+
Amazon Bedrock · Amazon EC2 · Amazon DocumentDBAgentic Service Operations
Rebuilding after-sales operations around Amazon Bedrock agents — warnings before failures, dispatch before phone calls, parts before demand. Fleet unplanned downtime fell 41%.
Zhejiang EP Equipment Co., Ltd.
In 2025 the company launched a Bedrock-based service-knowledge platform with AI-Driven. Answers got faster — but the operating model itself had not changed: break-fix service, manual dispatch, and judgment-based stocking.
Reactive Service, Idle Telemetry
Customers called only after breakdowns. Preventive work accounted for just 12% of maintenance orders, fleet unplanned downtime ran 14.6 hours per vehicle per half-year, and real-time telemetry from roughly 50,000 connected vehicles was only reviewed after the fact.
Manual Dispatch, Low First-Time Fix
Dispatchers hand-balanced skills, location, daily schedules, parts availability, and workload at about 25 minutes per order. Mis-assignment and missing parts kept the first-time fix rate at 61%, and repeat visits drove up both downtime and cost.
Coarse Global Parts Management
Eighteen overseas warehouses ran on three heterogeneous sources — a DAS system, an Australian WMS, and offline Excel. Stockouts hit 8.6%, triggering repeat visits and about $9,800 of emergency air freight per month, while multi-currency data and inconsistent part numbering blocked any global view.
Slow, Unauditable Settlement
Engineers spent about 20 minutes per order on paperwork with inconsistent fields. Reconciliation was manual, exceptions were caught by sampling, and nothing traced settlement back to the dispatch decision behind it.
A two-layer "deterministic engine plus LLM agents" architecture with Amazon Bedrock at the core, where three business lines share one governed cloud foundation
The Amazon Bedrock Agent Families
Sonnet-tier Claude models drive the reasoning chains, a Haiku-tier model handles high-volume classification, routing, and pre-scoring, and Amazon Titan Text Embeddings V2 vectorizes knowledge and cases — all invoked privately through VPC endpoints, with enterprise data never used for model training
Fault understanding → repair planning → parts coverage → an LLM ranking agent that re-ranks the rule pipeline's top candidates and writes its rationale, fused 0.6/0.4 with deterministic scores
Daily autonomous fleet inspection and anomaly attribution, opening preventive work orders at least 48 hours before predicted failures
A replenishment agent reviews every MRP proposal and issues cited purchase proposals; daily agents tune dispatch parameters and a weekly agent evolves strategy from closed-loop data
Proactive Service: Warnings Before Failures
TBox telemetry from roughly 50,000 vehicles enters through Elastic Load Balancing into self-managed stream processing and a complex-event engine on Amazon EC2 inside private VPC subnets, where trend analysis plus an expert-rule base raises risk warnings that Bedrock agents attribute and turn into work orders.
≤5s processing latency, ≥1,000 events per second
50+ scenario expert-rule base layered on trend analysis
Preventive orders opened 48+ hours ahead automatically
Agentic Dispatch & Settlement: Dispatch Before Calls
The Bedrock agent chain resolves fault features and parts coverage first; a deterministic dispatch pipeline on EC2 (parallel hydrators, hard filters, multi-dimension scorers, selector) produces candidates, and the LLM ranking agent issues the final order with its rationale. Settlement runs on AI pre-fill, scan-to-verify, and slider confirmation, reconciled by a dual rule-plus-LLM engine.
Technician profiles across 12+ dimensions, filtered and scored in parallel
Every decision fully traced in Amazon DocumentDB
AI pre-fill and scan-to-verify; 96% automated reconciliation
Global Parts Intelligence: Parts Before Demand
Three data sources across 18 overseas warehouses — DAS, WMS, and offline spreadsheets — feed a daily ETL into Amazon RDS (PostgreSQL, Multi-AZ). A two-step MRP engine computes safety stock and replenishment proposals, and a Bedrock replenishment agent reviews each one against logistics lead times, promotions, and product transitions.
~54,000 inventory snapshot rows per day, unified on a USD base
Every proposal issued as a cited purchase recommendation
Service Catalog replicates environments per brand and region
Agents don't switch to autopilot overnight. A shadow → suggestion → autopilot rollout, backed by circuit breakers, full traceability, and hard safety limits, keeps every expansion of autonomy explainable, auditable, and reversible.
Shadow Mode
Agents decided alongside dispatchers without acting. Double-blind review of every disagreement surfaced undigitized tacit rules — contract-exclusive technicians, customer blacklists, restricted products — which were encoded into the pipeline's hard filters.
Suggestion Mode
Agents propose a ranked shortlist with natural-language rationale for dispatchers to confirm; every acceptance and override flows back as a training signal for strategy evolution.
Autopilot Mode
Rollout is segmented across four dimensions — product SLA, customer tier, region, and order type — so management can reclaim autonomy along any one of them at any time.
Out-of-range parameters or malformed output degrade to suggestion mode with an alert — no silent failures
Every candidate's elimination reason, dimension scores, and LLM rationale are retained in Amazon DocumentDB
Live-electrical and lifting work is always routed to humans; customer-facing output is PII-masked
Bedrock is reached from private subnets, and enterprise data is never used for model training
LLM ranking is fused with deterministic scores rather than replacing them, preserving predictable fallback behaviour
Weekly strategy evolution requires p<0.05 validation, backtesting, and a 10% canary before full rollout
Measured January–June 2026 from the customer's work-order system, telemetry platform, dispatch decision logs, and global parts platform. All figures are measured, not projected.
Fleet Unplanned Downtime
Telemetry platform + work-order system14.6 hrs per vehicle per half-year
8.6 hrs per vehicle per half-year
Preventive Work Orders
Work-order system (order-origin tagging)12%
47%, triggered automatically by risk warnings
Failure Risk Warnings
Proactive service platform warning logsNo system in place
48+ hours ahead; 87% precision; 81% alert-handling rate
First-Time Fix Rate
Work-order system61%
78%
Dispatch Handling Efficiency
Dispatch decision logs (100% traced)25 minutes per order, manual
40 seconds per order in autopilot; P95 under 3 minutes
Global Parts Stockout Rate
Daily snapshots across 18 warehouses + DAS/WMS8.6%
5.2%, with inventory turns up 17%
Settlement & Reconciliation
Settlement system~20 minutes per order; manual reconciliation
28 seconds per order (P90); 96% automated reconciliation
Degree of Autonomy (Adoption)
Dispatch system operating statisticsNone
84% suggestion-mode adoption; autopilot covers 62% of orders
87% warning precision and an 81% alert-handling rate in canary
Elimination reasons and LLM rationale remain fully auditable
Three heterogeneous sources feed one governed platform daily
TBox telemetry streaming continuously at ≤5s processing latency
Agent invocations grew about 2.8× as autopilot expanded, yet average model cost per work order fell 47%. As autonomy coverage widens and parts working capital declines, three-year total TCO is expected to run about 50% below a self-hosted-GPU-plus-headcount path (projected).
Measured January–June 2026 via AWS Cost Explorer. LLM ranking runs only over the pipeline's top-5 candidates, Haiku-tier pre-scoring absorbs high-volume light tasks, and strategy agents batch off-peak — together holding unit costs down.
The cost pool of dispatch and settlement staffing plus emergency air freight fell from $40,700 to $23,300 per month including incremental AWS spend — $17,400 saved monthly, a ~43% reduction — as dispatch and reconciliation staff moved to exception handling and rule governance.
Bedrock spend averages about $2,500 per month against an estimated $22,000 per month for an equivalent self-managed GPU inference cluster plus 0.5 MLOps FTE.
Let Agents Run Your Service Operations
See how AI-Driven builds explainable, auditable, and reversible agentic operations platforms on Amazon Bedrock