This week on the EUVC Podcast, Andreas Munk Holm sits down with Mikael Johnsson, Co-Founder and General Partner at Oxx, one of Europe’s leading specialist B2B software investors.
Mikael has a very clear-eyed view on the current AI wave: he’s seeing valuation discipline slip, fundamentals being stretched, and a real risk that the market mistakes pilot-driven excitement for lasting enterprise value.
In this episode, they explore how to tell hype from substance, what “real” AI adoption looks like inside a business process, and how both founders and investors can keep their heads when everyone else is losing theirs.
🎧 Listen on Apple Podcasts and Spotify
🎯 This Episode’s Themes
Pilot vs Production: Why 92% of “AI adoption” is still pilots, and what real enterprise deployment looks like.
Quality of Revenue: How to distinguish experimentation from durable, expanding usage and ROI.
Application Layer Reality: Why it’s fine to build on public LLMs—if you own the workflow, not the model.
Moats in AI: Orchestrating business processes and owning transactional data vs relying on “speed of execution”.
Valuations & Vintage Risk: Why AI fund returns for 2024–2026 could be brutal for many, great for a few.
Team Composition in the AI Era: Smaller, more technical, with an even more critical lead architect.
Oxx’s Hill-Climbing Strategy: Avoiding hype rounds, backing “dark horses” with real business value.
Here’s what’s covered:
01:21 The “circular economy” moment: Nvidia, OpenAI, Oracle and a very unhealthy funding loop
02:30 Gartner’s 4,000-company survey: only 8% with live genAI in production, 92% pilots and sandboxes
03:30 Pilots vs real revenue: how Oxx tries to separate momentum from meaningful, workflow-embedded adoption
05:21 Signals that pilots convert: usage patterns, spreading across teams, and clear ROI logic
07:04 How founders should present quality: domain expertise, business case clarity, and evolving pricing models
08:54 Metrics that matter: retention, ICP-specific cohorts and reading product–market fit at “warp speed”
10:48 Application vs LLM: building on public models, avoiding lock-in and owning the business process
13:15 Moats in AI: why speed isn’t enough, and how orchestrated workflows + unique data build defensibility
17:13 Oxx’s investment principles: first-principles productivity and when a high multiple can still make sense
18:56 Bubble talk: generational winners vs “everyone funded like they’re OpenAI” and what that means for vintages
27:06 Who should pitch Oxx: domain-native founders, “dark horses”, and PMF appearing earlier than the old labels
You can listen to the full episode on Apple Podcasts and Spotify 🎧
✍️ Show Notes
Oxx in one glance
Focus: Specialist B2B software investor, with a strong lens on productivity, workflows and enterprise processes.
Stage: Invests when there is real product–market fit, whether the round is called “Seed”, “Series A” or something else.
Style: Valuation-conscious, fundamentals-first, and happy to climb a different hill rather than chase the most hyped deal in every cycle.
The red flags: circular funding & pilot-driven demand
Mikael’s “warning bell” moment came from two things:
The circular economy article:
Nvidia investing in OpenAI → OpenAI buying Nvidia chips.
Oracle investing in OpenAI → OpenAI buying Oracle capacity.
Capital, demand and valuation all chasing each other in a closed loop.
“That whole thing is just not very healthy. It’s insanely unhealthy.”
Gartner’s data from 4,000 enterprises:
Only 8% had live genAI applications in production.
92% were pilots, proofs of concept, sandboxes.
Yet the capex and opex going into AI infra is massive.
His conclusion: much of the apparent “traction” in AI is pilot-driven. Once companies move from “playing around” to structured deployment, the pace will slow—because integration into real processes and systems is hard, slow, and political.
Pilots vs production: how Oxx separates hype from substance
Oxx’s framework for distinguishing experimentation from durable value:
Usage spread:
Does usage stay with one curious user, or spread to teams and departments?
Does it expand from one task to multiple tasks until it starts to look like a business process?
Workflow embedding:
Is the product embedded in a critical workflow, integrating with other systems?
Does it orchestrate work across multiple stakeholders, not just a single user?
Economic logic:
Can you articulate, in plain terms, how the deployment yields clear ROI?
Are customers seeing measurable productivity or outcome improvements vs “this is cool”?
If an AI tool is sold to individual users on a credit basis and used lightly, Mikael views it as consumer-like behaviour. To be a true enterprise company, the product must:
move from individual experiments to multi-stakeholder adoption, and
show up inside recurring business processes.
How founders can show “quality revenue”
Mikael’s advice for founders pitching AI traction:
Lead with domain understanding.
If you’ve lived the problem—whether in sales, marketing, finance, healthcare, or another vertical—you can better articulate why your solution is enduring.
That domain insight is crucial to explaining why today’s pilots will become tomorrow’s core systems.
Show the commercial evolution.
Many AI products start with credits or usage-based trials.
Mikael looks for signs that relationships are shifting towards:
more credits (growing volume and importance), then
subscription or outcome-linked pricing that reflects long-term value.
Back it with retention and cohorts.
The “golden metric” is still retention.
In the AI era, you have to read it at warp speed: not annual cohorts, but monthly or even weekly patterns.
Where are you seeing strong, persistent usage across a specific ICP and use case? That’s where real PMF lives.
Building on LLMs: what’s fine and what isn’t
On the “should we just build on OpenAI?” question, Mikael is pragmatic:
Yes, build on public LLMs—don’t reinvent infra.
But:
Avoid deep, brittle dependency on any single provider.
Abstract your LLM layer so you can switch for cost, performance or strategic reasons.
Don’t be “just a UI on a model”.
The real opportunity is to:
embed deeply into business processes,
orchestrate workflows across multiple roles,
and become the system of interaction or intelligence for that process.
If you can do that, Mikael believes you can build application-layer businesses as big as the last generation’s Salesforce, Adobe, etc.
Moats in AI: beyond “execution speed”
Mikael pushes back on the idea that “execution is the only moat left”:
Orchestrating a business process across stakeholders—better and more flexibly than legacy ERP—can be a powerful moat.
The difference from old SAP-style systems:
Don’t force everyone into one rigid process.
Support moldable, personalized interfaces; my UI can be different from yours, while sharing the same underlying workflow.
If you do this well, and in the process generate unique transactional and insight data that continuously improves the product, you’ve built a moat that is:
tied to real work,
difficult to rip out, and
improving with every transaction.
Valuations, bubbles & what happens to vintages 2024–2026
On valuations, Mikael is blunt:
Oxx is on the conservative side, but will pay up if commercial proof justifies it.
The real issue is systemic:
A few AI companies will be truly generational (OpenAI likely already is).
But many are being funded as if they’ll be generational winners.
For society, he’s actually positive:
The overspending will help build out a lot of infrastructure and tooling we’ll all benefit from.
For VC:
When LPs look back at fund performance from vintages 2024–2026, many AI-focused funds will “look like shit”.
There will be a big hole of overfunded companies that didn’t work—and a small set of spectacular outliers.
Oxx’s response:
“Climb another hill.”
Don’t fight to get into the most oversubscribed, hyped rounds.
Look for under-the-radar companies with critical business value that may grow slower short term, but can become very large over time.
Team composition in the AI era
The availability of AI tooling changes team design, but not everything:
You need less raw coding capacity—copilots and tools help.
The lead architect becomes more critical:
They’re the brain behind the product.
Without a strong founding architect, you risk “sprawl hell”.
On go-to-market:
Early PLG and open-source traction can be built with tiny teams.
But once you enter enterprise sales, the job is the same as ever:
multiple stakeholders,
RFPs and procurement,
complex decision-making.
AI doesn’t magically simplify enterprise politics.
Are we in an AI bubble? Yes—and that’s okay
Mikael believes we are in a bubble around expectations, not around the underlying technology:
We’re overestimating how fast AI will turn into cash flows in the next 5 years.
We’re underestimating how transformative it will be over the longer term.
He leans on the Gartner Hype Cycle:
We’re at the peak of inflated expectations.
There will be a drop into the trough of disillusionment.
Then a climb up the plateau of productivity, as with every major tech shift.
Crucially: he’s a long-term AI bull—the critique is about timing and pricing, not about whether AI matters.
Who should come to Oxx?
The founders Oxx wants to back, especially in AI:
Deep domain experts who truly understand the business process and pain they’re addressing.
Often more experienced.
With scars from prior attempts or careers in that domain.
Teams configured for the new era:
Strong product architect,
agile development talent,
and a clear vision of how AI gets embedded into workflows, not just features.
Often “dark horses”:
Not the obvious darlings everyone is chasing.
Teams that had to fight harder, stayed true to their vision, and now sit on real product–market fit.
On stage, Oxx still optimizes for true PMF, not labels:
Yes, PMF is appearing earlier now (sometimes in “Seed” rather than “Series A”).
That forces Oxx to adapt:
fewer 12–24 month histories,
more reliance on usage patterns, ROI logic, and evolving commercial models.
The signals they look for are the same; the time window just got shorter.
💡 One-Liner Takeaway
AI will transform software and productivity—just not as fast as current prices suggest.
The winners will be those who embed into real workflows, own the data and stay disciplined while everyone else chases the peak of inflated expectations.
