Every warehouse has a classification problem it doesn't talk about enough.
High-risk SKUs get grouped with adjacent categories. The same item carries a different risk tag depending on which fulfillment center it sits in. Manual review cycles lag behind category changes, promotions, and assortment updates — so risk labels go stale. The result? Shrink exposure grows quietly, and intervention comes too late.
This isn't a data problem. It's an orchestration problem.
The real issue
Classification logic today is fragmented — static lists in one system, dynamic tags in another, tribal knowledge in a third. Each warehouse applies its own rules under time pressure, creating policy drift and inconsistent controls across the network. When an issue is identified or categories overlap, the time to reclassify is measured in weeks, not minutes.
The shift
What if classification wasn't a periodic review — but a continuous, intelligent process?
That's the idea behind an agentic approach to SKU risk classification. A semi-autonomous orchestrator — a planner agent — takes a structured SKU signal, breaks down the scope, and delegates to parallel specialist agents. Each agent evaluates the item through a different lens: industry frameworks like CRAVED and CAPTURED, internal rules and policies wrapped as agents, and organization-specific classification logic.
The outputs are synthesized through weighted signals into a single Risk Indicator Value (RIV), with configurable thresholds for flagging.
Built in — but only when policy gaps, ambiguity, or escalation requires it. Not every run.
What this unlocks
The same framework extends beyond baseline shrink patterns. Here's what the same agentic architecture can address:
Here's what makes it powerful: an agent can spot that two individually "safe" chemicals, when stored on the same shelf, create a fire risk. That's not something a static list catches.
Why this matters
Near real-time flagging. Consistent policy execution across every site. Multi-model triangulation that strengthens signal quality. Configurable weighting that adapts to the business. And human oversight exactly where it's needed — not everywhere.
This entire solution — from problem framing to working prototype — was built in hours, leveraging AI as a co-pilot. The architecture is model-neutral — works with any foundation model, keeping the focus on the orchestration logic, not vendor lock-in.
That's the mindset shift: spot the problem, prototype it, and bring something tangible to the table. The faster the invisible becomes visible, the faster stakeholders trust the direction and decisions move forward.