Dharmesh Shah: Find startup opportunities beyond the AI labs’ path

ChatGPT and Claude Cowork are becoming a new interface for software, but HubSpot co-founder Dharmesh Shah argues that useful work and customer relationships can retain value behind that interface. He joins a16z partner Joe Schmidt to discuss where software companies can build as AI labs expand.

Shah says HubSpot began by helping small businesses use the internet without being overwhelmed by fragmented tools. Two decades later, he sees a similar mission in helping the same customers use AI. Market uncertainty is real, but he is concentrating on customers and the opportunities to build.

He sees AI as another step in software’s long move toward higher levels of abstraction: builders can express an intended solution more directly. The problems businesses face have not disappeared, but the tools available to address them and the number of people who can build have expanded.

Faster iteration matters because a founder can put an idea in front of real customers and learn sooner. Shah does not think careful analysis alone can reliably identify a winning idea; even a successful company often ends up looking different from its initial concept.

Distribution may become harder as building software becomes easier, Shah says, because more products compete for attention. He advises starting with a narrow customer and problem, getting the first dollar of revenue and shortening the time and cost of each learning cycle.

The early HubSpot founders saw small and medium businesses as a market between long enterprise sales cycles and consumer businesses that could be difficult to charge. They explored a law-firm management product, but conversations showed that finding clients interested firms more than some of the administrative features.

The team chose marketing software for small businesses. Shah says HubSpot did not begin with a detailed plan to become a CRM platform; the later expansion emerged as the company learned more about customer needs and adjacent products.

A customer survey found roughly 20% used Salesforce, another 20% used a different CRM, and 60% used none. These were already paying HubSpot customers generating leads, so the founders asked what happened to those leads and why those businesses had not adopted a CRM.

Shah says complexity and cost were the barriers HubSpot focused on. It built an easy-to-use CRM and made it free, using the existing marketing business to support the investment. The goal was to create customer value before asking to capture more of it.

Although founders are often told to do one thing well, HubSpot initially offered a broad set of marketing tools: content management, search optimization, landing pages and forms. Shah says that breadth was necessary to address the customer’s actual problem, which was managing a fragmented collection of tools.

A platform ecosystem has to be treated like a product, Shah says. Developers need to know what the platform is likely to build itself and how quickly they can get an integration working and earn their first revenue. He points to the damage done when platforms first attract developers and then restrict them.

Schmidt’s “Yellow Brick Road” describes a tempting path on which frontier labs are advancing quickly. Shah agrees there is opportunity outside it, but warns founders against building only on a capability a leading model lacks today: adding the word “yet” may expose how temporary that gap is.

Founders should consider the company behind a product as well as the product’s present features. Shah looks for areas where a larger company faces a structural constraint or has little incentive to pursue a specific customer problem, while recognizing that the frontier labs have the ability to expand.

He uses CRM as an example of an area a frontier lab could enter, while arguing that its priorities and incentives make a focused effort less likely. This is his assessment of the competitive landscape, not a promise that a major AI company will stay away.

The labs have invested heavily in models and compute and need ways to sell their capabilities. They may serve consumers and large enterprises directly, while application companies can package that infrastructure for specific workflows and customer needs.

Shah compares AI interfaces to the browser’s role in an earlier software transition. Applications ran through browsers without the browser owner owning every application. In his view, ChatGPT and Claude Cowork may similarly become places where people access value created by other companies.

The interface matters, but the underlying work can matter more. Shah compares it to food delivery: an app is useful, yet the relationships, restaurant network and delivery logistics also create value. He argues that business software should be judged by the outcomes customers achieve.

He recommends using a deterministic database query for a question that structured data can answer, rather than asking a language model to infer the answer from a mass of messages. AI can be valuable without being the right mechanism for every operation.

HubSpot has long built reusable primitives: its marketing automation workflow could be reused when the company expanded into sales automation. Shah says the current question is which underlying components the company needs for the next two decades in an AI-enabled world.

He expects structured records to remain useful alongside a “system of context” containing unstructured information. A context graph could connect that information so agents can draw on both known facts and surrounding business context to do work.

Agentic coding could let an established company explore larger architectural alternatives, Shah suggests. He imagines spending money on model usage to replay an old implementation with different choices and then using the existing tests to assess the result; he presents this as an experiment, not a completed rewrite.

HubSpot Next operates as an innovation lab, with teams given freedom to test new products, markets or technologies. Shah recalls applying a similar approach to the original CRM: the team did not have to use the existing stack or brand if another route better solved the problem.

His current project, uSpot, is a CRM for one-person businesses. At the time of recording, he says the product was 30 days old. It is a small setting in which to test ideas such as a “second brain” that connects information and helps an individual get work done.

Rather than building a generic fix for AI memory that the model companies might soon supply themselves, Shah wants to learn what a particular customer needs and where context can be inserted into that customer’s work. A narrow setting can make the experiment cheaper and faster.

Successful experiments can be absorbed into the main company, remain separate or contribute selected ideas. Shah cites an earlier agent builder that was brought into HubSpot. He says the benefit is combining a startup’s freedom from organizational baggage with access to an incumbent’s customers, brand and experience when useful.

With uSpot, Shah began charging early to learn about pricing and perceived value. He says he was surprised by customers’ willingness to put down a credit card for usage-based pricing, compared with other projects he had launched.

He points to US business-formation data and argues that one-person companies are a growing market. He says more than 70% of new formations fall in a category not expected to hire an employee within two years. He sees this as an opportunity for software serving very small businesses.

Shah’s response to predictions of a software collapse is that AI itself is software and can address problems older tools could not. Some established companies may fail through the transition, he says, but he expects the overall set of opportunities to grow.

Asked about the best day he has had at HubSpot, Shah says it was about a week before the recording: ideas were starting to fall into place. He says his role lets him draw on the benefits of a large company while continuing to work hands-on on new projects.