Multi-Agent Intelligence for Warehouse Operations
Product Manager — WMS, FoodsUp Inc.
Built a specialized AI agent team that continuously monitors warehouse KPIs across inbound, outbound, inventory, labor, and cost — transforming reactive dashboard-checking into proactive, AI-driven operational decision-making.
Problem Discovery
Dashboards and BI tools generate charts and numbers, but translating a dip in pick accuracy into a specific corrective action requires experience, context, and time that managers don’t have mid-shift.
By the time a manager notices a KPI anomaly through manual dashboard checks, the impact has already cascaded — a missed put-away deadline becomes a stock-out, which becomes a late shipment.
Existing alerting tools can tell you “cycle count accuracy dropped to 92%” but cannot reason about why it dropped or what to do about it — the diagnosis and prescription step is entirely manual.
Inbound, outbound, inventory, labor, and cost KPIs are tracked in separate dashboards by different people — nobody has a cross-domain view.
User Research
Oversees daily operations across inbound, outbound, and inventory at FoodsUp’s Toronto and Montreal sites. Responsible for hitting KPI targets but spends the first hour of each shift manually reviewing dashboards.
Frontline shift lead managing 10–15 associates. Receives tasks from the manager but lacks visibility into why priorities change mid-shift.
Senior leadership overseeing multi-site operations. Needs weekly/monthly KPI trends for strategic decisions.
Strategic Bet
“A team of specialized AI agents — each owning a KPI domain — will outperform a single monolithic dashboard because agents can reason about their domain, detect anomalies in context, and prescribe specific actions, not just surface data.”
Positioning: Traditional BI tools visualize but need human interpretation. WMS vendor alerting uses simple thresholds. WarehouseIQ adds an intelligence layer: anomaly → diagnosis → action plan.
Why Now: AI agent frameworks matured + FoodsUp’s data instrumentation layer already built.
✓ What We Chose
Deeper domain reasoning over general-purpose simplicity.
Keep humans in the loop during pilot, build trust before automation.
✕ What We Said No To
Managers need to trust recommendations before handing over control.
Integrates into existing WMS rather than replacing tooling.
Solution
Used impact/effort scoring. Started with outbound (highest business impact) and inventory (most data-rich), then expanded.
Each agent specializes in one KPI domain with its own monitoring rules, anomaly thresholds, and action playbooks. Mirrors how warehouse teams are organized.
Real-time dashboard alerts for urgency. Conversational interface for depth — managers ask “why?” and get explanations.
Every anomaly comes with: what happened, why, what to do, and expected impact. Closes the gap between “wrong” and “fix.”
Achievements
Established the pattern for AI-assisted operations across the company.
Zero new instrumentation required — leveraged Tableau/Metabase infrastructure.
Architecture designed for Toronto and Montreal from day one.
Shift from skepticism to active engagement during pilot.
Reflection
Every analytics investment is a future AI investment. The Tableau/Metabase dashboards built earlier became the foundation for WarehouseIQ.
Splitting into domain specialists dramatically improved recommendation quality. Domain expertise matters even for AI.
The conversational interface was critical. When managers could ask “why?” and get a clear explanation, adoption followed.
Involve warehouse supervisors earlier. They execute the action plans and their feedback would have improved agent playbooks faster.