A particular thanks goes to the Verizon Join crew who’s been working very onerous on the mission: Matteo Simoncini, Luca Bravi, Alberto Rossettini, Martin Villarruel, Ceyhun Unlu, Adriel Zuquini, Andrea Benericetti.
Fleet managers right now face an awesome problem: reworking knowledge overload into actionable insights. While you’re managing hundreds of autos, every producing a whole lot of every day knowledge factors, figuring out important patterns turns into practically not possible by way of handbook evaluation. Verizon Connect, a worldwide fleet administration options supplier serving companies worldwide by way of its Reveal platform, encountered this actual problem at scale.
With over 1.2 million lively automobile subscriptions producing over 500 million knowledge factors every day throughout 80,000 distinctive knowledge indicators, fleet managers had been drowning on this knowledge and compelled to hunt for anomalies throughout fragmented paper logs and reactive spreadsheets. The sheer quantity made it not possible to establish rising questions of safety, upkeep wants, or operational inefficiencies earlier than they grew to become expensive issues.Fairly than constructing one other static dashboard or rule-based automation system, which solely catches predefined patterns, Verizon Join selected agentic AI to exchange that handbook guesswork with a centralized intelligence resolution. Agentic AI dynamically investigates new patterns, asks follow-up questions, and adapts its evaluation primarily based on what it discovers, making it effectively fitted to the unpredictable nature of fleet operations.
On this put up, we present you the way Verizon Join constructed and scaled an agentic AI resolution to remodel overwhelming fleet knowledge into clear, actionable insights for 100,000 customers every day. We stroll you thru the architectural selections, implementation challenges, and measurable outcomes that may information your personal data-to-insights transformation.
Constructing scalable structure
The answer handles knowledge at scale whereas sustaining cost-efficiency. The next determine describes the core elements. Later on this part, we stroll by way of and talk about the varied elements of the answer and tie them collectively within the ‘General structure’ part.
Determine 1 – Excessive-level resolution structure
The high-level description of the elements and the logical move is as follows.
- Orchestration begins – The every day set off initiates the workflow by activating the anomaly detection module inside the evaluation layer.
- Knowledge ingestion – The module pulls structured data from the uncooked knowledge retailer. That is the place the computationally heavy lifting occurs.
- Focused outputs – Fairly than asking a big language mannequin (LLM) to “discover needles in a haystack,” this module identifies particular anomalies and writes them to a devoted anomalies desk. By offloading numerical evaluation to specialised code, we keep away from the size and accuracy points LLMs face with uncooked tabular knowledge.
- AI agent activation – After the anomalies are prepared, the supervisor triggers the AI brokers. To enhance efficiency, a number of brokers can run in parallel, every specializing in a special buyer or knowledge phase.
- Reasoning & context – The AI agent orchestrates the ultimate evaluation. It queries the anomalies for the what and refers again to the uncooked knowledge for the why, utilizing the LLM to synthesize these inputs right into a coherent narrative.
- Perception supply – The ultimate reasoning is saved as generated insights, that are then served to the end-user by way of the Reveal software.
Elements of the answer
Anomaly detection
A typical pitfall in AI engineering is asking an LLM to carry out numerical evaluation on large-scale uncooked tabular knowledge. As AWS Prescriptive Steerage notes, LLMs can wrestle with advanced desk constructions and numerical extraction at scale. To handle this, we constructed a serverless statistical mannequin utilizing AWS Step Capabilities and AWS Lambda (See Determine 4). This mannequin performs the computationally intensive work of anomaly detection on structured knowledge. It identifies what the anomaly is, so the AI agent can concentrate on why it occurred and the right way to tackle it.
AI agent
We chosen Strands Agents, an open supply SDK for constructing and executing AI brokers, working in a serverless AWS Lambda setting. This deployment sample scales horizontally primarily based in your demand. The AI agent operates by way of a dynamic reasoning loop, autonomously figuring out the required investigation path quite than following a hard and fast set of steps. From the next description you possibly can discover that the AI agent is stateless, the context required for the insights technology is retrieved recent at evaluation time.
The AI agent makes use of particular instruments to:
Two-stage agentic structure
The perception technology follows a two-stage strategy, every utilizing the LLM’s reasoning capabilities in another way:
Stage 1: Abstract technology (anomaly aggregation & prioritization)On this first stage, the agent receives a set of uncooked anomalies detected throughout the fleet. Fairly than processing every anomaly individually, the LLM autonomously decides the right way to combination them into coherent perception candidates. It might probably group anomalies by:
- Frequent root causes (reminiscent of a number of autos exhibiting the identical habits sample)
- Temporal correlation (occasions focused on particular dates)
- Categorical similarity (associated security or effectivity metrics)
Each the grouping logic and the choice standards are totally on the LLM’s discretion. The system doesn’t impose fastened guidelines on how anomalies ought to be mixed. After aggregation, the agent assigns a relevance rating to every candidate’s perception primarily based on elements reminiscent of severity, recurrence, fleet-wide impression, and actionability. From these scored candidates, the agent selects the highest 4 most related insights to proceed to detailed technology. With this strategy prioritization adapts to the precise context of every person’s fleet, quite than counting on static enterprise guidelines which may miss rising patterns.
Stage 2: Detailed technology (agentic tool-based Investigation)The second stage is the place the agentic nature of the system turns into important. For every abstract perception, a separate agent occasion is spawned with entry to knowledge retrieval instruments. The agent autonomously decides which instruments to name, in what sequence, and what number of instances—iterating till it has gathered enough proof to provide a data-backed perception. Now that the agent execution is defined, let’s study why an agentic strategy is crucial for this use case. Fleet administration includes numerous variables and unpredictable eventualities that require dynamic investigation quite than predetermined logic, that creates two basic limitations:
- Finite sample protection: Code can solely detect patterns that had been explicitly programmed. Edge instances, novel correlations, or sudden knowledge distributions go unnoticed.
- Inflexible investigation flows: A script follows predetermined steps no matter what the info reveals mid-analysis.
In distinction, the AI agent can uncover patterns of any nature, together with edge instances that weren’t anticipated throughout improvement. If the info suggests an sudden correlation (reminiscent of harsh braking occasions correlating with particular time-of-day patterns, or a automobile’s habits altering after a selected date), the agent can pivot its investigation technique in actual time, making extra device calls to discover these emergent hypotheses. This flexibility is especially useful in fleet administration, the place:
- Driver habits varies unpredictably
- Exterior elements (climate, visitors, street circumstances) create non-obvious correlations
- Fleet composition and utilization patterns evolve over time
- New anomaly sorts would possibly emerge that weren’t current in historic knowledge
Instance of move orchestration
- Preliminary discovery: The agent identifies a abstract perception reporting a 30 % enhance in harsh braking occasions throughout all the fleet.
- Establishing a baseline: It autonomously decides to drag the fleet-wide historic common for the final 30 days to substantiate if this spike is a real anomaly or only a typical seasonal development.
- Sample recognition: Upon seeing that the spike is focused on particular dates, the agent shifts its focus to a every day breakdown. It discovers that the majority of those security occasions occurred on the identical two days.
- Focused drill-down: Based mostly on these dates, the agent identifies that autos 1015, 1142, and 1032 had been answerable for practically 70 % of the occasions.
- Comparative evaluation: It then makes a last, focused request to check the historic efficiency of Automobile 1015 towards its present habits to see if the motive force’s profile has essentially modified.
- Synthesis: solely after “connecting these dots” does the agent terminate the loop and generate a last, data-backed perception that highlights the precise drivers and dates concerned.
Massive language mannequin
To optimize price-performance, we first used the high-tier Claude 4.5 Sonnet to validate logic and perception high quality. Put up-validation, we transitioned to the extra cost-efficient Claude 4.5 Haiku for our manufacturing use case. Additional price-performance optimization led us to Amazon Nova 2 Lite, a lightning-fast multimodal mannequin, which delivers comparable perception high quality whereas lowering enter token prices by 70 % in comparison with Claude 4.5 Haiku. This substantial saving is important for the reason that workload is dominated by enter tokens (telematics knowledge, anomalies, context). The effectivity of Nova 2 Lite allows Verizon Connect with ship AI insights extra cost-effectively to its whole person base. High quality was maintained by way of an automatic testing suite and a gold-standard dataset, guaranteeing a battle-tested resolution upon full launch. LLMs are hosted at scale in Amazon Bedrock, a totally managed service with complete generative AI functionality, safety, privateness and accountable AI options.
Day by day set off administration and concurrency
To supply insights prepared initially of their enterprise day to the 100,000 customers, we use Amazon Easy Queue Service (Amazon SQS) to handle execution. By controlling the utmost concurrency of the SQS-to-Lambda set off, we are able to:
- Clean out spikes in API demand
- Keep inside Amazon Bedrock quotas: Tokens Per Minute (TPM) and Requests Per Minute (RPM)
- Present dependable supply with out over-provisioning sources
As an example, think about a state of affairs delivering insights for patrons throughout all the United States. The goal supply of insights is 8:00 AM ET, primarily based on knowledge generated as much as Midnight PT yesterday. Given the three-hour time zone distinction, the end-to-end course of should be accomplished inside a five-hour window. Allocating one hour for anomaly detection leaves a four-hour window for the AI Agent and LLM to generate insights. At a charge of 1,500 RPM (adjustable), the perception technology part will take roughly 1.25 hours, effectively inside our operational necessities.
Determine 2 – Most Concurrency SQS-to-Lambda particulars
Generated insights
The insights the agent generates are prepared for the Reveal software to devour. Upon login, new insights seem in a devoted panel on the dwell map, Reveal’s most visited web page, so that each person sees related insights instantly. Every perception is clickable, resulting in an in depth web page with the complete evaluation.
Determine 3 – Reveal’s most visited web page consists of hyperlinks to Operational Insights outcomes.
General structure
The general structure places collectively these 4 elements: Anomaly detection, parallelization of requests, insights technology engine, and storage of generated insights for consumption by Reveal software.
Determine 4 – General structure
To know how these elements work collectively, think about the next workflow: An insights request is triggered, together with the record of buyer IDs for which insights should be calculated. The statistical mannequin performs anomaly detection and shops leads to Amazon S3.
- N parallel requests are triggered as described earlier.
- N brokers analyze knowledge, create insights, and save them to Amazon S3. On this part, brokers invoke Amazon Bedrock APIs to achieve the chosen mannequin.
- The Reveal software can entry the insights saved in Amazon S3.
The outcome: actionable intelligence
The Operational Insights function was rolled out to Verizon Connects users in November 2025,and has served fleet managers with clear, pure language narratives like the next:
- Security sample detection: “Your fleet noticed a 100% enhance in harsh braking this week. Apparently, this coincides with a discount in harsh acceleration, suggesting driver fatigue or elevated congestion.”
- Operational effectivity: “Automobile #90000 is idling for 50% of its engine-on time, considerably above your fleet common. This represents pointless gasoline price.”
- Fleet efficiency: “Day by day mileage is down 59%, however dashing occasions are up 54%. This means autos are touring shorter distances at increased speeds—think about route optimization.”
Conclusion and searching forward
On this put up, we confirmed how Verizon Join constructed a scalable Agentic AI resolution on AWS that transforms uncooked IoT telematics knowledge into actionable fleet insights for over 100,000 customers. The structure combines Amazon Bedrock, Strands Brokers, AWS Step Capabilities, Amazon SQS, and a multi-tier knowledge layer to ship dependable, cost-efficient insights at scale.
Because the AI panorama evolves, we plan emigrate from AWS Lambda primarily based agent deployment to Amazon Bedrock AgentCore Runtime to additional streamline our AWS Lambda execution and use Mannequin Context Protocol (MCP) for quicker device integration.
To implement an agentic AI resolution successfully, start with a small-scale pilot to validate a fundamental use case and set up cost-efficiency. After the preliminary worth is confirmed, broaden the system by integrating automated workflows and data-driven personalization. The ultimate stage includes transitioning to a full enterprise deployment that helps superior orchestration and real-time processing throughout all the group.Begin constructing right now:
- Begin with the foundations: Amazon Bedrock for managed AI companies and AWS Lambda for serverless agent internet hosting
- Add orchestration: AWS Step Capabilities for workflow administration and Amazon SQS for scalable queueing
- Retailer and retrieve knowledge: Amazon S3 for object storage, Amazon Aurora for structured knowledge, and Amazon DynamoDB for quick lookups
- Discover agent frameworks: Strands Agents documentation for open supply agent improvement
In regards to the authors
