CHAPTER 03 · BUILD
Agents Shipped Inside the Existing Stack
The first system, and the one right behind it, shipped inside the existing ERP — no new platform, no rip-and-replace. Each one is real and running.
THE AGENTS · 3 FOR THIS WORKFLOW
This network needed three. The count isn't a package — it's however many distinct jobs the work splits into. Each agent is a separate identity with its own scoped tools and permissions.
Demand-Driven Reorder Recommendation
Supplier Risk Monitoring
Warehouse Slotting & Pick-Path Optimization
IT DOESN'T STOP AT A REPORT
The Agent Takes the Action
This one barely looks like a report at all — it runs the reorder. Most of what these agents do is take the action inside the ERP, with a buyer approving anything that spends money.
Run the demand plan
Re-forecasts every SKU against real orders each morning — nobody rebuilds a spreadsheet.
Create the purchase order
When a SKU crosses its live reorder point, the agent drafts the PO — supplier, quantity, terms — for the buyer to approve.
Send the PO to the supplier
On approval it goes out through the existing EDI or email path automatically.
Flag the dead stock
Opens a markdown-or-return recommendation on slow movers before the cash is stranded.
PRODUCTION TRACE
Demand-Driven Reorder Recommendation
Every step below runs against the plant's real data — the exact sequence, in order.
142
Reorder recommendations generated
121
Recommendations accepted
5
Supplier risk flags reviewed
8
Slow-mover alerts carried forward
6 items need your judgment
Ingest
AutomatedLive sales and inventory data pulled from the ERP every morning across all three DCs — no new sensors, no new data entry.
Fan out by DC
AutomatedOne data pull becomes three reorder recommendation sets — one per distribution center, each grounded in that DC's own demand pattern.
Ground every claim
AutomatedEvery recommended reorder cites the exact demand trend or supplier lead time behind it, not a general opinion.
Buyer review
Reviewer inputEach recommended order is one-click accept or reject. Rejections become training signal for tomorrow's recommendation.
Tomorrow's diff
AutomatedThe next morning's list opens with a diff against today's accepted orders — did it work, what changed.
REAL OUTPUT · SUPPLIER RISK MONITORING
One Example, Fully Processed
When a job calls for a written artifact, this is what the system produces — everything it finds, structured for the action it triggers next.
Supplier Risk Scan — Bearings & Fasteners Category
30 monitored suppliers · scanned against OFAC, Federal Register, and delivery-performance data · updated hourly
Supplier Risk Signals Detected
| Supplier | Signal | Source | Detail | Confidence |
|---|---|---|---|---|
| Supplier SR-118 | Delivery delay pattern | Historical EDI | 3 consecutive late shipments on the bearings line | 93% |
| Supplier SR-204 | Sanctions list name match | OFAC | Partial name match — resolved as false positive on manual review | 71% |
| Supplier SR-092 | Price volatility | Federal Register + FRED | Raw-material tariff change flagged upstream | 88% |
| Supplier SR-140 | Financial distress signal | Public filing scan | Late-payment pattern on trade references | 82% |
| Supplier SR-063 | Capacity constraint | Delivery performance | Lead time extended 40% over trailing quarter | 90% |
Supplier Risk Summary — for Procurement
Five suppliers carry a flagged signal this week. One is a resolved false positive (name-match only); the other four warrant a closer look before the next PO cycle — particularly SR-063, where lead time has grown 40% in a quarter with no communicated cause.
Suppliers to Watch This Week
WAVE 2 — IN BUILD
Warehouse Slotting & Pick-Path Optimization — What's Next
High-velocity SKUs are slotted far from the packing stations in two of the three DCs — a layout that made sense two inventory cycles ago and hasn't been revisited since.
The agent proposes a re-slotting plan; a pick-path solver checks the math against real order-mix data. Nothing ships on an LLM's arithmetic — the solver's answer is the answer, the agent just explains it.
Worked example the solver evaluates today in testing
DC2's top 50 SKUs by pick frequency are compared against their current slot distance from packing. Four re-slotting approaches scored by projected daily pick-time savings:
WHAT CHANGED
Time returned
~85% of manual reorder review time
Buyers got the time back for supplier negotiation and exception handling, not for a longer spreadsheet.
Operating complexity
A once-set-and-forgotten min/max rule became a reviewed, one-click-accept recommendation
Where this stands
First system live in 30 days
UP NEXT · TEAM VIEW