Kvistlund Fund II closed in January 2024 at $47M. We've now been actively deploying for nearly two years, and it feels like the right time to put in writing what we learned from Fund I that shaped the Fund II thesis — and where we think the next layer of value creation in logistics AI is sitting.
This note is for founders who are building in the space and want to understand how we think. It's also a record for ourselves. Fund theses that aren't written down have a tendency to drift.
What Fund I taught us about the category
Fund I's eight investments spanned 2021 through mid-2023. The thesis going in was simple: AI would restructure European logistics from the software layer, and the founders who would win were operators who had built these systems from the inside. We deployed $35M across predictive logistics, freight visibility, shipping optimization, fleet intelligence, and logistics collaboration.
The thesis was correct in its core framing. Every investment we made in companies with deep operator founders has performed better than our expectations at the time of investment — not because operator founders are automatically better at building software, but because they are systematically better at identifying which problems are actually worth solving. The list of "logistics AI" problems that sound important in a pitch but turn out to have no real buyer is long. Operator founders have a filter for that list that was acquired through direct experience.
What we got wrong, or underweighted, in Fund I was integration complexity. Several of the companies we backed took significantly longer to reach production deployment than we anticipated because the integration layer — connecting their intelligent output back into legacy TMS, WMS, or ERP systems — was harder than the founders expected. The data ingestion side was usually manageable. The write-back side, where the model's recommendations needed to push decisions back into operational systems, was frequently where timelines extended. We learned to diligence this more carefully. For Fund II, integration architecture is a first-order diligence question, not a second-order one.
Where Fund II is focused
Fund II has a narrower target than Fund I. Rather than investing across the logistics AI stack from freight to warehouse to last mile, we've concentrated on three specific categories where we believe the value creation is front-loaded in this part of the cycle.
The first is decision science infrastructure. Logistics is a domain of continuous, interdependent decisions — pricing, routing, inventory positioning, labor allocation, supplier selection. The existing tooling for making those decisions is either spreadsheets, rule engines, or point solutions that optimize individual decisions in isolation. What's missing is infrastructure for modeling decision interdependencies — how a routing optimization decision interacts with an inventory positioning decision interacts with a carrier pricing decision. Nextmv, which we backed in late 2024, is building in this direction. Their core bet is that the companies who will win in logistics software are building on composable decision-science infrastructure rather than bespoke models for each decision type.
The second category is reverse logistics intelligence. Returns are structurally broken in European e-commerce. Return rates in apparel run 25–40%. The cost of processing a return — sorting, condition assessment, restocking or liquidation routing, carrier reconciliation — is frequently higher than the original fulfillment cost. The tooling for managing this is primitive: manual inspection, rules-based triage, and liquidation channels that capture a fraction of the goods' actual value. Covariant, which we backed in 2024, is applying AI to the warehouse robotics challenge — building intelligent robotic systems that can handle the physical manipulation, identification, and sorting of goods that currently requires expensive manual labor. Their system learns from operational data to improve pick-and-place accuracy over time, making the economics of automated returns processing viable at scale.
The third category is physical world instrumentation — sensor data and computer vision that turns previously dark operational events into supply chain intelligence. Greyparrot, which we backed in 2025, is the example here: computer vision on waste streams as a real-time signal for manufacturing and distribution performance. We wrote a longer note on the Greyparrot thesis separately. The general pattern is: physical events that have always occurred but were never captured digitally, now becoming available as model inputs because sensor cost and computer vision accuracy have crossed the threshold of economic viability.
What we're not doing in Fund II
We're deliberately not investing in general-purpose TMS or WMS software. The large incumbents have defended their positions more effectively than we expected going into Fund I, and the sales cycle and implementation cost for displacing a core operating system in a logistics company remains prohibitive for a seed-stage company. There are better entry points than attacking the core system directly.
We're also not investing in logistics AI companies that are building horizontally — targeting all industries, all supply chain types, all geographies simultaneously. The companies that have worked in our portfolio have had a specific, defensible beachhead. The companies that have struggled have been trying to be too many things to too many buyers before they'd proven the core value proposition in any one of them.
And we're not investing in companies where the AI is a feature rather than the core product. This was always part of our thesis, but Fund I reinforced it. A logistics software company that has added a prediction model to a workflow product is a different animal from a company whose core product is the prediction model and whose workflow elements exist to surface that model's output. The former is easier to sell early (the workflow value is obvious) and harder to defend long-term (the workflow commoditizes, and the prediction model is an add-on that can be replaced). The latter is harder to sell early (the value is less immediately tangible) and harder to displace once it's embedded in operations.
The check size and what it reflects
Fund II operates with a seed check size of $1M–$3M and pre-seed of $500K–$1.5M. The upper end of our seed range is higher than Fund I's, which reflects both the larger fund size and the fact that companies we're meeting now at seed have more infrastructure built — cleaner data pipelines, more mature model architectures, further-along design partner relationships — than the comparable stage company in 2021. The baseline expectation has shifted. We've adjusted our deployment accordingly.
We take a board seat at every seed investment. That's not changing. The board engagement is where we provide the most direct value — architecture reviews with Anders, commercial introductions through Maja's network, and the fundraising prep work that starts at month six. None of that is scalable at the portfolio concentration that would come from making smaller checks into more companies. We'd rather own our thesis deeply than spread thinly.