Building a Supply-Chain AI Coworker with Databricks Genie One
For Brightbasket Market, a regional grocery chain, I wanted the supply-chain lead to go into the weekly supplier review with a clear call list. Which suppliers needed a delivery follow-up, and where did the retailer and supplier need to reconcile their demand plans?
Preparing that list requires information from purchasing, inventory, and demand planning. I built a coworker with Databricks Genie One that answers from those records and shows the sources behind each number.
How I’d prepare for the supplier review
Supplier 00014 would be my first follow-up. In the February 14, 2026 snapshot, one store was out of Product 000044 and still waiting for its 224-unit order, due on February 13. I’d want a revised delivery commitment for that order.
For Supplier 00005, I’d bring demand planning into the review. Brightbasket planned 445 units of Product 000103 for fiscal week 8, compared with the supplier’s 360. I’d want to understand the assumptions behind that 85-unit difference. The product was already out of stock in three stores, and the supplier also had overdue orders. Those orders needed reviewing, but the data didn’t link them to the stockouts.
How I connected the records
The Open Retail Data Model supplied the structure for purchasing, inventory, suppliers, and demand plans. I deployed the data in Unity Catalog, where access to the tables and metrics is governed.
I set up 12 metric views to define measures the dashboard and Genie One could both query. For the supplier review, mv_supplier_review combines overdue orders, store availability, and upcoming demand-plan disagreements in one row per supplier.
mv_supplier_review summarizes each source by supplier before the join, so the dashboard and Genie One read one row per supplier.I summarized each source by supplier before joining it to the supplier list. Joining the detailed order, inventory, and planning records directly could multiply rows and inflate the totals.
The detailed records remain available for follow-up. For Supplier 00014, those records show the stockout and overdue order for the same product at the same store.
I kept a dashboard as the reference for the supplier review. Genie One supports follow-up questions using the same measures and the underlying records.
How I organized the data by team
I organized the data and Genie Agents by business capability rather than by the source model’s package folders. The Brightbasket Supply Chain and Logistics domain has three subdomains:
- Supplier Collaboration brings together delivery performance, supplier risk, and joint demand plans.
- Replenishment and Inventory Management covers stock availability, replenishment, and inventory measures.
- Demand Planning and Forecasting covers forecast accuracy and planner adjustments.
Shared supplier, product, store, and fiscal-calendar dimensions sit in the parent domain. I used governed tags to associate each asset with its domain or subdomain.
I built one Genie Agent for each subdomain, selecting the relevant tables and metric views and adding SQL examples and business instructions.
Get Saurabh Shukla’s stories in your inbox
Join Medium for free to get updates from this writer.
The Supplier Collaboration Agent, for example, includes a query that matches overdue orders to out-of-stock products at the same store. Its instructions anchor relative dates to the business snapshot and distinguish purchase-order delays from shipment delays.
In my test runs, Genie One retrieved an Agent’s SQL examples and instructions and executed its own queries.
I wrote 11 Pages containing business definitions, examples, and links to related assets. The Plan disagreement Page explains when a difference between the two demand plans warrants attention: more than 20% of the supplier’s planned quantity. The 85-unit gap for Supplier 00005 exceeds that threshold.
Pages contribute explicit business context to Genie Ontology, which Genie One can retrieve when answering. Pages are currently a Beta feature.
I added workspace instructions to identify Brightbasket’s assets and anchor relative dates to the business snapshot. They also specify that a question about future demand should not inherit the date filter from an earlier question about last month’s supplier performance.
How Genie One answered the question
My test runs showed how Genie One used that context. It retrieved a past conversation, four Pages, and relevant data assets, then checked the columns in mv_supplier_review.
It ran three queries in parallel: one for suppliers with overdue orders, one for out-of-stock products with outstanding orders, and one for demand-plan disagreements. A follow-up query checked the joint demand plans for three out-of-stock products. The answer cited its sources and used the February 14, 2026 business snapshot.
How I used the same context in ChatGPT Desktop
I connected ChatGPT Desktop to the Genie One Model Context Protocol (MCP) server. From a Codex chat in ChatGPT Desktop, I could ask the same supplier-review question. Genie One answered using Brightbasket’s data and business context, with the signed-in user’s permissions.
In my tests, I asked ChatGPT to send the question to Genie One word for word. That kept it from adding its own dates or definitions. The returned answer still needs checking against the supporting records.
How I made the review a weekly routine
I saved the supplier-review steps as a Genie One skill, brightbasket-weekly-supplier-review. A skill is a set of instructions that Genie One loads when a request matches it, or when you type / and its name.
Then I turned the finished analysis into a scheduled task, which Genie One lists under Schedules. Genie One suggested what the task should do and when it should run. I set it for every Monday at 7:00 AM, with instructions that call the skill. Each run posts the review in a chat thread and sends it by email.
On Monday, the supply-chain lead can call Supplier 00014 about the 224-unit order that was due on February 13, and Supplier 00005 about the 85-unit difference in the fiscal week 8 plan. Before each call, the lead can open the order, inventory record, or demand plan behind it.
Data note: Brightbasket Market and its operating figures are demonstration data generated from the Open Retail Data Model. No customer or real business results are described.
