agentic ai · procurement · in development
Procurement OS
An autonomous procurement agent that performs procurement work — expediting, tendering, spend analysis — with the domain expertise encoded as skills in natural language. A main agent plus two subagents mirror a human procurement team. V1.0 covers expediting.
The architecture
agents → skills → tools, on a managed harnessTriggers
Chatthe user asks
Schedulecron cycles, always-on
Inbound emaila supplier replies
↓wakes
Agents
Main agentcoordinates, talks to the user, delegates
Subagentsspecialists — one supplier-facing, one data-facing
↓load
Skills
Procurement know-how written in natural language, not code — 14 skills in v1.0, loaded on demand. Adding a capability means writing skills, not rebuilding the system.
↓act through
Tools
Emailwrites to suppliers
Databasereads & writes records
ERP filesparses extracts
Schedulerplans its own follow-ups
The stack
built with
Claude Code
harness
Deep Agentsby LangChain
model
Claude Sonnet
memory + audit
PostgreSQL
observability
LangSmith
Every action lands in the database — multi-user, auditable end to end.
What to automate first
every procurement capability, mapped · v1.0 highlightedadaptive — shaped by the counterparty
procedural — follows rules
transactional
lower priority
Supplier Onboarding
Procedural on paper; every supplier presents different exceptions.
build first
Expeditingv1.0
RFQ Tactical
RFP Tactical
RFI Tactical
High volume, rule-driven, already outsourced today — the agent replaces expensive human time.
strategic
human domain
NegotiationStakeholder ManagementSupplier ManagementRFP/RFQ Strategic
The agent prepares — data packages, scenarios — and the human executes.
Process depth adjusts to where the category sits on the Kraljic matrix — routine, leverage, bottleneck, critical.