What has actually changed, and how it was measured.
Two deployments, both live. Every figure below is a before-and-after on a named process, with the measurement method stated. Where a customer is not named, it is because the name is not yet cleared for external use — not because the deployment is hypothetical.
Regional demand and pricing forecast cycle
JBSProfitability increase in Year 1, coordinated agent system
JBSSKUs price-simulated across markets in minutes
JBSMaster production scheduling cycle
Industrial manufacturerA 28-hour forecast cycle now takes under an hour.
The company
JBS — one of Brazil’s largest beef producers. Operations spanning multiple countries, 65 production facilities, 17 sales channels, more than 150,000 customers, and tens of thousands of SKUs against rising volatility in demand, pricing and logistics.
Before
Pricing analysts spent hours manually extracting data into Excel, running calculations offline, and distributing static recommendations that were frequently outdated by the time they reached the sales teams. Reacting to a change in the data could take days.
What A2go deployed
Agentic applications for price optimization, demand forecasting and inventory allocation, governed by the customer’s own business rules and running on data integrated from warehouse management, ERP and CRM systems plus external feeds — weather, market movements, feed costs. Built on the Databricks Data Intelligence Platform, governed through Unity Catalog.
After
- Regional demand and pricing forecasts: 28 hours to under one hour
- Pricing scenarios across 60,000+ SKUs simulated in minutes rather than days
- At least 25 users across market intelligence now run complex simulations daily and concurrently, in a governed environment, without data engineering support
- A centralized, batch-driven process became a self-service capability across distributed teams
- $200 million profitability increase in Year 1 from the coordinated agent system running across demand, production, S&OP and pricing
Read the A2go–Databricks customer story →
Measurement: elapsed wall-clock time for the named forecast cycle, before and after, on the customer’s own instrumentation. Published as a customer story by Databricks, December 2025.
| Production facilities | 65 |
| Sales channels | 17 |
| Customers | 150,000+ |
| SKUs modeled | 60,000+ |
| Users, daily and concurrent | At least 25 |
| Source systems | WMS, ERP, CRM, external feeds |
This deployment is documented publicly by Databricks. Read their write-up.
Master scheduling: 18 hours across 24 spreadsheets, to one 15-minute run.
The company
A roughly $500M industrial manufacturer and distributor. Multi-plant, shared supply and shared customers across sites, with a master production schedule rebuilt by hand every cycle.
Before
Twenty-four spreadsheets, maintained by different people, reconciled into a single master schedule over roughly 18 hours of planner time per cycle. Any material change after the run meant starting again, so in practice the schedule was frozen well before it should have been.
After
- Master production scheduling cycle: 18 hours to 15 minutes
- Rescheduling became something the team does when conditions change, rather than once a cycle
Measurement: planner hours logged against the scheduling cycle before and after; inventory and cash figures from the customer’s own finance reporting over the deployment period.
This customer’s name is pending clearance for external use. Reference calls can be arranged under NDA — ask during a working session.
Three agents carried this decision, in the order they run.
- Master Production Scheduling — sequences across sites that share supply, capacity and customers
- Lead-Time and Safety-Stock Optimization — re-evaluates what has to be on hand against the schedule that will actually run
- Promise-Date Jeopardy — returns a commitment that cannot be held as a constraint on the plan, rather than as an escalation to a person
Neither one started with a migration.
Existing systems stayed
No ERP replacement, no data warehouse consolidation, no master-data program completed first. ADIP read from what was already there.
One decision first
Each began with a single decision the business could name a number against, not a platform rollout. Expansion followed the result.
A person kept the call
Recommendations arrived with the tradeoffs and the expected impact attached. Approvals and overrides both fed back into the decision memory.
Ask for the number that matters to you.
Bring your own process and we will tell you, in one session, whether it looks like these two and what a measured first deployment would target.