A procurement agent for a holding: faster approvals, leaner stock
An agent carries a purchase request from the limit check to the supplier order in 1C and the systems around it. A person confirms only the disputed steps. The approval cycle is 35–45% shorter, excess stock 12–18% lower.
What changed in procurement
Master data, requests, limits and approval routes; exchange with 1C and the bank without retyping.
A forecast of how much to order by item and promotion, using the minimum order and lead time.
Spend and investment screens, budget variance, supplier discipline, and an alert if the 1C exchange stops.
What the numbers show
Procurement approval cycle −35–45%; excess stock −12–18% at the target service level.
Management reporting prep about 2× faster, because the figures come from one system of record.
Highlights
- The request, the spending limit and the approval route live in one process, not in email.
- The agent suggests an order quantity from lead time and the minimum order.
- Critical decisions stay with a person: the agent prepares the pack, a manager confirms.
- One screen shows approval time, budget variance and missed deliveries.
Stack
Python 3.12
Django + Django REST Framework
PostgreSQL
Redis + Celery
Apache Kafka
Camunda / BPMN (orchestration)
Keycloak (OIDC)
OpenSearch
PyTorch / scikit-learn (forecasting)
React + TypeScript
Prometheus + Grafana
Want a similar architecture for you ?