minute for questions. I
thought I was going to do that a little quick,
[18:45] so I apologize. Um but I do want to leave a
little bit of time. If there are any questions,
[18:48] be more than happy to answer them.
[18:55] And if not, then you guys
get a minute back. Thank you.
