Grupo Traxion Automates Truck Dispatch to Cut $60M in Downtime Losses

The Last Mile of Data How AI Turns Enterprise Data Into Decisions with Grupo Traxion — Ai4

Ai4

Grupo Traxión reduces unscheduled fleet downtime by up to 48% and eliminates $60 million in annual losses by replacing passive dashboards with autonomous AI agents that lock failing trucks from service before highway breakdowns occur.

AI-assisted summary of The Last Mile of Data How AI Turns Enterprise Data Into Decisions with Grupo Traxion — original source. Check the original for full context.

Source published: 2026-08-23

Detailed notes

## Background Published on YouTube by Ai4 on August 23, 2026, this presentation features Joshua Bernal, CIO of Grupo Traxión, addressing 'the last mile of data'—the critical gap between collecting enterprise data and executing timely operational decisions. The session explores how AI-native Decision Intelligence platforms, conversational analytics, and autonomous AI agents replace static dashboards to deliver contextual insights and rapid business actions.

## TL;DR Joshua Bernal outlines the immense data challenge of managing Grupo Traxión's 25,000-vehicle fleet, highlighting how traditional BI tools often stall at retrospective insights and leave up to 80% of corporate dashboards abandoned. To solve this last-mile disconnect, the organization deployed Northstar, a conversational decision intelligence layer that achieved over 97% predictive production accuracy across critical metrics like component failure and churn. The platform transitions analytics from passive alerts to closed-loop autonomous execution, such as proactively locking at-risk vehicles out of dispatch systems and scheduling maintenance. This automated intervention reduces unplanned fleet downtime by 23% to 48%, avoids more than $60 million in annual breakdown expenses, and consolidates fragmented analytics tools to cut roughly $470,000 in software licensing.

## Atmosphere & style The presentation carries the authoritative, pragmatic tone of an executive operating at massive industrial scale, anchoring modern AI discussions in hard logistics realities. Bernal moves briskly from the tangible operational headaches of metric reconciliation and costly highway breakdowns to crisp architectural diagrams and measurable unit economics. The delivery is structured, candid, and devoid of empty hype, giving viewers an empowering, practical roadmap for turning enterprise data into decisive action.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=4s) Modern enterprises generate massive real-time operational streams, but cross-departmental silos prevent teams from operating on a single version of the truth. - [01:37](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=97s) The fundamental enterprise bottleneck is the 'last mile problem'—the disconnect between possessing raw data (such as 400M daily GPS pings across 25,000 trucks) and translating it into decisive operational execution. - [03:31](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=211s) Between 60% and 80% of corporate dashboards are abandoned within months because traditional BI tools present retrospective facts rather than context-aware operational guidance. - [07:11](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=431s) Proactive scheduled maintenance costs 3x to 5x less than unplanned highway breakdowns ($2,000 per incident), making real-time intervention critical to managing operating costs. - [08:32](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=512s) With an average of 8.7 days of unplanned downtime per vehicle annually, Traxión faces an annual financial risk exposure exceeding $155 million across its 25,000-unit fleet. - [08:32](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=512s) Enterprise AI adoption hinges on trust and traceability, requiring verified data sources, clear transformation lineage, granular permissions, and immutable audit logs. - [10:55](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=655s) Conversational natural language analytics over enterprise data stores achieves over 97% predictive accuracy in production and slashes decision latency by 65%. - [13:27](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=807s) Autonomous execution engines that automatically schedule maintenance and lock at-risk vehicles in the TMS reduce unplanned downtime by 23% to 48%, saving over $60 million annually. - [15:21](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=921s) Replacing fragmented BI tool sprawl (Power BI, Tableau, Excel) with a unified conversational intelligence layer eliminates up to $470,000 in annual licensing fees while accelerating decision cycles by 29%. - [18:03](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=1083s) Closing the data gap requires focusing on unit-economic decisions over raw dataset ingestion, assigning single decision owners, and objectively measuring decision latency in hours.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=4s) The Disconnected Enterprise and the Vision of Northstar

Modern enterprises operate in a relentless flood of real-time operational data. Every single second, fleet vehicles stream telemetry, financial transactions close, customer service tickets resolve, factory sensors report measurements, and sales teams log closed deals. Yet despite this massive volume of generated information, business reality remains fundamentally fractured across isolated systems. Operations looks at one set of operational metrics, finance evaluates a disconnected ledger, and executive leadership attempts to steer using conflicting versions of the truth.

To bridge this divide, organizations need an intelligence layer capable of ingesting every operational signal, uncovering latent relationships, and converting millions of discrete events into a unified operational picture. Rather than introducing another static reporting tool, this approach—embodied by Northstar—acts as an intelligence layer that reads underlying signals to turn pervasive operational uncertainty into actionable confidence.

[01:37](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=97s) Enterprise Scale and the Reality of the Last Mile Problem

Joshua Bernal, CIO of Grupo Traxión and an enterprise product development founder, oversees one of the largest logistics and mobility operations globally. Managing a fleet of 25,000 vehicles in daily operation creates astronomical data scale: the fleet generates roughly 400 million GPS positions every single day, with every vehicle tracked at five-second intervals. Modern telematics devices stream over 250 distinct data points per second per truck, capturing fine-grained mechanical metrics across engines, fuel consumption, tire wear, hard turns, route deviations, and driver behaviors.

``` +--------------------------------------------------------------------+ | GRUPO TRAXIÓN FLEET TELEMETRY | | 25,000 Active Trucks | 400M GPS Pings/Day | 250 Pts/Sec/Truck | +--------------------------------------------------------------------+ | v +--------------------------------------------------------------------+ | RAW TELEMATICS STREAMS | | Engine Telemetry * Fuel Levels * Tire Wear * Driver Ops | +--------------------------------------------------------------------+ | v +--------------------------------------------------------------------+ | THE LAST MILE TRANSLATION GAP | | 5 Disparate Formats * Reconciliation Delays * Siloed BI | +--------------------------------------------------------------------+ ```

The fundamental bottleneck in enterprise data operations is not a lack of data, but the inability to extract timely business utility from it. Bernal recounts observing a fleet operations team spend three full business days attempting to reconcile an operational metric that should have taken three minutes to verify. The breakdown occurred because the underlying data was fragmented across five incompatible formats across disparate systems with no shared enterprise consensus. As Bernal emphasizes, this failure is not a technical ingestion problem, but a classic "last mile problem"—the profound operational disconnect between possessing raw enterprise data and translating it into decisive operational execution.

- Fleet scale: 25,000 commercial vehicles tracked continuously every 5 seconds. - Data volume: ~400 million daily GPS coordinates and 250 real-time telemetry variables per second per vehicle. - Operational friction: 3 days spent manually reconciling basic fleet metrics across 5 disjointed software systems.

[03:31](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=211s) The Dashboard Graveyard and the Evolution from Insight to Action

A pervasive paradox across enterprise IT is that despite aggressive tooling investments, active analytics utilization remains remarkably low. Only 25% of enterprise employees actively interact with the business intelligence and analytics tools their organizations procure. Consequently, between 60% and 80% of corporate dashboards fall into disuse within months of being deployed—a phenomenon Bernal describes as the "dashboard graveyard." Although 87% of enterprises have prioritized analytics adoption over the past decade, actual engagement metrics have remained largely stagnant for seven years because the core challenge has migrated from data access to semantic translation.

``` +--------+ +-------------+ +---------+ | DATA | ---> | INFORMATION | ---> | INSIGHT | +--------+ +-------------+ +---------+ | [ THE OPERATIONAL GAP ] (Projects Stall Here) | +--------+ +-------------+ +----------------+ | ACTION | <--- | DECISION | <--- | RECOMMENDATION | +--------+ +-------------+ +----------------+ ```

Enterprise analytics has progressed across three broad eras since 2010: shifting from static, immovable reporting dashboards to interactive self-service BI, and finally moving toward conversational, action-oriented intelligence where systems act directly on plain-language inquiries. The critical breakdown in existing business intelligence occurs along the cognitive value chain connecting data to action. Traditional BI projects stall at the insight stage. As Bernal notes, "An insight is a fact about the past. A decision is a commitment about what happened next. And that's the crossing most projects never make." When operators open a standard dashboard, they are forced to locate metrics, interpret structural schemas, correlate multiple disconnected sources, and determine an action plan entirely on their own.

- Analytics adoption gap: Only 25% of staff actively use procured enterprise BI solutions. - Dashboard churn: 60% to 80% of analytical dashboards are abandoned within months of deployment. - Decision paralysis: Dashboards merely present retrospective historical figures without providing context-aware operational guidance.

[07:11](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=431s) Fleet Economics: The True Cost of Unplanned Downtime

In industrial transportation, the financial delta between reactive and proactive operational decisions is substantial. A single unplanned highway breakdown incurs approximately $2,000 in direct expenses once emergency roadside repairs, vehicle towing, and driver downtime are factored in. Performing that identical repair as part of a scheduled, planned maintenance cycle costs three to five times less. Fleet maintenance expenditures account for roughly 9% of total fleet operational budgets—amounting to approximately $0.20 per operating mile according to international transportation benchmarks.

Detecting mechanical risk inside a data pipeline is useless unless an organization possesses the automated mechanisms to act upon that signal instantly. Bridging this operational gap requires transforming the user interface entirely. Instead of navigating complex database schemas, business users describe their operational intent in plain language ("Ask"), the AI layer identifies and reconciles the necessary underlying data stores ("Discover"), and specialized autonomous agents deliver contextual answers paired with direct operational workflows ("Act").

- Unplanned breakdown cost: ~$2,000 per incident including tow, repair, and driver downtime. - Planned maintenance savings: Proactive scheduled repairs cost 3x to 5x less than roadside failures. - Operating cost impact: Maintenance represents ~9% of total fleet operational expenditures ($0.20/mile).

[08:32](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=512s) Fleet Exposure and the Enterprise Trust Architecture

The financial exposure of fleet downtime becomes massive when evaluated across Traxión's operating footprint. The company's 25,000-vehicle fleet spans three primary business divisions: 10,000 dedicated passenger mobility units (the most service-sensitive division due to transporting personnel), alongside extensive operations in general freight and dry goods logistics. Across the fleet, unplanned vehicle downtime averages 8.7 days per unit per year. Across 25,000 assets, this represents an annual financial risk exposure exceeding $155 million in lost operational productivity and repair overhead.

``` +--------------------------------------------------------------------+ | ENTERPRISE TRUST FRAMEWORK | +--------------------------------------------------------------------+ | 1. SOURCE -> Explicit verification of original data stores | | 2. LINEAGE -> Comprehensive tracking of every transformation | | 3. PERMISSIONS -> Fine-grained role-based view and action rights | | 4. AUDIT TRAIL -> Immutable logging of all queries and updates | +--------------------------------------------------------------------+ | "You can trust it because you can trace it." | +--------------------------------------------------------------------+ ```

AI initiatives in enterprise logistics frequently fail during the pilot phase not because the underlying machine learning models are mathematically flawed, but because operational managers lack the transparency required to trust the outputs. Establishing operational trust requires strict visibility into the data supply chain: tracing the exact data source, mapping the transformation lineage, enforcing role-based permissions, and maintaining an immutable audit log. Bernal summarizes this governance imperative cleanly: "You can trust it because you can trace it."

- Fleet distribution: 10,000 passenger mobility vehicles, supplemented by large dry van and general freight divisions. - Fleet downtime: Average of 8.7 days of unplanned vehicle downtime annually per asset. - Financial risk: Over $155 million in annual downtime exposure across the total enterprise fleet.

[10:55](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=655s) Northstar's Learning Curve and Natural Language Analytics

Northstar implements this decision architecture by interfacing directly with enterprise data lakes, data marts, and warehouse structures. When deployed, the system begins training on raw enterprise schemas without manual context injection, organizing and understanding relationships iteratively. Day-one predictive accuracy begins at a baseline of roughly 48% and systematically increases by approximately two percentage points per training iteration.

``` +--------------------------------------------------------------------+ | NORTHSTAR MODEL ACCURACY TRAJECTORY | +--------------------------------------------------------------------+ | Day 1 (Initial Ingestion): [==== ] 48% | | Iteration 5-6: Operational Baseline Reached | | Iteration 22: [=================== ] 92% | | Production Maturity: [==================== ] >97% | +--------------------------------------------------------------------+ | Target Predictions: Churn, Operator Attrition, Component Failure | +--------------------------------------------------------------------+ ```

By iteration 5 or 6, the system achieves an operational baseline, and by iteration 22, it reaches roughly 92% predictive accuracy across high-impact business variables, including customer churn, driver/operator attrition, and historical component repair cycles. In production environments, accuracy surpasses 97%. Rather than configuring specialized dashboards, operational managers ask plain-language questions as if speaking to an analyst colleague. Direct conversational querying against underlying enterprise data stores slashes overall decision latency—the critical duration between identifying an insight and executing a business decision—by an estimated 65%.

- Baseline trajectory: 48% accuracy on Day 1, climbing ~2% per iteration to ~92% at iteration 22, and exceeding 97% in production. - Iterative ramp: Requires 5 to 6 training iterations to achieve an operational baseline against enterprise databases. - Efficiency gains: Natural language querying reduces decision latency from insight to action by 65%.

[13:27](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=807s) Autonomous Execution: From Telematics to Automated Dispatch

The true power of an operational intelligence engine emerges when analytics move from passive alerts to automated workflow execution. Telemetry streams and mechanical fault events continuously feed a rolling health score computed for every individual truck in the fleet. When an asset's telemetry breaches a designated risk threshold, Northstar avoids merely outputting a dashboard flag; it takes programmatic control of the dispatch workflow.

``` +------------------+ +-------------------+ +--------------------+ | Telemetry Stream | ---> | Risk Threshold | ---> | Autonomous Booking | | & GPS Pings | | Breached in Model | | & TMS Asset Lock | +------------------+ +-------------------+ +--------------------+ | v +-------------------------------------------------------------------------+ | OPERATIONAL OUTCOME | | * 23% to 48% reduction in unplanned fleet downtime | | * >$60 million saved annually in breakdown avoidance | +-------------------------------------------------------------------------+ ```

The system automatically schedules an inspection with the appropriate maintenance facility and simultaneously places an operational hold on the asset inside the Transportation Management System (TMS), preventing dispatchers from assigning the vehicle to a new route prior to service. Proactively taking high-risk units off the road before mechanical failures occur on the highway yields an estimated 23% to 48% reduction in unplanned fleet downtime. For Grupo Traxión, this single automated closed-loop intervention saves over $60 million annually in avoided breakdown costs and preserved cargo schedules.

- Automated dispatch integration: Autonomously books service appointments and blocks TMS route assignments before breakdowns occur. - Fleet downtime reduction: Drives a 23% to 48% drop in unscheduled maintenance events. - Bottom-line savings: Generates over $60 million in annual cost avoidance across fleet operations.

[15:21](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=921s) Consolidating Tool Sprawl and Licensing Overhead

Enterprise reporting environments frequently suffer from severe data sprawl, where identical operational metrics live across five or more conflicting mediums: Power BI, Tableau, Excel spreadsheets, email chains, and disconnected legacy ERP reports. Supporting this sprawling infrastructure incurs heavy direct software overhead, often costing upwards of $470,000 annually across 2,800 licensed corporate BI seats.

``` BEFORE: FRAGMENTED TOOL SPRAWL [Power BI] + [Tableau] + [Spreadsheets] + [Legacy Reports] + [Emails] * 2,800 Seats * $470,000/yr Licensing * Conflicting Metric Versions

AFTER: UNIFIED CONVERSATIONAL INTELLIGENCE (NORTHSTAR) [ Single Conversational Board / Real-Time Dynamic Dashboards ] * $0 Per-Seat BI Licensing * 29% Faster Decision Cycles * 1 Source ```

Northstar consolidates these fractured analytics systems into a single enterprise conversational interface trained strictly on internal company data without exposing intellectual property externally. By replacing per-seat static BI licenses with real-time, dynamically generated agent workspaces, organizations can eliminate up to $470,000 in recurring annual software licensing costs while simultaneously accelerating core business decision cycles by 29% in time-sensitive logistics environments.

- Tool consolidation: Replaces disconnected instances of Power BI, Tableau, spreadsheets, and email reports with a unified interface. - Direct cost reduction: Eliminates approximately $470,000 annually in per-seat software licensing across ~2,800 enterprise seats. - Organizational speed: Achieves a 29% faster end-to-end operational decision cycle across time-critical business units.

[18:03](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=1083s) The Monday Morning Framework for Closing the Last Mile

To successfully bridge the enterprise data gap, leaders must change how they approach technology implementation. Bernal outlines a pragmatic framework for teams to execute immediately:

- **Focus on Decisions, Not Datasets:** Prioritize high-value operational decisions that directly impact unit economics rather than undertaking massive, unfocused data lake ingestions. - **Establish Single Decision Ownership:** Clearly identify the specific operational role and individual accountable for acting on each decision type. - **Quantify Latency in Hours, Not Opinions:** Objectively measure the end-to-end friction and delay required to move from data ingestion to operational execution using real time metrics rather than subjective team feedback. - **Ground Context and Traceability:** Connect underlying source systems while enforcing end-to-end data lineage and governance so operators have complete confidence in automated recommendations.

Ultimately, the last mile of data is never about the physical distance across networks. It is defined entirely by what happens in the final few feet: whether high-fidelity operational intelligence successfully reaches the human or automated system responsible for making the decisive business choice.

## Notable quotes > [03:31](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=211s) An insight is a fact about the past. A decision is a commitment about what happened next. And that's the crossing most projects never make. — Joshua Bernal

> [08:32](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=512s) You can trust it because you can trace it. — Joshua Bernal

## Mentioned Joshua Bernal, Grupo Traxión, Northstar, Ai4, Power BI, Tableau, Excel

## Video description Joshua Bernal - CIO, Grupo Traxion

Organizations have invested heavily in collecting data, building dashboards, and deploying AI, yet many still struggle with the final step: turning information into confident operational decisions.

This session explores the "last mile of data"—the gap between enterprise data and real business action.

Through real world logistics and transportation examples, attendees will discover how AI native Decision Intelligence platforms, conversational analytics, and specialized AI agents are transforming the way organizations interact with their data. Instead of waiting for reports or navigating complex dashboards, business users can simply ask questions, receive contextual insights, and make faster, more informed decisions.

The session also explores the role of data governance, trusted enterprise knowledge, and AI agents in creating a new generation of operational intelligence.

[Source](https://www.youtube.com/watch?v=5XhFtrK_yC8)

中文简报

Grupo Traxion实现卡车调度自动化以削减6000万美元停运损失 Grupo Traxión通过以自主AI代理取代被动仪表板，在车辆于公路上抛锚前将其锁定停运，从而将车队非计划停运时间减少最高达48%，并避免每年6000万美元的损失。 Joshua Bernal概述了管理Grupo Traxión旗下25,000辆车车队所面临的巨大数据挑战，并指出传统BI工具往往局限于事后洞察，导致多达80%的企业仪表板被弃用。为解决这一“最后一公里”的脱节问题，该组织部署了对话式决策智能层Northstar，在零部件故障和流失率等关键指标上实现了超过97%的实际预测准确率。该平台推动分析工作从被动告警转向闭环自主执行，例如主动在调度系统中锁定有风险的车辆并安排维护。这种自动化干预使车队的非计划停运时间减少了23%至48%，每年避免了超过6000万美元的抛锚开支，并通过整合零散的分析工具节省了约47万美元的软件许可费用。现代企业产生海量的实时运营数据流，但跨部门孤岛阻碍了各团队基于单一事实源开展协作。 (0:04) 企业的核心瓶颈在于“最后一公里问题”——即拥有原始数据（例如25,000辆卡车每天产生的4亿次GPS定位信号）与将其转化为果断的运营执行之间的脱节。 (1:37) 60%至80%的企业仪表板在数月内被废弃，因为传统BI工具呈现的是回顾性事实，而非具备情境感知的运营指导。 (3:31) 前瞻性的定期维护成本比公路上意外抛锚（每次事故花费2,000美元）低3到5倍，这使得实时干预对于控制运营成本至关重要。 (7:11) 由于每辆车年均非计划停运时间为8.7天，Traxión庞大的25,000辆车队每年面临超过1.55亿美元的财务风险敞口。 (8:32)

Reading summary

Grupo Traxión CIO Joshua Bernal argues that enterprise analytics initiatives fail not from insufficient data collection, but at the "last mile" where analytical insights must translate into decisive operational execution. Addressing this gap requires shifting from static, retrospective dashboards to conversational intelligence layers and autonomous, closed-loop workflows. The Operational Last Mile and Dashboard Obsolescence (1:37) Modern industrial operations produce massive telemetry streams, exemplified by Grupo Traxión's 25,000-vehicle fleet generating roughly 400 million daily GPS pings and streaming over 250 data points per second per truck across engine performance, fuel use, and driver behavior. However, Bernal notes that raw volume frequently encounters a "last mile problem": operational teams can spend days reconciling basic metrics simply because underlying data is fragmented across incompatible software formats and disconnected departmental silos. Traditional business intelligence infrastructure has failed to solve this disconnect. Despite 87% of enterprises prioritizing analytics investments over the past decade, active employee engagement with BI tools has stagnated at 25% for seven years, and between 60% and 80% of corporate dashboards fall into disuse within months of launch. Bernal highlights that dashboards stall at reporting retrospective facts about the past rather than generating forward-looking operational commitments, leaving personnel to correlate disconnected schemas and determine actions without automated guidance. Fleet Breakdown Economics and the Trust Architecture (7:11) The economic penalty of delayed operational intervention in industrial transit is severe. In commercial trucking, an unplanned highway breakdown costs approximately $2,000 when factoring in towing, roadside repairs, and driver downtime—three to five times more expensive than planned maintenance. Across Traxión's 25,000 assets, which include 10,000 service-critical passenger mobility units alongside dry freight fleets, unplanned downtime averages 8.7 days per vehicle annually, creating over $155 million in combined operational exposure. Despite the clear incentives for predictive maintenance, Bernal observes that predictive AI models frequently stall during enterprise pilots because operational managers distrust opaque algorithms. To ensure operational adoption, organizations must implement a verifiable trust architecture comprising four controls: source verification, transformation lineage tracking, role-based access management, and immutable audit logs. Operational teams will only rely on automated recommendations if the underlying data supply chain can be fully traced and audited. Conversational Intelligence and Autonomous Dispatch (10:55) To bridge the latency between data ingestion and decision-making, Traxión deployed Northstar, an intelligence platform that interfaces directly with underlying enterprise data stores. The system achieves a baseline operational accuracy of roughly 48% on day one and systematically improves by roughly two percentage points per training iteration, reaching an operational baseline in five to six iterations, approximately 92% accuracy by iteration 22, and over 97% in production. Allowing operational managers to query data using natural language rather than static BI dashboards reduces decision latency by an estimated 65%. Beyond conversational query response, the architecture automates operational dispatch. Northstar continuously tracks mechanical fault streams against rolling risk thresholds; when an asset breaches safety limits, the system autonomously books an inspection slot and triggers an operational lock in the Transportation Management System to halt route assignment. Bernal reports that this automated closed-loop intervention reduces unplanned fleet downtime by 23% to 48%, generating more than $60 million in annual cost avoidance. Consolidating Tool Sprawl with Decision-First Governance (15:21) Deploying a unified intelligence layer also resolves enterprise software sprawl. In legacy setups, identical operational metrics are often duplicated across Power BI, Tableau, spreadsheets, and ERP reports, generating conflicting versions of truth and incurring up to $470,000 annually in per-seat licensing across 2,800 users. Replacing fragmented seats with a consolidated, internal conversational workspace eliminates these software overheads while speeding up core operational decision cycles by 29%. To institutionalize these gains, Bernal recommends a pragmatic implementation framework focused on decisions rather than dataset accumulation. Organizations must prioritize specific decisions that impact unit economics, assign single operational ownership to each decision type, objectively measure ingestion-to-execution latency in hours rather than subjective sentiment, and enforce context traceability across every underlying data source. The source demonstrates that closing the data gap in logistics relies on linking telemetry directly to automated execution workflows rather than building more reporting layers. However, long-term success remains contingent on whether enterprises can maintain end-to-end data lineage and establish clear human accountability for automated recommendations.

精炼阅读摘要

Grupo Traxión 首席信息官 Joshua Bernal 指出，现代企业的数据瓶颈不在于数据采集规模，而在于无法将原始数据快速转化为业务执行；通过引入类似 Northstar 的决策与自主智能层，企业能够打通“最后一公里”，将静态洞察直接转变为自动化运营动作。数据最后一公里的瓶颈与看板坟场 (1:37) Grupo Traxión 运营着包含 25,000 辆商用车的庞大车队，每辆车每 5 秒回传一次数据，每天生成约 4 亿个 GPS 定位坐标，且每辆卡车每秒产生超过 250 个机械指标数据点。然而，面对如此庞大的数据吞吐量，业务端却经常面临严重延迟。Bernal 举例指出，由于底层数据分散在 5 种不兼容的系统格式中，且缺乏全企业共识，车队运营团队曾耗费整整 3 个工作日去人工核对一项原本只需 3 分钟即可验证的基础指标。这种脱节表明核心障碍并非技术摄取能力不足，而是数据与业务执行之间的“最后一公里”断层。这一断层直接反映在传统商业智能（BI）工具的低利用率上。尽管过去十年中有 87% 的企业将分析工具列为优先事项，但仅有 25% 的员工会主动使用这些工具，且活跃度已停滞七年。其结果是，60% 至 80% 的企业看板在部署数月后便被废弃，沦为“看板坟场”。Bernal 强调，传统 BI 往往停留在提供回顾性历史事实的阶段，迫使操作人员自行寻找指标、理解复杂模式并拼凑解决方案，无法跨越从静态“洞察”迈向具体“决策”与“行动”的关键鸿沟。非计划停机的经济损失与信任架构 (7:11) 在工业运输领域，被动处置与预防性维护的财务差异极为悬殊。单次意外公路抛锚由于包含现场紧急抢修、拖车及司机误工，直接成本约为 2,000 美元；而如果纳入计划性维护周期，相同维修的支出可降低至三分之一至五分之一。车队维护支出约占整体运营预算的 9%（国际标准约为每英里 0.20 美元）。在 Traxión 涵盖 10,000 辆客运车及大量干线货车的庞大资产池中，单车年均非计划停机时间达到 8.7 天，整体年化财务风险敞口超过 1.55 亿美元。针对物流 AI 项目常在试点阶段停滞的现象，Bernal 指出症结不在于算法模型的数学缺陷，而在于业务主管对模型输出缺乏信任。为建立操作信任，企业必须推行“可追踪即可信任”的数据治理框架，包含对底层数据源的显式验证、端到端的数据血缘追踪、细粒度的角色访问控制，以及防篡改的审计日志，从而为基于算法的自动化决策提供制度保障。自然语言交互与闭环自主调度执行 (10:55) 智能决策系统 Northstar 直接对接企业数据湖与数据集市，通过在原始架构上迭代学习构建业务关联。其第一天的预测准确率基线约为 48%，随后以每次迭代约 2 个百分点的速度提升，在第 5 至 6 次迭代达到运营基线，并在第 22 次迭代时在客户流失、司机离职及零部件维修周期等预测上达到约 92% 的准确率，生产成熟期更超过 97%。业务人员通过自然语言进行人机交互，将从获取洞察到执行决策的延迟缩短了约 65%。该架构的核心突破在于将分析告警直接转化为闭环工作流。系统结合遥测流与机械故障事件动态计算单车健康分，一旦数据触发设定的风险阈值，系统不再仅发送被动警报，而是自主接管调度：一方面自动向对应维保中心预约检查，另一方面在运输管理系统（TMS）中对高风险车辆施加操作锁定，防止调度员为其指派新路线。通过在车辆发生公路故障前提前干预，车队非计划停机时间减少了 23% 至 48%，为 Grupo Traxión 每年避免了超过 6,000 万美元的故障损失。治理工具蔓延与落地实施路径 (15:21) 传统企业报告环境普遍存在工具蔓延问题，相同指标常并存于 Power BI、Tableau、Excel 电子表格与 ERP 报告中。在 2,800 个企业 BI 席位下，每年的按席位软件许可成本高达约 470,000 美元。通过将分散的报表整合进单一的会话式智能工作区，企业不仅能在保护内部知识产权的前提下完全消除这笔周期性许可支出，还能在对时间高度敏感的物流环境中将核心决策周期加快 29%。为推动数据最后一公里的落地，Bernal 提出了“周一早晨框架”：首先，企业应优先聚焦直接影响单位经济效益的高价值运营决策，而非展开宽泛且无重点的数据湖建设项目；其次，明确单项决策的责任角色与负责人；第三，用以小时计的端到端耗时等客观指标量化业务延迟，取代主观反馈；最后，严格维护底层数据系统的上下文关联与血缘治理，确保操作人员对自动化决策建议保有充分信心。数据最后一公里的本质不是物理网络跨度，而是信息能否在决策前的关键环节转化为行动。该方案展示了通过自主智能调度与自然语言交互消解系统孤岛的可行性，但模型高度依赖底层遥测传感器的高频覆盖与数据血缘机制的完整性，在数字化基础较弱或流程缺乏标准化约束的场景中，其落地成效仍有待检验。

Detailed notes — zh

## Background 本内容由 YouTube 频道 Ai4 于 2026 年 8 月 23 日发布，主讲人为 Grupo Traxion 的首席信息官（CIO）Joshua Bernal。视频页面简介指出，本次演讲聚焦于企业数据与实际业务行动之间的“数据最后一公里”痛点，通过真实物流案例探讨如何利用 AI 原生决策智能平台、对话式分析与专有 AI Agent 赋能一线决策。 ## TL;DR Grupo Traxion 每日管理 25,000 辆卡车并产生数亿条实时数据，但以往跨系统的碎片化导致数据核对极为缓慢，且传统 BI 仪表板面临高达 60%–80% 的闲置废弃率。为此，公司引入了决策智能平台 Northstar，不仅将决策延迟缩短 65%，还将预测准确率提升至 97% 以上。通过打通从预警到自动预约保养并在 TMS 中锁定车辆的闭环，Traxion 将未计划宕机时间减少了 23% 至 48%，每年成功避免超过 6,000 万美元的损失。演讲最后提炼出一套以“关键决策”而非“数据集”为核心的企业 AI 落地四步法。 ## Atmosphere & style 整场演讲务实干练且直击痛点，Joshua Bernal 兼具企业高管与产品开发者的双重视角，既有宏观的架构思考，又有精确到单车损失的硬核财务数字。演讲叙事层层递进，从剖析“仪表板坟墓”与低效核对的行业共性困局，转向以数据追溯和自动闭环重构信任的破局方案，情绪从对海量沉淀数据的焦虑转化为对 AI 决策价值落地的笃定与振奋。观看过程节奏紧凑、充满实战说服力，是一场极具启发性的企业数字化与 AI 转型实录。 ## Key takeaways - [01:37](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=97s) 企业并不缺少数据，但面临严重的“最后一公里”难题：海量信息分散在互不相连的业务系统中，导致耗时数天跨系统核对、无法快速形成决策。 - [01:37](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=97s) 以 Grupo Traxion 为例，车队每日管理 25,000 辆卡车、生成 4 亿条 GPS 定位数据且每秒捕获 250+ 数据点，庞大数据流对转化应用提出了极高要求。 - [03:31](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=211s) 传统 BI 陷入采用率停滞困境：仅 25% 员工积极使用分析工具，60%–80% 的仪表板在上线数月后被废弃，项目普遍卡在“过去洞察”而无法迈向“未来决策与行动”。 - [07:11](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=431s) 商业分析正从静态与交互式仪表板跃迁至 AI 决策智能时代，新范式让业务人员通过自然语言直接提问，由专业 AI Agent 提供上下文并驱动行动。 - [08:32](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=512s) 车队维护占运营总成本的 9%，未计划故障损失比计划维护高 3 至 5 倍；Traxion 单车年均 8.7 天的未计划宕机曾带来高达 1.55 亿美元的年化财务风险暴露。 - [09:53](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=593s) 企业级 AI 落地的核心在于信任，而信任来自对数据来源（Source）、数据血缘（Lineage）、权限控制（Permissions）与审计轨迹（Audit Trail）的全链路可追溯性。 - [10:55](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=655s) Northstar 智能层基于真实企业数据展开自适应学习，预测准确率从初始的 48% 快速迭代至 97% 以上，并将洞察到行动的决策延迟缩短约 65%。 - [13:27](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=807s) 实现决策自动化闭环：系统在健康评分异常时自动预约保养并在 TMS 中锁定车辆，使未计划宕机减少 23%–48%，每年避免超 6,000 万美元损失。 - [13:27](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=807s) 单一对话界面取代了大量无人使用的报表与系统，消除每年超 47 万美元的 BI 软件许可浪费，并将业务整体决策周期加快 29%。 - [18:03](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=1083s) 落地 AI 决策智能的四步法则：聚焦关键决策而非单纯扩充数据集、锁定决策责任人、以“小时”量化决策阻力、在连接数据源时保持业务上下文与可追溯性。 ## Detailed walkthrough 以下是根据视频演讲内容整理的详细中文逐段分析与记录： ### [00:04](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=4s) 介绍与数据的爆发增长演讲者 Joshua Bernal（Grupo Traxion 的 CIO）首先向观众致谢，并简要介绍了 Grupo Traxion——这不仅是拉丁美洲最大的物流公司之一，在全球范围内也名列前茅。随后，他播放了一段视频，展示企业在日常运营中所产生的海量数据及其碎片化现状。企业每时每刻都在产生海量信息：从车队车辆传输的远程测电设备数据、闭合的财务交易、客户互动、运营过程、传感器读数、生产制造，到发送的电子邮件、交付的服务以及完成的销售。然而，这些信息往往处于碎片化状态，分散在几十个互不相连的系统中。财务部门看到的是一方面，运营部门看到的是另一方面，而客户成功部门、销售部门和管理层各自基于不同的“现实版本”开展工作。针对这种痛点，演讲提出了“Northstar”概念——一个能够理解全局的单一智能层（Intelligence Layer）。它并非另一个仪表板或数据工具，而是一个能够读取每种信号、发现隐性关系、并将数以百万计的业务事件转化为统一且连贯理解的智能网络，从而将不确定性转化为决策信心。 - 车辆远程测电设备（Telematics）持续传输数据 - 跨部门（财务、运营、销售、客户成功）的数据严重碎片化与孤岛化 - Northstar 的定位：打破孤岛的统一智能层，而非传统的 BI 仪表板 ### [01:37](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=97s) 物理海量数据与“最后一公里”难题 Joshua Bernal 介绍了自己的背景，他不仅担任 Grupo Traxion 的 CIO，还通过自己的公司从事企业级产品开发。接着，他用一组具体数字展示了 Traxion 所面临的极高运营复杂度与数据规模：公司每天管理着 25,000 辆卡车，每天生成大约 4 亿条 GPS 位置数据，整个车队每 5 秒就被追踪一次。现代远程测电设备每辆车每秒可捕获超过 250 个数据点（涵盖发动机状态、油耗、轮胎、驾驶行为以及急刹车、急转弯等各种事件）。然而，“拥有数据”与“使用数据”完全是两码事。Joshua 回忆起几年前的一个场景：车队运营团队花费了整整三天时间去核对一个本来只需 3 分钟就能搞定的数字。这不是因为缺少数据，而是因为数据分散在 5 个不同格式的系统里，且没有任何人能达成一致。他强调：“这当然不是技术问题，而是数据的‘最后一公里’问题（Last Mile Problem）。” ```mermaid flowchart TD accTitle: 物流数据生成与处理流程 subgraph DataSources["数据源生成"] A["25,000 辆卡车车队"] --> B["每天 4 亿条 GPS 数据"] A --> C["每秒 250+ 数据点/车 (发动机/油耗/轮胎/驾驶行为)"] end subgraph LastMileIssue["最后一公里瓶颈"] B & C --> D["分散在 5 个不兼容系统中"] D --> E["数据核查耗时 3 天 (本应只需 3 分钟)"] end E --> F["决策延迟与商业损失"] ``` - **车队规模**：每日管理 25,000 辆卡车 - **数据采集频率**：车队每 5 秒追踪一次，远程测电设备每秒采集 250+ 数据点，每日产生 4 亿条 GPS 定位数据 - **核心痛点**：数据格式不统一、系统割裂，导致极其低效的跨系统核对 ### [03:31](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=211s) 数据分析工具的困局与三代演进 Joshua 提出了一个令人震惊的事实：企业面临的从来不是“数据短缺”。目前，即使企业支付了昂贵的软件许可费用，也只有 25% 的员工在积极使用商业智能（BI）和业务分析工具；60% 到 80% 的仪表板在上线几个月内就被废弃，变成了无人问津的“仪表板坟墓”（Dashboard Graveyard）。尽管 87% 的组织表示在过去 5 到 10 年里增加了分析投入，但整体的用户采用率在 7 年内几乎毫无增长。瓶颈早已不是“如何获取数据”，而是“如何转化数据”。商业分析经历了三个时代的演进： 1. **2010 年左右的传统仪表板**：展示静态信息，数据无法流动，存在严重的“最后一公里”隔离。 2. **过去 10 年的交互式仪表板与自助式 BI**：虽然实现了更高的交互性，但依然止步于展现数据。 3. **AI 驱动的决策智能时代**：用户无需直接面对复杂仪表板，只需自然提问，系统即可给出解答并直接采取行动。 Joshua 强调了从数据到行动的全链路：数据（Data）$\rightarrow$ 信息（Information）$\rightarrow$ 洞察（Insight）$\rightarrow$ 建议（Recommendation）$\rightarrow$ 决策（Decision）$\rightarrow$ 行动（Action）。当今绝大多数项目都卡在了“洞察”这一步。然而，“洞察是对过去事实的陈述，而决策是对未来行动的承诺。”数字本身并不是决策，仪表板展示了数字，却无法替人做出决策。 - **利用率低下**：仅 25% 的员工积极使用已采购的分析工具；60%–80% 的仪表板上线后被废弃 - **采用率停滞**：过去 7 年间分析工具的普及率几乎未变 - **决策断层**：项目普遍停留在“洞察（过去）”阶段，无法跨越到“决策与行动（未来）” ### [07:11](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=431s) 车队维护场景与交互界面的革新为了直观展示数据与决策之间的鸿沟，Joshua 以车队物流的实际成本为例进行了算账： - **故障成本差距**：一次未计划的卡车故障拖车、维修加上司机停工成本，单次平均损失约 2,000 美元；而如果按计划进行预防性维护，相同的维修成本要便宜 3 到 5 倍。 - **运营占比**：根据国际组织统计，仅维护一项就占到了整个车队总运营成本的 9%（相当于每英里消耗 0.20 美元）。发现风险与采取行动是两码事，而这其中的差距就是资金流失的地方。要弥合这一差距，必须彻底改变人机交互界面： 1. **询问（Ask）**：用户直接用平实的自然语言提出业务需求。 2. **探索（Discover）**：系统自动识别并定位包含答案的数据源。 3. **行动（Act）**：专门的 AI Agent（智能体）提供带有业务上下文的解释与解决方案，而非仅扔出一个图表。 - **未计划故障损失**：单次约 2,000 美元，成本比计划维护高出 3 至 5 倍 - **维护成本占比**：占车队运营总成本的 9%（或 0.20 美元/英里） - **新交互范式**：自然语言提问 $\rightarrow$ 系统自主定位数据 $\rightarrow$ AI Agent 上下文解答并执行行动 ### [08:32](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=512s) Grupo Traxion 的业务规模与宕机风险 Grupo Traxion 是全球第二大同类物流企业（在墨西哥及拉丁美洲居首）。Joshua 详细拆解了 Traxion 的车队构成与未计划宕机所带来的巨大财务风险： | 业务线分类 | 车辆数量 | 业务特点与敏感度 | |:--- |:--- |:--- | | **客运出行 (Passenger Mobility)** | 10,000 辆 | 最敏感的业务线（涉及人员运输与安全） | | **普通货运与干货运输 (Freight & Dry Goods)** | 15,000 辆 | 支撑基础物流网络 | | **总计** | **25,000 辆** | - | 在过去，每辆车年均未计划宕机时间高达 **8.7 天**。将这一数字乘以整个 25,000 辆车的庞大车队，公司每年因未计划宕机带来的潜在财务风险暴露高达 **1.55 亿美元**。这是经过审计的内部真实数据，证明了在“最后一公里”创造真正业务价值而非仅做仪表板的紧迫性。 - **车队结构**：10,000 辆客运车辆 + 15,000 辆货运/干货车辆 - **宕机指标**：单车年均未计划宕机 8.7 天 - **年化风险暴露**：超过 1.55 亿美元 ### [09:53](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=593s) 信任构建机制：透明度与数据血缘大多数 AI 试点项目之所以夭折，并非因为 AI 模型不够准确，而是因为没人敢足够信任它并付诸行动。信任并非来自空洞的信心，而是来自全面的“可见性（Visibility）”： - **来源（Source）**：明确数字究竟来自于哪里的原始数据。 - **数据血缘（Lineage）**：追踪并记录数据在传输与转换过程中的每一步变化。 - **权限（Permissions）**：清晰界定谁有权查看该数据并据此做出行动。 - **审计轨迹（Audit Trail）**：对每一次数据访问和修改进行日志记录。 Joshua 强调：“你能信任它，是因为你能追溯它（You can trust it because you can trace it）。” ```mermaid flowchart LR accTitle: 企业级 AI 信任四大支柱 subgraph TrustFramework["企业级 AI 信任框架"] A["数据来源 (Source)"] --> B["数据血缘 (Lineage)"] B --> C["权限控制 (Permissions)"] C --> D["审计轨迹 (Audit Trail)"] end D --> E["可追溯的 AI 决策 (Trusted AI Decision)"] ``` - **试点夭折根源**：缺乏信任导致不敢将决策付诸实践 - **信任四大要素**：数据来源、数据血缘、权限控制、审计轨迹 ### [10:55](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=655s) Northstar 平台机制与渐进式学习 Joshua 现场展示了 Northstar 平台的运作机制。他澄清这只是原理的演示，而非推销。在部署初期，Northstar 的初始预测准确率从 48% 开始，但在连接企业数据库后，它无需人工标注上下文即可基于真实数据展开训练，并围绕企业自身的业务逻辑进行自适应学习。系统通常需要 **5 到 6 次迭代** 即可达到基础的运营标准。迭代所需时间取决于 Data Lake、Data Mart 或 Data Warehouse 中的数据量。 - **初始准确率**：第 1 天基于真实数据，预测准确率为 48%。 - **迭代提升**：每次迭代大约提升 2 个百分点；到了第 22 次迭代时，预测准确率可达 **92%** 左右。 - **关键预测领域**：涵盖客户流失（Customer Churn）、司离职率（Operator Attrition）以及基于历史记录的故障维修预测。 - **当前水平**：随着数据源源不断进入，目前系统的迭代准确率已提升至 **97% 以上**。不仅如此，通过自然语言进行查询，将“洞察”到“行动”之间的决策延迟（Decision Latency）降低了约 **65%**。 - **准确率演变**：48%（Day 1）$\rightarrow$ 92%（第 22 次迭代）$\rightarrow$ 97%+（当前实际运行） - **决策延迟减少**：缩短 65% 的决策响应时间 ### [13:27](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=807s) 自动化闭环与极致的经济效益 Northstar 的核心能力在于它不只停留在警报阶段，而是能直接打通业务流程：系统会结合每一个 GPS 信号和机械事件，为每辆车计算滚动健康评分（Rolling Health Score）。当某辆车的风险值超出阈值时： 1. Northstar 不仅仅抛出红色预警，而是**自动预约保养维修**。 2. 自动在 TMS（运输管理系统）中**锁定该车辆**，防止系统在故障发生前为其分配新的运输任务。在 Traxion 的车队中应用此机制后，未计划宕机时间减少了 **23% 至 48%**。转化为财务收益后，公司每年可避免超过 **6,000 万美元** 的损失，这在行业内树立了标杆。 ```mermaid flowchart TD accTitle: Northstar 自动闭环维护流程 subgraph SignalProcessing["实时信号处理"] A["GPS 信号 & 车辆机械事件"] --> B["计算车辆滚动健康评分"] end subgraph AutoAction["自动化决策与执行"] B -- "超越风险阈值" --> C["Northstar 自动预约保养时间"] C --> D["TMS 系统自动锁定车辆 (禁止分配新行程)"] end subgraph BusinessImpact["商业价值"] D --> E["未计划宕机减少 23%-48%"] E --> F["每年避免 6000万+ 美元损失"] end ``` 同时，针对软件许可费用的浪费，过去同样的数字散落在 Power BI、Tableau、电子表格、邮件和旧系统中，仅 2,800 个席位的软件许可费用每年就高达 **470,400 美元**（其中 80% 的报表无人使用）。通过 Northstar 的单一对话界面，软件许可费用被彻底清理，CEO 为之振奋，同时整个业务的决策周期加快了 **29%**。 - **宕机减少**：未计划宕机降低 23%–48% - **直接经济收益**：每年节省超过 6,000 万美元 - **软件许可节省**：消除传统 BI 软件 2,800 个席位共计 470,400 美元/年的授权费 - **决策周期**：整体加快 29% ### [18:03](https://www.youtube.com/watch?v=5XhFtrK_yC8&t=1083s) 实施建议与“最后一公里”的本质最后，Joshua 为希望在下周一就开展此类落地的企业管理者提出了四点具体建议： 1. **优先选择关键决策，而非数据集**（Pick critical decisions, not data sets）。 2. **明确每一个决策的责任人**（Identify who owns each decision）。 3. **以“小时”为单位衡量当前的沟通与决策阻力**，而不是凭主观感觉（Measure today's friction in hours, not in opinion）。 4. **带着业务上下文连接正确的数据源**，并在扩展规模的同时始终保持可追溯性（Traceability）。 Joshua 总结道：“数据的‘最后一公里’从来就不是关于物理距离，而是关于在最后那几英尺里发生了什么——智能是否真正传递到了那个必须做出决策的人手中。如果智能能够触达真正做决策的人，这才是最值得投资的最后一公里。” - **四步落地法**：聚焦关键决策 $\rightarrow$ 确定责任人 $\rightarrow$ 量化阻力时间 $\rightarrow$ 带上下文连接数据源并保持追溯 - **演讲核心结论**：智能必须直达决策者本身，这才是数据的终极价值 ## Video description Joshua Bernal - CIO, Grupo Traxion Organizations have invested heavily in collecting data, building dashboards, and deploying AI, yet many still struggle with the final step: turning information into confident operational decisions. This session explores the "last mile of data"—the gap between enterprise data and real business action. Through real world logistics and transportation examples, attendees will discover how AI native Decision Intelligence platforms, conversational analytics, and specialized AI agents are transforming the way organizations interact with their data. Instead of waiting for reports or navigating complex dashboards, business users can simply ask questions, receive contextual insights, and make faster, more informed decisions. The session also explores the role of data governance, trusted enterprise knowledge, and AI agents in creating a new generation of operational intelligence. [Source](https://www.youtube.com/watch?v=5XhFtrK_yC8)

Source quotes

[3:31] Joshua Bernal: An insight is a fact about the past. A decision is a commitment about what happened next. And that's the crossing most projects never make. [8:32] Joshua Bernal: You can trust it because you can trace it.

Original transcript

[0:04] Thank you all for coming here to say I will present one of the the largest [0:12] logistics company around the world, not only in Latin America. Just starting with one video [0:18] just to show you what we're going to do. Every second your business generates an [music] enormous [0:24] amount of information. Fleet vehicles transmitting telemetry, [music] financial transactions closing, [0:31] customer interactions happening, operations running, sensors reading, manufacturing [music] [0:36] producing, emails sent, services delivered, sales made. [music] Everything creates information. [0:45] But that information remains fragmented, [music] scattered across dozens of disconnected systems. [0:51] Finance sees [music] one part of the picture, operations sees another. Customer [0:57] success, [music] sales, executives, everyone works from a different version of reality. [1:03] What if there was a single intelligence [music] layer that understood all of it? [1:08] Not another dashboard, not another data [music] tool, an intelligence that reads every signal, [1:14] finds every [music] relationship, and transforms millions of business events [1:19] into one coherent [music] understanding. Patterns that were invisible become clear. Uncertainty [1:27] becomes confidence. Disconnected [music] signals become one intelligent network. This is Northstar. [1:37] Okay. I'm Joshua Bernal, CIO at Traction. I would like to Sorry. I had that one. I I also [1:49] work on enterprise product development through my own company. Then Um I'm we manage 25,000 trucks [2:00] every day. [snorts and clears throat] Before I get into the problem, let me show you the scale we are [2:07] we actually talking about. 25,000 trucks generating roughly 400 million GPS position a day. [2:18] The whole fleet tracked every 5 seconds. Modern telematics captures over 250 data points a second [2:27] per vehicle. Engine, fuel, tires, driving behavior. On top of that, every down, every [2:36] hard turn, every every event or every unit, every route, every client. That's the raw material. [2:44] What do you do with it? It's completely different question. At the end, the data exists. [2:53] A few years ago, I watched a fleet operations team spend three more days reconciling or trying [3:00] to reconciling a number that should have taken no more than 3 minutes. Not because [3:07] the data didn't exist. It exists effectively in different systems, in five different formats, [3:15] agreed on by no one. That's not a technology problem, of course. That's a last mile problem. [3:23] And that gap between having data and actually using it is what I want to talk about today. [3:31] Uh here's Here's the part that surprise uh all the people hold on here. This the pri- [3:40] here's the part that surprises people that this isn't data shortage. Uh-huh, only 25% of employees [3:48] actively use real real real use the the BI and business analytics. Mhm. All the tools that the [3:56] company already pay for. Between 60 and 80% of dashboards go in use within months of launch. [4:06] It's crazy. People call it the dashboard graveyard and you know what I mean, right? 87% of [4:14] organizations report increased analytics adoptions 5 10 years ago. Yet hasn't moved in 7 years. [4:25] The bottleneck move is not access anymore. It's translation. The shift moves through three eras. [4:41] Sorry. I'm just familiarizing with that. It moves in three eras, right? [4:49] If you could see here, we have 2016 I started in 2010. That is starting in 2010. Um and [5:02] if you could see the traditional dashboards with the static information with all the data that [5:09] are not moving to any anywhere and moving around the the different the different time. [5:19] You could see what is the last mile problem here. [5:24] Um really between these 10 years, we are moving really faster. We are [5:34] we we have more than interactive dashboards, self-service BI and beyond. [5:40] You ask, you get answer and the system can act on it. No dashboard required. [5:51] Um Follow the chain. Here, it's really important. Data becomes information. [5:59] Information becomes an becomes an insight. Insight becomes a recommendation. [6:05] The recommendation becomes a decision, and a decision becomes an action. Mhm. [6:11] But what happened what really happened with the projects right now? The projects stop at [6:18] the insight part. Mhm. An insight is a fact about the past. A decision is a commitment [6:27] about what happened next. And that's the crossing most projects never make. That's the thing. Mhm. [6:37] Numbers aren't decisions, right? Here's why dashboards don't close the gap on the round. [6:43] Every time someone opens one, they have to know where to look, uh understand the structure behind [6:50] the numbers, interpret it correctly, relate it to two or three other sources, and then [6:56] decide what to do. Alone. Really. A dashboard shows a number. It doesn't make a decision. [7:11] Let me make this concrete with logistics. A single unplanned breakdown cost around 2,000 US dollars. [7:19] When you Once you add the repairs, towing, and driver downtime, [7:25] that same repair done on an unplanned schedule costs three to five times less. [7:33] And maintenance alone eats up 9% of total fleet operating. [7:39] Uh 20 cents on on the dollar per mile according to international organizations. Detecting a risk [7:49] and acting on it are two completely different things. And that gap is where the money is. [7:58] So, what does it take to close that gap? The interface has to change. [8:05] You ask describe what you really need to know in plain language. [8:10] The system discovers. It identifies which sources actually hold all the answers. And then it acts. [8:19] A specialized agent explains the result with context, not just a chart. [8:32] Here at Grupo TIERRA Auto Scale, one of the largest companies in the world, not [8:38] only in Latin America, they actually is the second largest in the world. Apply it to our own fleet, [8:45] the number gets real. 25,000 vehicles across three line of business, 10,000 in passenger mobility, [8:53] the most sensitive of three, of course, because we we have people transportation. [8:59] And and the rest is split between general freight and dry goods. Unplanned downtime at [9:06] 8.7 days per vehicle is crazy. Apply that across our fleet, if you are looking at an estimate, [9:17] maybe are more than one 155 million US dollars a year in exposure to unplanned downtime. [9:28] I want to be precise. This is the past of our fleet size. [9:34] And audited internal figure that you could see here. And it tells you the scale of what at stake. [9:44] And why we need to create value, not a dashboard in last mile. [9:53] An AI answer isn't enough if you don't know where it came from. This is where most pilots quite die. [10:02] Not because the model was wrong, really. But because no one trusted it enough to act on it. [10:18] The Trust doesn't come from confidence. It comes from visibility, [10:25] where the number actually comes from, the source. How it got there, every transformation track it. [10:31] The lineage. Who allowed to see it and act on it, permissions. And every access and change locked. [10:41] The audit trail. You can trust it because you can trace it. [10:51] Okay. [10:55] This is where I want to show you what this looks like. [11:04] When it's actually built, this is Northstar. And what I want to be clear, this is a demonstration [11:11] of the principle, not a pitch. It just starts at 48% and it never stop learning. [11:19] Every deployment starts the same way. Northstar goes to into your databases, into our databases. [11:27] It starts training on real data without any context and organize it around your own context. [11:36] Because it it is learning. Uh-huh. Every time, every iteration, it's learning. [11:43] It takes five to six iterations to reach an operational baseline. [11:49] And it keeps learning after that. How much time you would you need to spend for each iteration? [11:55] It depends of all the data that you have into our data lake or into our data mart or data warehouse. [12:03] And it keep learnings after that. Um [12:09] Predictive accuracy uh around 48% as I told you. [12:15] Day one, real data, no shortcuts. Each iteration adds roughly two points by inter by iteration 22, [12:24] you are the 92% approximately. Mhm. Across the predictions that actually matter, [12:32] customer churn, operator attrition, repair prediction based on the history, [12:37] and it doesn't stop at 92. It keeps climbing as long as the data keep coming in. [12:51] This is a live screenshot. You You can see here. You can see here. Uh one question, ask it in plain [12:58] language, the way you will ask a colleague. The system queries the underlying sources directly. [13:07] There's no dashboard to build first. And querying in natural language isn't just faster to look at. [13:14] It cuts a decision latency, the time between insight to action, by an estimated of 65%. [13:27] What is next? It don't stop here. Mhm. This is where it gets interesting, really. [13:34] Every GPS being, every mechanical event fits a rolling health score for each unit. [13:42] When a unit crosses the risk [snorts] threshold, North Star doesn't just show a red flag. [13:51] It books a maintenance visit automatically. Mhm. And it blocks the TMS, uh transport management [13:58] system, from assigning that a newly that unit another trip. Before the breakdown happens, not [14:05] after, applies to our fleet. That's an estimated of 23% to 48% reduction in unplanned downtime. [14:14] If we transform that numbers into money, we are avoiding more than 60 million dollars per year. [14:27] In this industry is is a benchmark. Uh-huh. Uh this this numbers are really acquiring [14:36] more and more and more performance each with each iteration. Right now, we have more than [14:45] 97% of accuracy in the inter- in the into the iterations. [14:54] And if you will see, we can add it a new a new step into our [clears throat] change value. [15:02] The recommendation. Because why? Because the the agent, North Star, reads a lot of data, [15:10] tons of data, and after that, uh they made some recommendations to do with that data. Um [15:21] yesterday, the same number. mhm, [15:27] Yesterday, the same number lives in different five places. Power BI, [15:32] Tableau, spreadsheets, email threads, legacy systems, legacy reports. [15:37] And licensing alone runs uh estimated of 470,000 US dollars. [15:47] This is really important. Mhm. The alert that already books itself [15:53] it's sending to our TMS or sending to our maintenance system automatically, [16:01] and then everything goes well. Um the unit goes into the maintenance with a planet site. [16:11] Okay. Uh after that, the licensing. The licensing is is a matter that [16:20] for all the companies, for the CEO. Licensing in the past, as I told you, Power BI, Tableau, [16:28] spreadsheets. All the reports that you have, all the 80% of the reports that that you really have. [16:36] And licensing alone runs an estimated 470,000 400 US dollars a year. Across 2,800 seats, [16:47] more or less, in our company. Maybe you have more in your company. With Northstar, it's one board, [16:54] one conversation, no per seats. Uh dashboard built in real time for whatever the business needs. [17:03] Agents trained us on our company data. This is the most important thing. In our company data. We are [17:10] not trying outside of the world of our company. It has our data. Uh-huh. Not only integrated in our [17:19] in our data lake, but also in the sources, but also in different ways into our company. That's [17:27] an estimated of $470,000 a year in licensing eliminated. Your CEO will be happy for with this. [17:37] Plus 29% faster decision cycle. This is the most important thing when you are talking about [17:46] a business that depends on real time. [17:50] Five places becomes one. So, the answer is trusted the months the moment it's seen. [18:03] So, if you are trying if you are really trying to close in this gap starting Monday morning, [18:10] here's where to start. Pick critical decisions, not data sets. Identifies who owns each decision. [18:20] Measure today's friction in hours, not in opinion. And connect the right sources with context. [18:29] Keep traceability as you scale. The last mile was never really about distance. [18:37] It was about what happened in those final few feet. Whether the intelligence actually reaches [18:44] the person who has to decide. This is the main thing. If intelligence [18:50] could reach the person that really decides, that's the mile worth investing in. Thank you.