General-purpose AI tools keep disappointing founders who use them for marketing — because 90% of marketing is generic but 100% of the value lives in the 10% that’s specific to your business. a16z’s latest essay argues that vertical software survives AI disruption through process engineering, not switching costs. The “last mile” in marketing isn’t deploying the tool — it’s encoding which of 19 channels to prioritize, how your brand voice differs from competitors, and when to post on LinkedIn vs. Hacker News for your specific audience.

Key Takeaway: Vertical AI beats horizontal AI because the value is in domain-specific workflow encoding, not generic capability. For marketing, this means AI that understands your specific stage, channels, tone, and audience — not a general chatbot that writes blog posts for any company the same way.

George Sivulka — founder and CEO of Hebbia, an AI company building vertical AI for knowledge work — just published “In Defense of Vertical Software” on the a16z newsletter. His thesis: vertical software doesn’t survive AI disruption through switching costs or familiar UIs. It survives through process engineering — deep understanding of how specific teams in specific industries actually work.

His examples come from finance. But the argument maps perfectly to marketing — and explains why general-purpose AI tools keep disappointing founders who try to use them for growth.

The Last Mile Isn’t Deployment. It’s the Workflow.

Sivulka’s central reframe: the “last mile” problem in enterprise software isn’t about getting the product to the customer. It’s about encoding the actual process — the idiosyncratic, domain-specific way work gets done.

His examples from finance are illustrative:

Different compliance flags between private credit and private equity teams at the same firm
Individual managing directors with varying standards for CIM summaries
Divergent diligence processes — forty-page templates versus email spreadsheets

The 90% of work that looks generic across firms isn’t where the value lives. The value is in the 10% of idiosyncratic preferences where “deals get done and careers get made.”

Marketing has the exact same structure.

Marketing’s Last Mile Problem

Every marketing team thinks their workflow is unique. They’re right — but not in the way they think.

The 90% that’s generic: create content, post it somewhere, track metrics. Any tool can do this. ChatGPT can write a blog post. Buffer can schedule a tweet. Google Analytics can show you traffic numbers.

The 10% that actually matters:

Workflow	What Generic Tools Miss
Channel selection	Which of the 19 traction channels to prioritize for this specific business at this specific stage
Content calibration	The difference between enterprise SaaS tone and DTC brand tone — not just vocabulary, but argument structure
Distribution timing	When to post on LinkedIn vs. when to submit to Hacker News vs. when to send a cold email — varies by industry
Cross-channel orchestration	A blog post that feeds a Twitter thread that feeds an email sequence that feeds a webinar funnel — the connections are domain-specific
Optimization loops	Knowing that a 2% click rate on a SaaS email is good but on a consumer email is terrible — context that generic tools don’t encode

Sivulka writes: “The value of enterprise software comes from understanding the process and the organization well enough to make the software do exactly the right thing.”

Swap “enterprise software” for “marketing AI” and you have the thesis behind vertical marketing agents.

Why Foundation Model Providers Can’t Solve This

Here’s Sivulka’s sharpest argument: general-purpose AI providers — Anthropic, OpenAI, Google — cannot be opinionated about specific organizational workflows. They build for everyone simultaneously.

Claude is brilliant at writing. It can draft a blog post, compose an email, generate ad copy. But ask it to:

Decide which channel to prioritize for a pre-seed B2B SaaS company
Distribute that content across platforms with channel-specific formatting
Monitor performance and reallocate effort based on what’s working
Orchestrate a multi-step campaign across content, email, PR, and community

It can’t. Not because it lacks intelligence — because it lacks the process engineering layer that turns intelligence into execution.

This is the same argument we made about Anthropic’s own marketing team. They use Claude to save 100+ hours a month on content creation. But they still need humans for distribution, scheduling, monitoring, and cross-channel orchestration. The foundation model handles the generic 90%. The domain-specific 10% requires something built specifically for marketing.

Stronger Models Make Vertical Software More Valuable

This is Sivulka’s most counterintuitive argument — and the one most relevant to marketing AI.

You’d expect that as foundation models get smarter, horizontal tools would subsume vertical ones. The opposite happens. Sivulka points to legal AI: after OpenAI’s o-series models launched, legal AI companies thrived rather than collapsed. Why?

Because the orchestration layer matters more than the base model. Knowing when to trust the model, what data to feed it, how to route outputs, and when to escalate to a human — that’s the vertical layer. Better base models make the orchestration layer more powerful, not less relevant.

In marketing, this means:

A smarter model writes better copy. But a vertical marketing agent knows which copy to write for which channel and when to publish it.
A smarter model understands more about your market. But a vertical agent knows how to turn that understanding into a campaign that runs across 19 channels simultaneously.
A smarter model can reason about strategy. But a vertical agent executes the strategy autonomously — research, create, distribute, measure, adjust, repeat.

Every improvement in Claude or GPT makes Lane more capable, not less relevant. The foundation model is the engine. The vertical agent is the vehicle.

The “90% Right Is 100% Wrong” Problem

Sivulka’s most striking claim comes from finance: “90% right is the same as 100% wrong.” In compliance-heavy industries, almost-right output is worse than no output.

Marketing has a different version of this problem. It’s not about compliance — it’s about coherence.

A marketing campaign that’s 90% right — great copy, wrong channel. Perfect email, wrong send time. Beautiful landing page, no distribution. Each piece works in isolation. The campaign fails as a system.

This is why content generators — tools that produce 90% of the work — don’t solve the marketing problem. The 10% they miss (distribution, timing, channel selection, optimization) is the 10% that determines whether anyone sees the content.

Generic AI tools optimize for the 90% that’s easy. Vertical marketing agents optimize for the 10% that matters.

What This Means for How You Choose Marketing AI

Sivulka’s framework gives founders a clear decision tree:

Use horizontal tools (ChatGPT, Claude, etc.) when:

You have a marketing team that knows which channels to use
You need help with content creation specifically
You’re doing one-off tasks (write this email, draft this post)

Use vertical marketing agents when:

You don’t have a marketing team and need end-to-end execution
You need cross-channel orchestration, not just content
You want the system to decide what to create, where to distribute it, and how to optimize — based on what works for businesses like yours

The horizontal tool is a pen. The vertical agent is a marketing department.

This is why we built Lane on the 19 traction channels framework rather than as a general-purpose writing assistant. The framework encodes decades of process knowledge about which channels work for which types of businesses at which stages — exactly the kind of domain-specific intelligence that Sivulka argues horizontal players can never replicate.

The Vertical Bet

Sivulka closes by arguing that vertical software companies should lean into their domain expertise rather than trying to compete on model quality. The AI labs will always have better base models. The vertical players will always have better process engineering.

The same is true for marketing AI. OpenAI will always have a smarter general-purpose model. Anthropic will always push Claude further. But neither will build a system that knows the difference between a pre-seed SaaS marketing strategy and a Series B e-commerce marketing strategy — and executes both autonomously.

That’s the vertical bet. And it’s the one that a16z just told you to make.

Source: In Defense of Vertical Software — George Sivulka, a16z newsletter

References

In Defense of Vertical Software — George Sivulka, a16z newsletter, February 2026
Related: Anthropic’s Marketing Team Saves 100+ Hours With Claude. Here’s What They Still Can’t Automate. — The copilot ceiling in practice
Related: The 19 Traction Channels, Explained — Domain-specific channel framework Lane uses
Related: Dario Amodei Predicts a Country of Geniuses in a Data Center. — The model capability exponential
Related: Software Is Eating Marketing Labor — a16z’s broader take on AI replacing services