Most of the warehouse automation conversation has been about hardware. AMRs, goods-to-person systems, automated storage and retrieval — the robotics story has dominated the past decade of warehouse investment. That framing made sense when the bottleneck was physical throughput. But the bottleneck has shifted, and the industry hasn't fully caught up with the implications.
The new constraint isn't move speed. It's decision quality. And decision quality is a software problem.
What automation actually gave us — and what it didn't
The generation of warehouse automation that started scaling around 2016–2020 solved a real problem: it increased throughput per square meter and reduced the reliance on peak-season labor pools that were becoming structurally unreliable. A 3PL running a goods-to-person system in a 25,000 sqm facility can process two to three times the order volume of a comparable manual facility at steady state.
What it didn't solve: the intelligence layer. Automated systems are extremely good at executing decisions. They are almost entirely dependent on external systems — WMS, OMS, slotting optimization rules — to tell them which decisions to execute. The robots move fast. The orchestration logic still runs on rule engines built in 2008.
Here's what that looks like in practice. A growing FMCG brand running a mid-scale 3PL operation — let's say 8,000 SKUs, 30,000 orders a day at peak — has their goods-to-person system running at utilization rates that look fine on average. But the rule engine governing putaway and slotting is optimized for their SKU mix as it existed 18 months ago, when they set it up. Their top-velocity SKUs have changed. New product lines have different dimensional profiles. The rule engine doesn't know this. Nobody has touched the slotting logic because it "works" and the person who knew how to tune it left. The automation is doing exactly what it's told. It's just being told the wrong thing.
The intelligence gap is structural
Traditional WMS vendors have not been motivated to solve this. Their business model is the licensing fee on the WMS itself — complex, highly configured implementations that take 12 months to go live and create switching costs that last a decade. Embedding adaptive ML into the core logic would mean: (a) acknowledging that the rule-based configurations they sold you are suboptimal, and (b) introducing model behavior that's harder to support, document, and blame-assign when something goes wrong. Neither of those outcomes is in their interest.
The result is that warehouse intelligence has evolved as a layer above the WMS, not inside it. Slotting optimization engines, demand-driven replenishment models, labor productivity forecasting — these have all emerged as standalone software products that sit between the WMS and the automation layer, pulling operational data and pushing back better decisions.
We think this is the right architecture. Not because integrating intelligence into WMS is technically impossible, but because the replacement cycle for WMS is a decade and the replacement cycle for a decision intelligence layer is much shorter. A warehouse operator can deploy a slotting optimization module without migrating their WMS. The land-and-expand motion is cleaner, the integration surface is smaller, and the value is demonstrable in 90 days rather than 18 months.
What "cognition" means in this context
When I use the word cognition, I mean specifically: the ability to update the decision logic based on outcomes observed in the same system. This is distinct from automation, which executes fixed decisions, and from analytics, which describes past decisions. A cognitively capable warehouse system notices that SKU velocity for a particular product category spikes every Thursday between 2 PM and 6 PM, correlates that with order placement patterns upstream, and adjusts pre-positioning logic before the next Thursday — without a human writing a rule.
The technical components needed to do this well are: a clean operational data stream (scan events, order release patterns, equipment utilization logs), a model that can extract signal from that stream on a rolling basis, and a feedback loop that validates whether the decisions the model recommended actually improved throughput or reduced exception rates. None of this is research-level ML. It's engineering discipline applied to a domain that has historically been resistant to it.
Kargo, which we backed in 2025, is building in this direction. Their core product sits between WMS and automation orchestration, building a continuous model of intraday demand patterns and equipment state, and adjusting wave release and slotting recommendations dynamically. In retrospective testing against live operational data from their design partners, the throughput improvement from dynamic wave release versus static 4-hour wave windows was in the 12–18% range in high-variability environments. That's not a rounding error for an operator whose labor costs are their second-largest line item.
The cases where software intelligence is not enough
We're not saying that all warehouse problems are software problems. There are real physical constraints — dock doors, conveyor topology, ceiling height — that software can optimize around but cannot eliminate. A warehouse designed for palletized bulk storage cannot be made to efficiently handle each-pick e-commerce fulfillment through software alone, regardless of how good the WMS intelligence layer is. Physical architecture constrains the possibility space.
The honest framing: software intelligence maximizes utilization of existing physical infrastructure and makes the humans and machines operating in it more effective. It does not replace the capital decision about whether to build a mezzanine, add a sorter, or reconfigure dock doors. Those decisions benefit from intelligence-layer data — it should inform whether you need the capex — but the physical investment remains a separate question.
Where we're looking
The companies building in warehouse intelligence that we find most interesting in 2025 share a few traits. First, they have genuine operational data access — meaning they have design partners who are giving them live operational logs, not just retrospective data exports. Second, they are building the feedback loop, not just the model — there's someone on the technical team who has thought hard about how to validate that the recommendations the model makes are actually improving the metrics that matter. Third, they have a wedge that doesn't require displacing the WMS on day one — the integration surface is a specific workflow or decision type where the improvement is measurable in weeks, not quarters.
The category is still early. Most warehouse operators — outside of the top-tier 3PLs and the largest in-house e-commerce operations — are still running on rule engines and gut instinct. The intelligence layer hasn't penetrated mid-market warehouse operations in Europe at any scale. That's a large addressable base for companies that can demonstrate value without requiring a rip-and-replace of existing infrastructure.