Four investments into the portfolio we are building with Fund I, and the pattern that connects them is clearer to us now than it was in the original investment memos. We backed Parloa for conversational AI in enterprise contact centres, Workist for document processing automation in procurement, Candis for AI bookkeeping automation for small and medium enterprises, and Kenjo for HR workflow automation. These are four different verticals, four different go-to-market motions, and four teams with different technical and commercial compositions. But they share a structural characteristic that we have come to think of as the defining feature of the category we want to back: they are all automating workflows that create significant operational drag precisely because they are too variable for rule-based systems but too repetitive and high-volume for humans to handle without error accumulation.
Parloa is addressing a problem that every enterprise with a meaningful inbound call volume faces: the contact centre is expensive, the quality of service is inconsistent, and the best agents eventually leave for better opportunities. What makes the problem hard is that customer queries are expressed in natural language, often ambiguously, sometimes in multiple languages, and the system needs to handle the full distribution — not just the common cases that a scripted IVR can manage. The conversational AI architecture that Parloa is building needs to handle intent recognition, dialogue management, and backend system integration in a single coherent flow. That requires trained models, not decision trees. The enterprise buyer for this product is a head of customer operations or a CTO, and the purchase decision is driven by operational cost pressure rather than innovation appetite — which means the sales dynamic is fundamentally different from a new-capability SaaS pitch.
Workist and Candis are both in the document-heavy automation space, but for very different buyer profiles and document types. Workist processes purchase orders, delivery notes, and invoices flowing through B2B procurement at manufacturing and distribution companies — structured but highly variable documents where the field extraction problem is genuinely difficult at scale. Candis targets the SME bookkeeping layer, where a tax adviser or in-house finance manager is manually categorising bank transactions and matching them to receipts. In both cases the core value proposition is the same: replacing a high-volume, low-skill cognitive task that is too variable for rules and too tedious for humans to do at consistent quality. The difference is in the buyer and deployment context. Workist is sold into operations managers at Mittelstand firms; Candis sells through tax adviser networks, which changes the sales architecture entirely.
Kenjo is the most conventional-looking of the four from a product perspective — HR software is a mature category — but the angle is not conventional. They are not building an HRIS or a payroll system; they are building the workflow layer that sits between what an HR platform records and what a manager actually needs to do with that information: onboarding checklists, probation period follow-ups, absence management workflows that need to comply with different national labour regulations across a workforce that spans Germany, Spain, and the UK. The AI element is in the adaptive workflow intelligence — surfacing the right task to the right person at the right time based on employment context. It is a narrower thesis than building a full HR suite, but the workflow layer is exactly where AI-native architecture creates value that static form-builders do not.
What we learned from building conviction on these four: the common thread is not the vertical or the buyer type but the shape of the problem. High-volume unstructured data, variable but bounded domains, existing manual processes that are clearly costing money, and buyers who care about reliability more than novelty. We expect to keep backing this problem shape in new verticals over the life of Fund I, and we expect the pattern to hold: the initial commercial risk is always whether the product is reliable enough to displace the existing process, not whether the process is worth automating in the first place.