AI Labs Laid a Yellow Brick Road; a16z Urges Startups to Take a Detour

Frontier AI labs are reaching into the application layer; there is little debate about that. What's debated is how much room startups have left. a16z partner Joe Schmidt wrote an article using a Wizard of Oz metaphor to break this down. HubSpot co-founder and CTO Dharmesh Shah discussed his own version on the podcast.

Who Is the Yellow Brick Road For

In The Wizard of Oz, Dorothy follows the Yellow Brick Road to the Emerald City to find the wizard. Joe Schmidt borrows this image to refer to the path labs are rushing down with concentrated resources: horizontal tasks such as code generation, writing, and image generation. Their common trait is that every time model capability improves a notch, the product gets better too. Spending money on pretraining and post-training yields linear returns.

Walking this road is the most tempting and most dangerous thing for startups. Take a high-performance model, connect a few off-the-shelf connectors (G Drive, Slack, Salesforce, Notion, GitHub), build an agent orchestration layer on top, and it looks like a business. The problem is that the labs are doing the same thing themselves—Cowork and Codex are examples. They have the models, better gross margins, pricing power, and control over architecture choices. Even if a startup can briefly outperform Codex on performance, it can't beat the other side's distribution channels and brand halo.

Joe Schmidt's conclusion is blunt: using the same set of connectors, with no sub-agents, no configuration layer, and no distribution, is a road to a dead end.

The Rest of the Land of Oz

Interestingly, the labs themselves admit they can't do everything. OpenAI and Anthropic are both aggressively promoting forward-deployed joint ventures to help enterprises configure and customize models. If the next model release could solve everything, spending billions of dollars would make no sense.

The remaining territory looks like this: multi-step, multi-participant work that requires pulling context across multiple systems, routing to different roles of human approval, often involving legacy systems, where one wrong input causes a disaster. Joe Schmidt offers three tests to judge whether you're on the Yellow Brick Road.

First, the tools-and-steps test. Take horizontal AI search over Google Drive as an example: one step, one tool; if the result is wrong, the user just asks again. Compare that with making redline changes against three years of case law: dozens of steps, across multiple tools, with output reviewed by a partner and possibly contested in court. Both are called agents doing work, but only one requires a dedicated team building software for years.

Second, the system test. Are you building the system customers use to run their business, or a tool layered on top of their existing systems? Systems usually have higher ACV, but high ACV doesn't guarantee a system. Ask: if the lab launches something that looks like direct competition, will customers still need you? If yes, you're a system. If no, you're a tool.

Third, the P&L test. Labs score by benchmark; companies in Oz score by the customer's income statement. Customers don't care what score you got on SWE-Bench; they care whether the contract got signed, whether the clause was changed correctly, and whether the policy was bound to the right one.

What a CTO Who Has Never Had a Direct Report in 30 Years Did

In that podcast episode, Dharmesh Shah talked about his own experience. He has worked in software for more than 30 years and has never had a direct report. His latest project is uSpot, a CRM for one-person companies. It started charging on day one and became his most successful launch at HubSpot.

There's a point on the timeline worth noting: the segment at 14:40 talks about the 60% of people who aren't using a CRM at all. That explains why it's SMB and why it's one-person companies. This market wasn't worth serving before because software was too expensive and customers were too fragmented. Now software is cheaper to build, and a market that previously couldn't support a software company suddenly can.

There's another judgment in the podcast: models are becoming infrastructure, and ChatGPT is the new operating system. That sounds like a slogan, but in the context of uSpot it becomes concrete. A one-person company doesn't need a full CRM installed; it needs something that fills in data in the background and, when needed, asks only one question: "Send it or not?" A commenter added under the tweet: there's only one reason a one-person company doesn't use a CRM—no one has time to type into it. The one that can read data from the inbox and calendar by itself wins.

Half the Work Has Nothing to Do with the Labs

The article cites the hands-on experience of 11x CEO Prabhav Jain and FurtherAI CEO Aman Gour, and this part is the most worth digging into.

Prabhav Jain says that about half of any real workflow is non-agent deterministic software, and labs are no better than startups on this half. The other half is agentic, but you need to tune, train, and constrain it for specific outcomes. Writing a cold email for an insurance agent and opening a call with an SMB owner are not helped by general model capability; they have to be fed in bit by bit through vertical-domain knowledge.

He gives an example that sounds silly: if a company is already a customer, you shouldn't send a cold outreach to its contacts again. That's not simple at all. Maybe the CRM stores a company domain, but the company has dozens of subsidiaries? Maybe the master record stores the parent company's domain? Could an outdated field in Salesforce cause a cold pitch to go to an existing customer's CRO? Real data is dirty. Models won't automatically cross that line.

Aman Gour talks about the same thing from the insurance angle. Two insurance companies go through the same process: submission, review, quote, underwriting. The process itself is easy. What's hard is what's inside: which risks to escalate, which loss signals matter, which underwriting rule wins when two conflict, and when a human signature is mandatory. This logic doesn't live in some clean rules engine; it's scattered across SOPs, manager reviews, underwriting philosophy, and years of experience, much of it never written in a form a model can directly read.

So FurtherAI builds agentic workflows, not pure agents. Workflows provide repeatability, auditability, and controllable costs; agents handle variation and how to recover when the happy path breaks; humans stay in the loop to sign off on judgments that require accountability. On day one, this thing automates only the manual labor. After thousands of runs, every escalation, every exception, every human correction becomes process memory. Aman Gour's formulation: the workflow delivered on day one is not the moat; the loop that grows out of production use is.

Models Below, Systems Above

The article's concluding judgment is that there will be big winners both on and off the Yellow Brick Road. Labs keep winning because they own both the models and the distribution channels for horizontal tools. Whether companies off the road can win depends on whether they own the work system: the interface where the work is actually executed, and the data that flows out of it.

Models are replaceable. Work systems are not.

Dharmesh Shah and Joe Schmidt's full conversation is here, nearly an hour long, ranging from how to pick startup ideas to how HubSpot internally used AI to replay 20 years of code.

The YouTube version is here, with more granular chapters. The Builder section and the one-person company section are worth watching first.

Joe Schmidt's original article was published on X under the title "Avoiding Death on the Yellow Brick Road"; the link is here.