A German version of this article is available on Benedikt's LinkedIn.

Three months ago, my colleagues and I set out four investment criteria we're now weighting more heavily given the AI opportunity: product-led leadership, deep product logic, critical outcomes, and proprietary data.

Since then, I've met with dozens of founders and management teams, mostly in Germany and the DACH region, as we look at new software investments.

There is a noticeably confident mood among the founders and software leaders I’ve met, and nothing in those conversations has shifted my view on the four criteria. What has changed is how we test for them.

The signal for product-led leadership isn't about AI

Ask a founder to walk through a typical day for their core user and you learn more than any AI feature walkthrough.

The strongest product leaders can describe exactly how a specific user spends their day inside the software, edge cases included. There is a very clear understanding what is the “job to be done” that the company has designed. When someone speaks the customer's language fluently, adoption of whatever they build next tends to come easily.

The second signal is where AI actually sits in the business. In the best companies, it sits inside the core roadmap, discussed the same way as any other feature.

Then there's pace and commercial impact. Across a number of our portfolio , we have seen product build times compress from around nine months to under three, and some of the earliest movers are reporting 10% or more of new bookings from AI products.

We are always looking for founders who can move at pace with a clear sense of which products are getting traction, which have been killed, and why.

Testing "deep and hard to replicate"

Almost every software leader tells me their product logic is deep and hard to copy. To assess this, it is often eye opening to e.g. walk through what happened to the product when the last regulatory change came into force. How did this get into the product, and how long did it take?

This question works because it tests three things at once:

Depth. Does the company actually encode the rule, or does it hand the customer or an implementation partner a configurable field?

Evolution. Is someone inside the business tracking changes continuously, or does the product only catch up once customers complain?

Speed. How quickly does the team turn its interpretation of a new rule into shipped, tested logic?

In Germany, the phased e-invoicing mandate is a good live example. A strong answer gets specific quickly. It covers which formats they support, how they handled transitional rules and edge cases like credit notes or cross-border invoices, and which cases still go to a human for sign-off and why. It usually also names the people involved, typically in-house tax or payroll experts who work side by side with product managers. A weak answer stays general: "we have a flexible rules engine" or "our partners configure that." That usually means the domain knowledge sits with consultants or customers, not in the product.

Another useful topic on product logic is to discuss what happens if foundation models get better – does that make someone excited or nervous? A model that reasons better still doesn't know how a specific regulation applies to a specific client, or which exceptions need escalating. That knowledge sits in the product, built up over years of rule changes and customer edge cases. A business with genuinely deep vertical logic should therefore be excited, because every model improvement makes its encoded knowledge more usable and gives it more to build on. If the honest answer is "worried", the value probably sat in the interface or in aggregating information, which is exactly what better models commoditise.

Founders overstate proprietary data more than any other claim

This is where I push back hardest, and where the most interesting conversations happen.

The most valuable proprietary data is generated by customers using the platform, but software execs sometimes confuse this with data they simply hold or have assembled. Customer-created data includes things like transaction records, exception logs, customer logic on decision making. What makes it valuable is context. A number on its own is easy to replicate. A number linked to the action a user took, the outcome that followed and the reasoning in between can't be rebuilt anywhere else.

Scopevisio, which provides cloud ERP software for the Mittelstand, is a good example from the Hg portfolio. Its ERP has lots of context, because all the financial and operational data sit in one database, so the system sees how a transaction connects to the order, contract, invoice, approval and payment behind it. Every answer is traceable to the underlying records and comes with an explicit confidence level, and after each interaction the user is asked how useful it was. This is not something that a generic model reading exported data could do. Customers are not asking for proprietary data, but they want a solution they can check. That's why the key questions for software leaders are how they build trust with users on their AI journey and whether the product gets better because customers use it.

Germany's strength is real, but with areas to improve

On deep product logic and critical outcomes, German software companies are structurally strong. Payroll, tax, accounting and compliance here are rule-dense and change constantly. Accuracy has never been optional.

Diamant is an interesting example from our portfolio: It is one of the only vendors that has focussed on just one product over decades. An organisation that lives and breathes accounting. Through on-prem, cloud and AI technologies, they have been the trusted partner of the German Mittelstand navigating local regulatory complexities and customer challenges year after year. Now keeping the same product ethos in an agentic age of finance.

The most common blind spot is conflating strong engineering capability with product-led leadership. Plenty of businesses I meet have strong engineering teams but an under-invested product function, and that holds back both their speed and ambition.

The other is proprietary data. The raw material is often excellent, made up of regulatory submissions and Mittelstand transaction histories, but usability sometimes lags because cloud migration came late, and the on-prem installed base is large.

The advantage is real and the AI opportunity is enormous. Incumbents’ ability to capture it depends on correcting these blind spots. The window won't stay open by default.

The views and opinions expressed in this blog are the author’s and should not be taken to represent the views or positions of Hg or its affiliates. This blog is for general information only. It is not investment advice, a recommendation, an offer or solicitation to buy or sell interests in an Hg fund, or an invitation or inducement to invest. Past performance is not a reliable indicator of future results and is not indicative of the performance of any Hg fund or investment. Statements contained in this blog are based on current expectations or estimates and are subject to a number of risks and uncertainties. Actual results, performance, prospects or opportunities could differ materially from those expressed in or implied by these statements and you should not place any undue reliance on these statements.