Stop Building ChatGPT Wrappers. Build Vertical AI.

A founder pitched me in Bangalore three months ago. SaaS product, AI-powered, serving small businesses. Clean demo. Nice UI. I asked him: “If OpenAI ships a slightly better model next month, does your product get better or does it get replaced?”

He didn’t have a good answer. That’s the wrong answer.

That question — almost word for word — is the filter that Elad Gil, one of the most successful early AI investors in Silicon Valley (early backer of Airbnb, Stripe, Square, and more recently Anyscale, Groq, and Harvey), uses to separate real AI companies from what he calls “thin wrappers.”

In a McKinsey interview, Gil put it this way: the key question is whether an increase in the foundation model makes your product better or worse. If it’s just taking and reformatting data from an LLM, it won’t be useful because the model will eventually get there. But if it provides deeper functionality, workflows, and integration with existing tools — that’s when it’s unique.

That test is deceptively simple and ruthlessly clarifying.

What a Wrapper Actually Is

Let me be precise, because “ChatGPT wrapper” has become a vague insult.

A wrapper, in Gil’s framework, is a product whose entire value proposition is a marginal UX improvement over a raw API call to a frontier model. Better prompting, cleaner interface, slightly more constrained output. The second OpenAI or Anthropic ships a feature that achieves the same thing — or the moment a user figures out they can do it directly — the wrapper has no value.

Wrappers exist because it’s fast to build them. I’ve seen teams go from idea to MVP in 72 hours by gluing a UI on top of an API. The speed is real. But the durability isn’t.

a16z data from 2025 makes this concrete: among the top 50 consumer AI products, companies with proprietary models raised an average of $98 million. Companies that fine-tuned existing models raised $20 million. “Wrapper” companies raised $9 million. The market is pricing defensibility already. Investors aren’t fooled by demos anymore.

The question isn’t whether your AI demo is impressive. Every AI demo is impressive in 2026. The question is what happens the day after the demo.

What Vertical AI Actually Means

Vertical AI is not just “AI applied to a specific industry.” That definition is too weak.

Real vertical AI means building around the workflow of a specific domain — understanding the data that flows through it, the decisions that get made, the integrations that matter, and the compliance context that surrounds it. The AI component is almost secondary. The primary asset is your depth of understanding of how a specific type of work actually gets done.

Gil’s example of compliance workflows is instructive: “Anytime you’re doing a lot of repetitive tasks with documents, you’re doing something that a language model can do very well.” But the opportunity isn’t just “use an LLM to read documents.” The opportunity is to map the specific document types, the specific regulatory framework, the specific outputs required, the specific downstream systems those outputs need to integrate with — and build something that works inside the actual operational reality of people who do that job.

Harvey — the legal AI startup that Gil called out as “just working” — is a clean example. It raised three massive rounds in 2025, going from a $3B to $5B to $8B valuation. Harvey isn’t a legal chatbot. It’s built around the actual workflow of law firm lawyers: the way cases are researched, memos are written, documents are reviewed, precedents are found. The AI is embedded in process, not floating on top of it.

The Boring Industry Playbook

This is where I spend a lot of my thinking time, and I want to be honest: the best vertical AI opportunities are in the industries that nobody at a Silicon Valley conference is excited about.

I’m talking about:

• Freight brokerage and logistics — relationships-driven, document-heavy, manual processes running on spreadsheets and phone calls in many markets. The TAM is enormous and the incumbents are slow.

• Trade finance — letters of credit, bills of lading, compliance documentation. Massive transaction volumes, extraordinarily paper-intensive, with enormous variation in formats and regulatory requirements across jurisdictions.

• Facilities management — maintenance scheduling, vendor coordination, compliance reporting, asset tracking. Running at scale in property companies, hospitals, and campuses everywhere. Almost entirely manual at the coordination layer.

• Professional services back-office — accounting firms, law firms, architecture practices. The tools exist but they’re not connected and the workflow is held together by institutional knowledge.

• Healthcare revenue cycle management — insurance claims, coding, denials management. An industry generating billions in administrative waste because the process is incomprehensible.

Every one of these is ugly. Every one has a sales cycle that involves procurement committees and security reviews and change management. Every one requires you to understand domain-specific terminology, regulations, and operational culture before you can build anything useful.

And that’s exactly why they’re good opportunities.

Paul Graham’s “Schlep Blindness” essay makes this point from a slightly different angle: there are great startup ideas lying around unexploited because people avoid them for being unpleasant or tedious. Stripe became one of the most valuable private companies in the world precisely because they were willing to deal with banks and fraud and payment regulations that everyone else avoided. Stripe didn’t just build a payments API — they built deep into the regulatory and operational substrate of money movement.

Vertical AI in boring industries is today’s version of that.

What Defensibility Actually Looks Like

I want to be specific about what makes vertical AI defensible, because “workflow integration” is easy to say and hard to achieve.

Proprietary data compounds. When your product is used in a real operational workflow, it generates data that a raw API call never sees. Edge cases, corrections, domain-specific outputs, feedback loops. That data, accumulated over months and years of real usage, trains better models and builds context that a competitor starting fresh can’t replicate.

Switching costs are real. When your AI is embedded in someone’s ERP, their compliance system, their client data, their internal workflow — switching isn’t a product decision, it’s a migration project. The bigger the workflow surface area, the higher the switching cost.

Trust takes time to earn in regulated industries. Healthcare AI and legal AI and finance AI don’t get adopted on the strength of a demo. They get adopted when someone internal champions the product, when it survives a security review, when the compliance team signs off. That process is slow and painful for early-stage companies — and it’s also exactly what makes the position hard to attack once established.

Domain expertise in the team creates compounding advantage. The companies building vertical AI that actually work have founders or early team members who come from the industry. They know what the documents look like, who the decision-makers are, what the failure modes are, what the compliance landscape demands. You cannot hire your way to that knowledge from the outside quickly.

How to Evaluate Whether You’re Building a Wrapper

I run this filter on every AI startup I talk to now, and I’d run it on your own company if you’re building in this space:

1. What specific operational data does your product generate or ingest that isn’t available from a model API? If the answer is “none,” you’re a wrapper.

2. If the underlying model improves by 30% in capability next month, do you win or does your moat narrow? If the model improvement makes a raw API call substitute for your product, you’re a wrapper.

3. What does your product know about a specific customer’s context, history, and workflow that a competitor starting fresh six months from now wouldn’t know? If the answer is “nothing,” you’re a wrapper.

4. What integration does your product have with the real systems — ERPs, CRMs, compliance databases, workflow tools — that your customers already use? If the answer is “none,” you’re a wrapper.

You can be a wrapper and make money in the short term. I’m not saying don’t ship. I’m saying know what you are and what you need to build to not be that anymore.

The founders I’ve seen win in AI over the past two years are not the ones with the best technology instincts. They’re the ones with the deepest operational instincts about a specific domain. They know their industry the way a veteran knows a battlefield — the terrain, the weather patterns, the mistakes that look obvious in hindsight.

Build that. The AI is almost the easy part.

---