ChatGPT Is the New Browser, but Browsers Don't Think for You

a16z recently invited HubSpot co-founder and CTO Dharmesh Shah to talk about software opportunities in the AI era. Dharmesh has a judgment: ChatGPT and Claude are becoming the new browser, or the new operating system.

His analogy comes from history. When desktop software moved to the browser, Internet Explorer and later Mozilla looked like they would eat everything, but in the end Google did not own the entire application layer. The browser was only the entry point; the real products, companies, and ecosystems grew on top of it.

The metaphor spread quickly, but someone in the comments gave a sharper take: browsers don't read your data and then decide what to do—these things do.

That comment points to the key difference. A browser is a passive window; ChatGPT and Claude are doing cognitive work for you. They are not just entry points; they are also moving upstream.

Dharmesh himself admits this: AI is doing real thinking, and that is different from the browser era. But he thinks the conclusion is unchanged: ChatGPT and Claude will not own the entire software stack, and a large number of companies will plug in like HubSpot.

The question is where to plug in.

The Yellow Brick Road: The One the Labs Are Building

a16z partner Joe Schmidt wrote an article called "Avoiding Death on the Yellow Brick Road," and later recorded a podcast episode with Dharmesh specifically about this.

The full interview is on YouTube and runs nearly an hour.

http://youtu.be/s4lXP9IuHjk

The Yellow Brick Road is the most conspicuous road in The Wizard of Oz. Joe uses it to refer to the direction AI labs are fully committed to: code generation, writing, image generation. What these problems have in common is that the stronger the model itself, the better the product. Every dollar put into pretraining and post-training directly improves quality.

If startups compete with the labs on this road, they basically have no chance of winning. The labs own the models, the distribution channels, and the brand halo. Putting together an orchestration layer with the same connectors and building a generic agent is, with high probability, helping someone else validate the market.

Joe's own words are more direct: if you are building an AI application, using the same connectors as the labs, with no sub-agents underneath, no configuration, and no distribution, then you are on a road to nowhere.

Elsewhere in Oz

But the Yellow Brick Road is only one road in Oz. Joe calls the rest "The Rest of Oz."

What does business look like there? Models are woven into complex tools, automation, and integration networks. The work is multi-step and multi-player, with sub-agents handling tasks for specific roles and industries. It requires gathering context across systems and then routing it to different people for approval at different stages. It often involves legacy systems, requires deterministic outcomes, and leaves little room for error.

Translation: dirty work, hard work, work that requires industry knowledge.

Joe gave a detail from sales. Prabhav Jain, CEO of 11x, described a seemingly simple problem: if a company is already your customer, you should not contact their employees again.

But how does that work in practice? The CRM may only have the parent company's domain, but the company has dozens of subsidiaries. The CRM record may be the parent company's domain. An outdated matching field in Salesforce may send a cold email to the CRO of an existing customer. Real-world data is messy, and the model will not automatically cross that gap.

Aman Gour, CEO of FurtherAI, talked about insurance. Two insurance companies run the same process: submission, review, quote, underwriting. The path looks the same, but what really distinguishes the two companies is what is inside the path: which risks get escalated, which loss signals matter, which of two conflicting underwriting rules wins, when a human signature is required, how external data is pulled in, and how the final decision is recorded.

This logic is not in a clean rules engine. It is scattered across SOPs, manager reviews, underwriting philosophies, company-specific preferences, and years of operating experience. Much of it is not written in a form a model can directly read.

Four Kinds of Defense

Joe summed up several reasons why "the rest of Oz" can hold its ground.

Data and learning flywheels. A lot of knowledge is not in any training set: unwritten industry conventions, undocumented standards, tribal knowledge in practitioners' heads. A company that has run through hundreds of legal red-line reviews, thousands of insurance underwriting cycles, and tens of thousands of sales outreach attempts has already internalized the shape of the problem. A newcomer throwing together an agent on the fly cannot replicate that.

Managing model change and complexity. The labs also route internally, with different models handling different requests. But they will not route across vendors, will not evaluate whether a competitor's model is better at a certain subtask, and will not use an open-source fine-tuned model for the narrowest slice. Application companies have to rerun evals, recalibrate prompts, and keep production from breaking every time a new model is released. Nobody wants to do this work, but customers need it.

Cost optimization. Sending every query to the most expensive model is a shortcut to negative gross margin. The best application companies route in tiers: frontier models for the hardest tasks, mid-tier models for high volume, and their own fine-tuned small models for sufficiently narrow scenarios. The labs set the floor price; application companies sell intelligence that is just good enough.

Governance. Become the control plane for how customers run AI in a given vertical: permissions, audits, what agents can do, and what they actually did all converge here. Someone also has to do the dirty compliance work: FRCP and lawyer rules in legal, HIPAA in healthcare, SEC and FINRA in finance, and state insurance regulations. A generic player cannot become one hundred vertical industries at the same time.

Three Tests

Joe gave three criteria for telling whether you are on the Yellow Brick Road or elsewhere in Oz.

Tool and step test. How many steps does this work have, and how complex are the tools needed to support it? Doing a search in Google Drive is one step with one tool, with loose results: the user reads a summary and asks again if it is wrong. Compare that with a multi-step legal red-line review using three years of law firm precedents: dozens of steps across multiple tools, output that must pass partner review, and possibly arguments in court. Both look like "an agent doing work," but only one requires a dedicated team spending years building deep software.

System test. Are you building a system that customers use to run their work, or a tool that sits on top of systems they already have? A system owns the workflow end to end: data capture, governance, and records of completion. When customers describe how the work actually happens, they are referring to the system. A tool only adds intelligence to a workflow the customer is already running. Tools can generate revenue, but the labs can take them away, because customers do not depend on you as the orchestration layer. Ask yourself: if the lab released a direct competitor tomorrow, would customers still need you? If yes, it is a system.

Hedge fund and P&L test. The labs' performance is measured by benchmark scores; the rest of Oz's performance is measured by customers' P&L. Customers do not care what score your model got on SWE-Bench or MMLU; they care whether the agent closed the deal, whether the red-line review was correct, and whether the policy was underwritten correctly. If they are looking at workflow-specific outcomes rather than generic capability scores, you are in the right place. The best agent businesses should run like hedge funds, winning on alpha in the customer's P&L, not on benchmark scores.

CRM for a Company of One

Dharmesh is also building something new himself. uSpot, a CRM for companies of one. It charges from day one, and it was his most successful launch at HubSpot.

That choice is interesting. A company of one has no sales team, and no one has time to type into a CRM. Someone in the comments put it precisely: the reason companies of one skip CRM is simple—no one enters data into it. A product that wins has to fill in data itself from the inbox and calendar, and only ask "is this okay?" before sending something.

Dharmesh worked at HubSpot for more than twenty years and never had direct reports. In the interview he said that now is the most fun he has had in more than thirty years of building software. Joe asked him how he views market sentiment. He said the worst period is also the best period; the fog of war is always there, but just focus on the customer.

As for whether ChatGPT is the new browser, Dharmesh's full answer is: it is the new OS, it is the new browser, but it will not own the entire stack. Companies will plug in like HubSpot.

Where to plug in matters more than whether to plug in.

There is no place for startups on the Yellow Brick Road. The rest of Oz is vast.