I spent most of 2018 and 2019 selling a programmatic advertising platform to mid-sized e-commerce brands in Iberia. The pitch was always some version of: better targeting, faster creative iteration, lower CPAs. The buyers were smart people with limited budgets who had already tried four other tools that promised the same thing. What they actually needed wasn't a better ad-serving layer. They needed a different way to think about how marketing content got made and measured. We didn't have that answer. We had a dashboard.
That experience is part of why I founded Telhaverde with a specific thesis rather than a broad mandate. The restructuring happening now in marketing technology is not incremental. It's the kind of change that makes some categories of software irrelevant while creating entirely new categories from scratch. And most of the founders we meet who are building in this space are underestimating how deep the restructuring goes.
What "restructuring" actually means
The marketing stack as most enterprises know it is a pile of point solutions assembled over fifteen years of tactical procurement. CMS, DAM, email platform, social scheduler, ad network integrations, analytics layer, attribution tool — each purchased to solve a specific bottleneck, each requiring manual workflows to connect to the others. The underlying assumption holding this stack together is that content is expensive and scarce, so you spend time optimizing its placement and distribution. You A/B test a subject line because writing a hundred variants isn't worth the cost.
Generative AI breaks that assumption at the root. When content generation becomes orders of magnitude cheaper and faster, the constraint shifts from creation to curation, quality control, and personalization at real scale. The tools built to optimize scarce content don't help you manage abundant content. The workflows built around expensive creative production don't make sense when production time collapses. You don't need a better version of your 2019 marketing stack — you need a different stack built on different assumptions.
This is not a claim that every existing MarTech vendor is going away. Some categories — CRM, customer data platforms, event infrastructure — are fairly durable because their core value isn't tied to the creation-scarcity assumption. But anything that primarily existed to help teams do more with less content, or to optimize scarce creative, is being pressured from below by foundation models and above by new specialized tools built to work with abundance.
Why the infrastructure layer matters more than the application layer right now
When we were raising Fund I in 2021, we had a thesis about generative AI in marketing that was, in retrospect, more right about the direction than the timing. The large language models of 2021 were good enough to be interesting, not good enough to be trustworthy in production marketing workflows. What happened between 2021 and now is that the models got genuinely good — good enough that the conversation shifted from "can AI generate usable marketing copy" to "how do we route, evaluate, and scale AI-generated content in our production systems."
That shift changes what the interesting infrastructure problems are. When content generation was unreliable, the hard problem was making it reliable. Now that generation is reasonably reliable, the hard problems are: How do you maintain brand voice consistency across thousands of generated variations? How do you build evaluation pipelines that catch quality issues before they reach customers? How do you integrate model-generated content into existing approval and compliance workflows without slowing them down? How do you attribute performance back to content variables when you're running ten thousand variants instead of ten?
These are infrastructure questions. They're not glamorous, they're not the kinds of problems that get covered in consumer media, and they require founders who understand both the NLP side and the enterprise systems side. We look for both — which is rare, and which is why we believe seed-stage infrastructure in this layer is undersupplied relative to the eventual demand.
The European dimension isn't a footnote
One thing I've observed in conversations with US-based investors looking at AI MarTech: Europe is often framed as "harder," meaning the buyer cycles are longer, the GDPR compliance requirements add cost, and the market is more fragmented by language and local context. All of that is true. But it misses what makes European enterprise MarTech genuinely interesting for infrastructure builders.
European enterprises — particularly in financial services, healthcare, and regulated retail — have compliance requirements that create natural forcing functions for better AI governance. A European bank buying an AI content generation tool for its customer communications isn't just doing brand safety review. It's managing regulatory exposure under multiple jurisdictions. That means the market is willing to pay for infrastructure that builds explainability, auditability, and version control into the content pipeline — features that US-market buyers often deprioritize until something goes wrong.
Founders who build for European compliance requirements first tend to build more defensible infrastructure. The compliance-first architecture isn't the overhead — it's the moat. We've seen this play out in portfolio companies where the compliance documentation and audit trail capabilities turned out to be the differentiator in enterprise procurement, not the generation quality per se.
Where we're not: what this thesis doesn't cover
It's worth being explicit about what we're not investing in, because the category language around "AI for marketing" has become imprecise enough that it covers very different things.
We're not primarily investing in horizontal AI writing assistants that compete on feature breadth. That category is important and has real companies in it, but the competitive dynamics favor platforms with massive distribution and the ability to bundle AI writing into adjacent products. We're not competing with that.
We're also not investing in companies whose differentiation is primarily prompt engineering on top of a foundation model. The barrier there is too low. What we look for is companies where the model — whether it's a fine-tuned language model, a domain-specific embedding layer, or a custom evaluation framework — creates a compounding data advantage over time. The model has to get meaningfully better as it processes more of the customer's content, audience data, and performance signals. Without that compounding, you're renting someone else's moat.
The companies we've backed in Fund I and are actively deploying into with Fund II all share one characteristic: they're building something that couldn't exist without deep technical work at the model layer, and that technical work creates a feedback loop the customer participates in. That's the durable version of this category.
What I'd tell a founder building in this space right now
The question we ask founders most often isn't about their technology — it's about their theory of distribution. Who is the buyer? What's the buying signal — when does a marketing team or a CMO decide they need this now rather than in six months? What's the switching cost once they're in? How does the product get stickier as usage grows?
The marketing technology buyers we know from our operating days are not early adopters. They're practitioners under budget pressure who will deploy new tools when the ROI is legible and the risk of failure is low. That means the best AI MarTech companies we've seen are ones that insert themselves into an existing workflow before replacing it — they generate one component of the content production pipeline, prove the quality, earn trust, and expand from there. The companies that try to replace the workflow on day one have a much harder time.
The stack is being rewritten. The founders who understand it from the inside — who have run content operations, managed creative production, dealt with attribution hell — are building the most interesting versions of what comes next. That's who we're here for.