Closing the Inventory Loop: Autonomous Agents Inside a Multi-Site Manufacturer
Stockouts and overstock both fell sharply,
as agents took over reorder timing and quantity decisions
Cycle counts moved from quarterly events to continuous,
with anomalies surfaced and resolved before they hit production
Operations team reclaimed weeks of manual reconciliation,
redirecting effort toward supplier strategy and new-line ramp-up
About the Project
Services
Project Details
A multi-site industrial manufacturer with operations across Asia-Pacific engaged Webpuppies to fix a problem that no amount of ERP customisation had solved: inventory truth was fragmented across factories, warehouses, suppliers, and finance – and every reconciliation cycle revealed a different version of reality. The result was an agentic AI layer that owns the inventory loop end-to-end, with humans intervening only on genuinely ambiguous decisions.
The Challenge
Inventory was a fiction the company maintained quarterly. Between cycles, every team – production planning, procurement, finance – operated on a different snapshot, and the snapshots disagreed. Stockouts triggered emergency air freight; overstock tied up working capital and aged into write-offs. The cause wasn’t bad data – it was no single owner of inventory truth.
Key Challenges
Inventory truth fragmented across systems
Reorder decisions were reactive, not predictive
Manual reconciliation consumed weeks per cycle
No defensible audit trail
Without an agentic layer that continuously reconciled the picture and acted on it, the company was paying for inventory it didn’t have and storing inventory it didn’t need.
Our Solution
From Reactive Reconciliation to Self-Correcting Inventory: An Agentic Loop That Closes Itself
Agentic AI Development for Industrial Operations
Webpuppies designed and shipped an agentic AI layer that treats inventory as a continuously reconciled live state, not a quarterly snapshot. Agents own the boring middle of the loop – sensing, deciding, acting, auditing – and humans focus on the small set of decisions that genuinely require judgment.
Unified Inventory Graph
Every SKU, every location, every in-flight order, every supplier commitment lives in a single graph that ingests from ERPs, WMSes, supplier feeds, and shop-floor sensors. The graph is the source of truth that all agents read from and write to.
Specialised Operations Agents
A demand-forecasting agent, a reorder agent, a supplier-comms agent, and a discrepancy-investigation agent – each with a narrow surface, each reliable enough to ship in an industrial context where wrong calls cost real money.
Safety Primitives for Operational Decisions
Every agent action affecting purchase orders, write-offs, or production schedules passes three checks: source citation against the graph, confidence scoring with human-review thresholds for high-value decisions, and deterministic validation against business rules. No exceptions.
Continuous Audit by Construction
Every agent decision is logged with its inputs, reasoning, and outcome – so when finance asks ‘why did we buy this?’, the answer is structured, complete, and reproducible.
Ongoing Partnership
Beyond the first agents shipping, Webpuppies continues to expand the agent roster, harden the safety architecture, and onboard adjacent operational domains – turning the inventory loop into the foundation of a broader operational intelligence layer.
Agent Roster Expansion
New agents on the roadmap for production scheduling, supplier risk monitoring, and quality-event correlation. Each ships faster than the last because they share the same graph, tools, and safety primitives.
Safety Architecture Hardening
Continuous tuning of confidence thresholds and validation rules as the agents take on increasingly high-stakes operational decisions.
Knowledge Graph Enrichment
Continuous ingestion of supplier-performance signals, market price data, and shop-floor telemetry so the graph gets denser and the agents get smarter with every cycle.
Multi-Site Onboarding
Each new factory or distribution centre launches as a configuration pass, not a rewrite – the architecture localises through config, not code forks.
Results & Impact
Stockouts and Overstock Both Fell Sharply
Continuous Cycle Counts Replaced Quarterly Reconciliation
Discrepancy-investigation agents now surface and resolve mismatches as they appear, not three months later. Senior operations staff no longer lose multiple weeks per quarter to manual cycle counts.
Audit Trail Became a First-Class Output
Finance can now trace any inventory decision – purchase, transfer, write-off – to its source signal and reasoning. Discrepancy investigations that used to take days now take minutes.
Operational Capacity Reallocated Toward Strategy
The headcount that used to maintain inventory truth now drives supplier consolidation, new-line ramp-up, and SKU rationalisation – moving from reactive bookkeeping to forward-looking operations.
Related Projects
MyRepublic
Increasing Conversions with a Region-Specific, User-Friendly Website
See how we helped MyRepublic create an intuitive, multi-regional website that boosted engagement and conversions.
Ron Kaufman
Merging Two Websites into One Seamless Platform
Find out how we helped Ron Kaufman, one of the world’s leading customer service keynote speakers, integrate his two websites into one seamless platform for a better user experience.
