Fund I closed in March 2021. Seven months later, with the first three investments under our belt and active diligence running on a dozen more companies, I want to write down what we've observed about how European logistics operators actually engage with AI-native software — because a number of our pre-fund assumptions have been updated by direct contact with the market.
This is not a thesis revision. The structural argument — that European supply chain software is due for a rebuild from the ML layer up — is intact and in some ways more compelling than it looked in 2020. But the path from thesis to outcome runs through operator procurement behavior, implementation dynamics, and change management realities that are specific and worth being explicit about. We are investors who came from operations, and we try to hold ourselves to the same standard we apply to founders: be honest about what you're observing, update your models, write it down.
The integration burden is higher than we modeled
Before raising Fund I, I spent time mapping the technology stack of a typical mid-size European 3PL. On paper, the path from a raw carrier data stream to a trained ETA prediction model looked achievable for a competent ML engineering team in three to four months. In practice, we've watched this take twice as long at nearly every company we've spoken with, not because the engineering was harder than expected, but because the data pipelines connecting legacy WMS and TMS systems to cloud infrastructure are messier than any API documentation reflects.
European logistics operators have a wide distribution of technical debt. A Swedish 3PL that upgraded to a modern cloud TMS in 2019 is in a fundamentally different integration position than a comparable German operator running a 2011 SAP implementation with custom EDI connectors to 40 carriers. The AI product that works for the first operator is not automatically viable for the second. Founders who are building ML-native freight visibility or predictive demand products need to make explicit decisions about where on that technical-debt spectrum their early customers sit — and the ones who are honest about this in their go-to-market are building more durable early traction than the ones who assume they can serve everyone.
The corollary is that data quality standards vary enormously. We've seen carrier event data streams that are reliably timestamped and geographically precise, and we've seen carrier event data that is a collection of manually entered status codes with 6-hour update intervals and no location coordinates. Training a production ETA model on the second dataset produces something that underperforms naive heuristics. Founders who understand carrier data quality at the level of knowing which regional carriers are reliable API sources and which require data enrichment are doing something their less operationally grounded competitors are not.
Operations teams adopt differently than IT teams
A pattern that has shown up consistently across our portfolio companies and the companies we've diligenced is that the procurement journey for AI logistics software follows a split path depending on whether the initial champion is in operations or IT. Both paths can lead to a signed contract. They take different amounts of time and require different sales motions.
When the champion is an operations director — a head of transportation planning, a VP of supply chain — the adoption dynamic centers on whether the software produces better decisions than the current process. These buyers are outcome-oriented. They want to run a pilot on a real corridor with real carrier data and see whether the ETA predictions are more accurate than what their planners are currently producing. If the answer is yes by a meaningful margin, they will advocate internally for procurement. If the answer is marginal, they won't. The model quality is the story, and operations champions are good at evaluating model quality because they understand the problem domain.
When the champion is in IT or a CDO function, the conversation centers on integration architecture, data governance, and enterprise contract terms. These buyers are not evaluating the quality of the predictions — they're evaluating the security compliance documentation, the SLA structure, the implementation timeline, and whether the vendor has a credible enterprise support organization. The bar for a small AI startup to clear an enterprise IT procurement process is genuinely high, and it's particularly high in regulated industries. We've watched several promising companies spend six months in enterprise IT procurement cycles and come out the other side with a purchase order that was half the ARR they had expected, because the IT procurement process resized the scope.
We are not saying operations-led adoption is always better than IT-led adoption. Enterprise IT sponsorship creates organizational durability that an operations pilot doesn't. But the founders who understand which path they're on — and who calibrate their sales process, their proof-of-concept design, and their contract terms accordingly — are moving faster through their first ten customers than the ones who are treating all enterprise buyers as interchangeable.
Change management is part of the product problem
A consistent observation from our portfolio companies' implementations is that the human side of deploying ML-based logistics software accounts for more of the implementation timeline than the technical side. The average time from signed contract to first production model in active use has been longer than the average time from signed contract to technical integration complete.
This is a structural feature of deploying prediction-based software into organizations where predictions were previously produced by experienced humans. A freight planner who has been building carrier allocations by hand for eight years has a workflow, a set of implicit rules, and a professional identity that is organized around that task. Replacing part of that workflow with model recommendations requires that person to trust the model, understand when to override it, and feel that the new process is not making their job worse. The technical deployment is a prerequisite; the behavioral adoption is what produces the business outcome.
We've seen this addressed well and badly. The companies that address it well tend to have UX built around augmentation rather than replacement — the model makes a recommendation, the planner can see the inputs and confidence interval, and the planner retains clear authority over the final decision. The companies that address it badly tend to frame the product as automation of the planner's job, which triggers organizational resistance that no amount of product quality can overcome.
For seed-stage founders, the lesson is that your first customer success team hire matters as much as your second ML engineer hire. You need someone who can run a structured change management process inside a logistics operation and who understands that the human adoption problem is distinct from the integration problem. These are not the same person and they don't require the same skills.
The 2021 freight market created temporary tailwinds that won't persist
It is worth being explicit about something that has made our first year of investing easier than the next several years will likely be: the supply chain disruption of 2020-2021 made supply chain software a board-level priority in a way it normally is not. Logistics operators who would ordinarily take 18 months to evaluate new software and another 12 months to procurement approval were moving faster because the business case was urgent and visible.
Several of our portfolio companies have benefited from this. Sales cycles that would have been 9-12 months in normal conditions ran in 4-6 months. Budget decisions that would have required CFO sign-off were made by operations VPs. The pandemic-driven supply chain crisis functioned as an accelerant for the exact category of software we were backing.
That accelerant will slow. Freight rates have already begun normalizing in Q3 2021, and the operational urgency that made supply chain software procurement easy is subsiding. The companies that have used this window to build genuine customer relationships and demonstrate product quality will have a strong foundation when market conditions normalize. The ones that have grown primarily by riding the wave without building durable product value are exposed.
We are tracking this carefully across our portfolio. The companies we're most confident in are the ones where churn risk is low because the product is producing measurable outcomes that justify renewal — not the ones where growth was fast but the underlying product attachment isn't clear yet.
What this changes about how we evaluate new investments
Seven months into Fund I deployment, we have updated three things about our diligence framework. First, we are more explicit about asking founding teams to map their target customers' technical debt profile and demonstrate that they have a credible integration playbook for their actual customer segment. Second, we are paying more attention to go-to-market motion — whether the founder is targeting operations champions or IT champions, whether they have a proof-of-concept structure appropriate for each path, and whether they have the customer success capacity to handle behavioral adoption. Third, we are stress-testing financial models for a post-disruption market where enterprise sales cycles return to historical norms.
None of this changes the direction of the thesis. The European supply chain software stack is being rebuilt from the prediction layer. The best founders are the ones who came out of logistics operations and understand the problem at the level of real data and real user behavior. Those founders, building genuinely ML-native products for European operators, are the companies we want to back. That judgment is firmer in October 2021 than it was when we raised the fund.