A CRM agent for a distributor: who to call and what to offer
A manager opens the account and immediately sees orders, debt, the risk of losing the customer and what to offer next. The agent assembles that from the CRM, 1C and the correspondence. Qualified lead-to-deal conversion rose 8–14%.
What changed for the sales team
Customer card, pipeline and tasks; a lead score and the share of won deals by stage.
A hint of which item and promotion to offer; reports by manager and region; an alert on debt and a silent deal.
A short read of the negotiations and answers from the history, inside the employee’s permissions.
What the numbers show
Qualified lead-to-deal +8–14%; key-account churn −6–10% on the pilot.
Time spent on the sales lead’s report −35–50%.
Highlights
- One card holds orders, payments, debt, tickets and meetings — no hunting across five windows.
- The agent suggests win likelihood, the risk of losing the account and which item to offer this season.
- A sales lead sees conversion and each manager’s contribution without waiting for a Monday pack.
- Data arrives from the accounting system, phone and email; the negotiation history can be asked in plain language.
Stack
Python 3.12
FastAPI
PostgreSQL
ClickHouse
Apache Airflow
dbt
Redis
Qdrant (vector search)
LangChain / LLM API (RAG)
scikit-learn / XGBoost
React + TypeScript
Metabase / Grafana
Want a similar architecture for you ?