Every investment memo we write starts with the same question: what is the durable structural advantage here, and why does this team have it? With Retrain.ai, both answers came quickly. The structural advantage is the skills intelligence graph — a representation of workforce capability that compounds over time as more enterprise data flows through it. The team has it because they understand both the ML side and the HR operations side with unusual precision. That combination is rare enough that we moved fast.
This is the write-up behind that decision. Not a pitch summary — more an account of what we learned during diligence and why it shaped how we think about workforce intelligence as a category.
Why Skills Forecasting Is Structurally Hard
Most enterprise HR tools treat skills as a taxonomy problem. You define a list of skills, tag employees against the list, and run reports on coverage gaps. The problem is that skill taxonomies decay faster than organizations can maintain them. The half-life of a specific technical skill description is probably 18-24 months — shorter in domains like machine learning and cloud infrastructure. By the time an enterprise's HRIS has a stable, agreed-upon taxonomy for a role, the role has already changed in ways the taxonomy doesn't capture.
Retrain.ai approaches this differently. Rather than asking companies to maintain a skills taxonomy, the system infers skills from signals that are already being generated: job postings, internal mobility data, learning completion records, project assignments, and external labor market data. The inference layer continuously reconciles what the company says it needs with what people are actually doing and where the external market is moving. The result is a skills graph that is self-updating rather than maintenance-dependent.
This matters because the actual use case that enterprise CHROs want — "will we have the capabilities we need for the business we are building in 24 months?" — requires dynamic rather than static data. A static taxonomy gets you an annual gap analysis. A dynamic graph gets you ongoing signals that inform workforce planning as a continuous process rather than a once-a-year exercise.
What the Diligence Process Revealed
We spent about six weeks on diligence. Jaakko led the technical review — he spent two sessions going through the inference pipeline architecture with the CTO, examining how the model handles the core challenge of semantic drift (the same job function described differently across time and across organizations). The approach uses a combination of embedding-based skill normalization and a hierarchical ontology that is updated from external labor market signals on a rolling basis. It is not a solved problem, but the team's approach to the unsolved parts was thoughtful and honest about the limitations.
On the market side, we spoke with heads of people analytics at four growing European enterprises — a Nordic financial services group, two German mid-market industrial firms, and a Dutch logistics operator. All four had the same foundational complaint: their current workforce planning process relied on subjective manager surveys and annual compensation benchmarking data, giving them almost no forward-looking capability. The question was not whether they wanted something better — they all did. The question was whether they trusted an AI-native system to be the infrastructure beneath a consequential business decision.
Two of the four said they would pilot but not commit until the product had an audit trail that showed HR and line managers where each forecast came from. That single piece of feedback shaped how we framed the product roadmap in our investment thesis.
The Defensibility Argument
Workflow software in HR is not particularly defensible. If you build a nice ATS, a competitor can build a nicer one and your customers can switch. The switching cost is low because the core asset — candidate data — sits in the employer's ATS, not in the vendor's model.
Skills intelligence is different. The more enterprise data that flows through Retrain.ai's graph, the better the normalization becomes and the more accurate the forecasts get. The graph itself — the accumulated record of how skills relate, how they evolve, how they cluster into capability profiles — becomes an asset that improves with use. A new entrant starting from scratch doesn't have that graph. They start with a static taxonomy, which is what the existing players already offer and what enterprises are already frustrated with.
We are not saying this moat is unassailable. A well-funded competitor with access to a large proprietary labor market dataset could build a comparable graph from a different starting point. But that starting point is not trivial, and the time it takes to accumulate data at enterprise scale is measured in years, not quarters. For a seed-stage investment, that runway matters.
What We Are Watching Post-Investment
The risk we underwrite most carefully is the enterprise sales cycle. Skills intelligence is a new category that requires CHROs to believe two things simultaneously: that AI can infer accurate capability profiles without a manual taxonomy, and that those profiles are reliable enough to base consequential workforce decisions on. Getting both beliefs to coexist in the same buying committee requires a proof point that is specific and auditable.
We have been working with the team on a pilot structure that reduces initial commitment by scoping the first deployment to a single business unit rather than org-wide. A 300-person engineering org at a growing Nordic fintech ran the first structured pilot in early 2025. The output — a capability forecast showing projected skill gaps across four engineering functions over an 18-month horizon — was the first time the company's head of engineering had seen that kind of forward-looking data derived from signals other than his own intuition. That specificity is what makes the sales case.
We are watching customer expansion rate within accounts, forecast accuracy measured against actual attrition and hiring patterns, and the speed at which the sales motion transitions from champion-led to committee-led. The last one matters most: if the product stays dependent on a single internal champion to survive procurement, it is fragile. If the output is auditable and trusted by HR leadership independent of the champion, it has a different growth profile.
Why This Team
The founding team has an unusual combination of credentials. The CEO came from a workforce planning function at a large European logistics group — she has been the buyer of this kind of product, repeatedly frustrated by what the market offered, and built the founding thesis from that frustration. The CTO built the NLP infrastructure layer at a Nordic labor market platform before joining. Together, they understand the problem from the demand side and the ML architecture from the supply side, without needing a bridge between the two.
That combination is not common in B2B HR tech. Most teams are either strong on product and weak on ML (they build workflow-first, AI-later) or strong on ML and weak on enterprise distribution (they build technically interesting systems that procurement processes cannot handle). The Retrain.ai team occupies the middle ground by virtue of experience rather than strategy, which is a more durable position to be in.
The $2M Seed round Sammalkko led in 2023 was sized to get the product from concept to a structured set of enterprise pilots with data. We are now in the stage that matters — watching whether the thesis about compounding graph value and enterprise trust holds up against the friction of real procurement and real deployments. The early signals are positive. The skepticism we maintain is proportionate to how early it still is.