Supply Chain Operations, Run by Agents
How Stewdio rebuilt the operations layer of a $2.1B consumer goods distributor - from disruption handling to returns - in four months.
Company Profile
The Challenge
Every disruption triggered a scavenger hunt. A late container or a line-down supplier meant someone pulled stock positions from Oracle, checked DC capacity in Korber, opened the Kinaxis plan, then chased four colleagues over Teams to agree on a response. Typical time from signal to action: 3 to 5 days. The stockout rate sat at 5.9%. Ten planners and coordinators met for a 75-minute daily huddle built around a shared spreadsheet - most of it read-aloud status. Supplier performance lived in fragments across Coupa, the WMS, and quality logs, so nobody could answer a simple question: which suppliers are actually hurting us? The planning cadence made it worse. The weekly S&OP refresh was stale the day it shipped. DC allocation ran as an overnight batch that could not react to same-day signals. 19% of system-generated purchase orders were changed after issue. Dock scheduling depended on inbound ETAs that three people maintained by hand. Returns took 6 to 8 days to disposition, tying up sellable inventory the whole time. The operations team had become human middleware between seven systems, and every resignation took untransferable know-how out the door.
The Audit (5 Weeks)
A Stewdio forward-deployed team sat inside the operations org for five weeks: two weeks embedded in the daily huddle, site visits to 3 of the 7 DCs, and a full week tracing purchase orders from MRP suggestion to receipt. The team catalogued 39 recurring exception patterns across the department.
26 of 39 exception patterns are rule-expressible - The remaining 13 require real trade-off judgment and belong with humans
Context gathering consumed over half of coordinator time - Time measured from exception raised to first decision-relevant action
SKU-level forecast accuracy had plateaued - No planner had bandwidth to investigate systematic error sources
One in five purchase orders amended after issue - MRP quantities ignored live demand and supplier lead-time drift
Dock labor consistently over-scheduled - Estimated 12-18% excess hours driven by unreliable inbound ETAs
The Architecture
A single agent layer spanning the whole operations chain, from first demand signal to final returns disposition.
Exception Management
A detection agent watches Oracle, Korber, and Kinaxis for disruption signals and opens a structured case with every relevant data point already attached. A resolution agent closes the 26 rule-expressible patterns on its own. For the 13 judgment cases, a briefing agent assembles the situation, drafts two or three response options with cost and service trade-offs, and routes the package to the accountable owner.
Demand Planning
A sensing agent refreshes forecasts continuously, blending sell-through data, open order momentum, and returns velocity with outside signals such as freight lane congestion, commodity price moves, and promotional calendars. A planning agent assembles the S&OP deck automatically - variance analysis, risk callouts, and proposed adjustments included - so the meeting starts at the decision, not the data.
Procurement
A PO validation agent checks every MRP suggestion against live demand, lead-time variability by supplier, and open order coverage before release. A supplier intelligence agent merges Coupa, WMS receiving data, and quality records into one living scorecard per supplier. A compliance agent flags off-contract purchases and volume-rebate thresholds at risk of being missed.
Warehouse Coordination
An inbound tracking agent pulls live shipment status from carrier APIs and the EDI gateway, giving dock schedulers dependable ETAs and right-sizing labor plans. An allocation agent rebalances stock across the 7 DCs continuously, replacing the overnight batch. A returns agent classifies incoming units on arrival and routes each to restock, refurbishment, or disposal without queueing.
Quality
A quality signal agent links inspection holds and defect logs directly to supplier scorecards, surfaces emerging defect trends before they become claims, and auto-drafts corrective action requests. Quality history now sits beside price and lead time in every sourcing decision.
Implementation (4 Months)
Each month shipped working agents into production, not slideware.
Daily huddle dropped from 75 to 20 minutes. Dependable inbound ETAs for the first time. Dock over-scheduling began falling within two weeks.
Draft acceptance hit 91% in the first week. The 8 highest-volume exception patterns ran fully autonomously by month close. PO amendment rate on the top 40 SKUs fell from 19% to 7%.
Unified scorecards exposed 3 chronically underperforming suppliers nobody had quantified. Continuous allocation retired the nightly batch. S&OP became a decision meeting.
The quality agent caught a month-old creeping defect trend at one supplier before it reached customers. All agents connected end to end across every sub-function.
Results
Measured 90 days after full rollout, against the trailing-year baseline.
This case study is an illustrative scenario: a representative engagement model, not a named client. We publish real deployment numbers as engagements complete.
ROI: $6.4-8.4M Recurring + $5.6M One-Time
We report the two value types separately because they behave differently. The $5.6M working capital release came from shrinking safety stock that existed only to absorb slow disruption response - it lands once, and the freed cash stays freed. The recurring $6.4-8.4M is annual and tends to widen as agent accuracy compounds.
Stockout reduction - Recovered margin on prevented lost sales plus expedite freight avoided
Coordinators redeployed - 9 named roles moved from status-chasing to supplier development and network design
Capacity recovered - Dock labor hours and planner time returned to productive work
Procurement efficiency - Fewer PO amendments, captured rebates, reduced off-contract spend
Quality cost avoidance - Defects intercepted before shipment instead of after claims
Returns acceleration - Sellable inventory back in stock 5+ days sooner
Cross-Department Impact
The operations layer did not stay contained. Supplier reliability scores began informing finance's payment prioritization. Finance returned spend concentration analysis that reshaped the diversification roadmap. Demand sensing outputs now feed the commercial team's revenue forecast directly. Within three quarters, operations, finance, and sales were working from one shared intelligence layer - value we deliberately excluded from the figures above.