We’re a mid-sized biotech company that just wrapped a major overhaul of our demand planning system. For years we’d been running on spreadsheets and static models that couldn’t keep up with volatile demand signals across multiple product lines and geographies. Stockouts and excess inventory were both happening regularly, and our S&OP meetings felt more like finger-pointing sessions than planning. We finally bit the bullet and implemented a machine learning-based forecasting solution integrated with our ERP.
We took a sprint-based approach—six-week cycles, starting with a minimum viable product that had only standard forecasting functions. The goal was to simplify user understanding and testing before layering in complexity. After each sprint, we logged bugs, prioritized fixes, and gathered input from demand planners and supply chain ops on what features were must-haves for go-live versus backlog items. The AI system pulls in historical sales, pipeline data, market trends, competitor activity, and even social media signals to generate initial forecasts. Our planners review and adjust those forecasts based on their market knowledge, then we push a single consolidated forecast downstream to manufacturing and procurement.
Outcomes so far: forecast accuracy is up, inventory levels are more aligned with actual demand, and we’ve reduced safety stock without increasing stockout risk. The sprint model kept us on track and made sure the system actually fit our workflows instead of forcing us to adapt to rigid software. Biggest lesson learned: process design has to come before technology selection. We spent a lot of time upfront defining what planning activities could be automated and where human judgment was still critical. That foundation made all the difference.
This is a solid use case. One thing I’d add from our experience: make sure you define clear metrics upfront for what success looks like. Forecast accuracy is obvious, but also service levels, inventory turns, and cost metrics. Without baseline KPIs it’s hard to prove ROI and keep executive support when things get bumpy mid-implementation.
Six-week sprints is smart. We’re on S/4HANA too and looking at embedding ML forecasting into our IBP process. Did you integrate the ML models directly into the ERP or run them in a separate analytics layer and feed results back? Curious about your data architecture.
Good question. We went with a hybrid—ML models run on a separate analytics platform that ingests data from ERP, CRM, and external sources. Forecasts get written back into the planning module in near real-time. Keeps the ERP performance clean and gives us flexibility to iterate on the models without touching core system. On change management, we involved planners early in sprint reviews and made it clear the AI was a tool to assist them, not replace them. Transparency about what the model was doing helped a lot.
What data sources made the biggest difference in forecast accuracy? We’re debating whether to pull in social media and competitor data or just stick with historical sales and pipeline. Concerned about noise versus signal.
Appreciate the detail here. We’ve been stuck in analysis paralysis on vendor selection. How did you decide which planning vendor to go with? Did you prioritize fit with existing systems or best-in-class features?
This resonates. We tried a similar overhaul last year but didn’t do the sprint approach—went big bang and it was a mess. How did you handle change management with your planners? We had a lot of resistance from folks who didn’t trust the AI forecasts at first.