There is a recurring tension in how investors and founders talk about AI for marketing. On one side: the platform bet. Build a general-purpose AI layer that plugs into any marketing workflow, capture the whole funnel, sell up and down the org chart. On the other side: the vertical bet. Own one workflow deeply — email subject lines, display creative variation, SEO content briefs — and become so good at that specific thing that switching costs make you nearly immovable.
We have watched both plays unfold across the companies we have backed since 2021. The evidence is fairly clear, and it informs how we evaluate every new investment. This is an argument for the vertical bet — with the nuance that vertical does not mean permanently small.
What "vertical AI" actually means in this context
Vertical AI, as we use the term, does not mean industry-specific. It means workflow-specific. A tool that is purpose-built for writing paid social ad copy is vertical. A tool that writes all marketing copy for all channels is not vertical — it is a platform. The distinction matters because the quality bar for each workflow is set by the professionals who do that work daily, not by a general language model benchmark.
A paid social manager evaluating ad copy has precise opinions. The hook needs to load in under six words. The claim needs to match the creative format. The tone needs to match the platform — Facebook copy does not read like LinkedIn copy does not read like TikTok copy. A general-purpose language model that has seen the full distribution of internet text produces mediocre output across all of these. A model fine-tuned on high-performing paid social creative, trained with feedback from actual campaign performance data, produces output that a practitioner would not be embarrassed to run.
This is the core compounding advantage of vertical AI: the feedback loop is tighter. A platform that generates all content types cannot easily attribute output quality to specific use cases. A vertical tool that owns one workflow gets performance data that is directly relevant to refining the model for that workflow.
Why the data moat is real in narrow verticals
We are skeptical of generic data-moat arguments — every AI company claims one. But in specific marketing workflow verticals, the moat argument holds up in a way it does not for general content generation.
Consider ad creative specifically. A platform that generates ad creatives and measures downstream performance — click-through rate, conversion rate, cost-per-acquisition across channels and industries — accumulates a training signal that is genuinely hard for a general-purpose model to replicate. The reason is not volume. It is specificity. The platform knows not just what copy was generated, but what copy was actually deployed, what variants were tested against each other, and which variant won. That performance-linked training data is structural. It cannot be scraped. It can only be accumulated by being in the workflow.
We saw this with the companies we backed in the generative ad creative category. The ones that survived the 2023 consolidation wave were the ones who had accumulated the most performance-linked feedback data, not the ones with the broadest feature sets. Breadth was not the moat. Specificity was.
The failure mode of premature horizontal expansion
The platform trap is real, and it happens at a predictable moment. A vertical AI company achieves strong retention in its core workflow. Users start asking for adjacent capabilities. The sales team argues that a broader platform would close larger deals. The founder, understandably, starts roadmapping expansion.
We are not saying horizontal expansion is wrong. We are saying it is wrong at the seed stage, and it is wrong before the core workflow is defensible. The companies that expanded too early ended up competing with general-purpose AI platforms on those platforms' terms — breadth of capability, rate of feature release, integration surface area. Those are losing battles for a small team.
The companies that stayed vertical long enough to build a genuine performance feedback loop — typically 18 to 24 months of being deeply embedded in one workflow with one buyer persona — came out of that period with something no general-purpose platform had: conviction from their users that the tool actually worked better than anything else for that specific thing. That conviction translates into renewal, into referral, into pricing power.
How we evaluate vertical AI claims during diligence
When a founder pitches us on vertical AI for a specific marketing workflow, we have a short checklist of questions that gets to whether the vertical claim is substantive or cosmetic.
First: does the model actually know something about this workflow that a general-purpose model does not? We ask the founder to show us examples where their tool produces materially better output than GPT-4 on the same task. If the examples are not compelling, the vertical positioning is mostly marketing.
Second: what is the performance feedback loop? How does the model improve from deployment? If the answer is "users can rate outputs" we are less interested than if the answer is "we track downstream campaign performance and use that signal to fine-tune." The difference between those two is the difference between a UI preference and an actual compounding data advantage.
Third: who is the primary buyer and what is the switching cost at 18 months? If the tool is embedded in a workflow that the buyer owns — not just uses — and if the institutional memory of the tool (brand voice fine-tuning, product knowledge, historical performance data) would be painful to recreate elsewhere, that is a genuine switching cost. If the switching cost is "we would need to re-set up the integration," that is not a moat.
Vertical does not mean capped upside
The common objection to vertical AI is that the addressable market is small. A tool for writing email subject lines only — how big can that be? The answer, depending on how you count it, is quite large. Every marketing team that sends email has someone whose job involves writing or approving subject lines. The question is whether the tool is wedging its way into a workflow that expands over time, or whether it is genuinely limited to a narrow task forever.
In practice, the best vertical AI companies do expand — but they expand from a position of strength in the core workflow rather than from insecurity about market size. A tool that has become the standard for subject line optimization in a marketing team is very well positioned to own the full email content layer, then the full campaign messaging layer. The expansion is earned. It is not assumed in the original pitch.
What we look for at seed is a founder who is comfortable staying narrow long enough to build something genuinely defensible. That discipline is rarer than it sounds. The pressure to expand — from investors, from customers, from the competitive landscape — is relentless. Founders who can hold the line on their core workflow until they have built a real compounding advantage are the ones who end up building platforms that matter. The vertical discipline is what creates the foundation the platform stands on.