How Pillar AI’s River probabilistic planning software transformed inventory performance across all SKUs and distribution centers in a South American country.
Like most large consumer goods companies, our customer (a global beverage leader operating across dozens of distribution centers and hundreds of SKUs in a South American country) faced a persistent planning tension: hold too much inventory and you tie up working capital; hold too little and you lose sales and miss service targets.
In order to stay in balance, they run a relatively mature planning process focused on managing inventory corridors (min, max, target)
Despite this, in various geographies they struggle with a persistently high % of inventories falling outside the corridor each week.
At any given week, roughly 35% of SKU-location combinations held inventory outside their target range—either excess tying up cash or dangerously low risking out-of-stocks.
Pillar AI deployed River: a probabilistic simulation platform, in a 3-month pilot covering all finished-goods SKUs and DCs across the country.
River predicts the full probability distribution of demand and supply, and then runs a simulation to find the odds of inventory falling outside the corridor. It then runs an optimization to recommends the smartest corrective actions before problems occur.
Transformer-based AI models generate calibrated probability distributions for both demand and supply, capturing the full range of real-world outcomes, not just a single point estimate.
A “digital twin” of the supply chain runs 1,000+ simulations to play out possible inventory positions, identifying the odds of falling outside the corridors.
Optimization leveraging the odds from the simulation, identifies ~1,000 recommended actions per week to improve Expected Value of inventory positions.
River is designed from the ground up to avoid disruption to people and processes. It operates in the background, reading from existing systems and generating recommendations without replacing or modifying any part of the current process.
Planners continue to work exactly as they do today. There is no new interface to learn, no change to how supply orders are reviewed or approved, and no dependency on River being in the critical path.
River simply updates the plan to mitigate value loss.
Planners can review the updates, or not.
Planners can access supplemental information such as probability distributions, or not.
Planning remains managed entirely within the existing system. River runs in the background, layering probabilistic intelligence on top of the current process without requiring any changes to how planners operate day-to-day.
River's UI is entirely optional. Most customers do not want any additional interface to adopt or maintain.
The customer's planners spent years learning their planning system, and updating their processes to match. All integration work was focused on how River could push into the existing system.
The customer wanted to evaluate River's recommendations before going into production. They ran River's corrected plan as a second scenario within the planning system itself for direct head-to-head comparison and business case evaluation.
Together with the customer, we ran 6x back-test simulations spanning May–June 2025, covering all products and distribution centers in the country. Results were consistent across every run. The table below shows both single-country pilot results and projected savings at global scale.
| Metric | Single Country (6-Run Avg.) | Global Scale (Projected) |
|---|---|---|
| Out-of-Corridor Inventory Improvement | 5-15% (35% out to 20-25%) | 5-15% |
| Annualized Cash Savings | $18M | $100–200M |
| Annualized Operating Savings | $9M | $50–100M |
For every correction that misses the mark, four to five succeed in keeping inventory within range, reducing excess, or preventing a stock-out.
Probabilistic planning gives us something we’ve never had before: a clear view of the odds. Instead of reacting to inventory exceptions, we can get ahead of them. The pilot results show that leading with probability, rather than a single plan number, can unlock tens of millions in savings.
Jeff Alpert, CEO — Pillar AI
Though River is not a demand forecasting tool, improved demand prediction is a side benefit. River’s transformer-based AI demand model outperformed the existing planning system, reducing Mean Absolute Percentage Error (MAPE) by 7 percentage points.
By reducing excess inventory while also preventing stock-outs, Pillar delivered a dual benefit: $10–20M+ less working capital tied up in inventory, without sacrificing customer service levels.