Context
Dharmesh Shah is the co-founder and CTO of HubSpot, the CRM and marketing software company for small and mid-sized businesses, which he says is now 20 years old and has over 3 billion dollars in revenue. He talks with a16z partner Joe Schmidt, whose essay "Avoiding Death on the Yellow Brick Road" argues that there is plenty of room to build software outside the path the frontier AI labs are racing down. The conversation covers how to pick and test ideas, how HubSpot chose its market and grew from one product into a platform, and where application companies can still build lasting value as AI models become infrastructure. For PMs, it is a practical guide to finding problems worth solving, validating them cheaply, and deciding where a product can win.
The Big Idea
You cannot know in advance whether an idea is good, so the winners are the teams with the shortest and cheapest learning loops, aimed at a narrow market and at the whole customer problem.
Dharmesh backs this with HubSpot's own history: the original plan was ERP software for law firms, the CRM was never in the early pitch decks, and both pivots came from talking to customers rather than from analysis. AI makes those learning loops faster than ever, while distribution gets harder because everyone can now build.
Key Insights
1. Ideas are tested, not analyzed
- What: Dharmesh says a common founder mistake is treating idea selection as rigorous analysis, such as an 18-column spreadsheet or US Census counts of how many law firms exist.
- Why it matters: in his experience, and in talks with other founders, it is almost impossible to know an idea's quality or odds of success ahead of time.
- Evidence: even successful companies often end up with a product that looks different from the original idea, keeping only some of its "gene pool."
- So what: since you cannot predict, the smart move is to get the idea in front of real users as fast as possible and learn whether to keep going or try something else.
2. Start narrow; distribution got harder
Testing ideas fast only pays off if the test reaches real customers, and Dharmesh says that is where the challenge has moved:
- AI lets teams work at a higher level of abstraction, so lessons that once took six months to two years can now be learned by shipping something and testing it with real customers.
- Distribution is the one thing that is still hard, and he argues it may be harder than before, because many more people can now build software and it is harder to cut through the noise.
- His answer is to narrow the aperture early. The second common founder mistake, he says, is starting with a massive total addressable market and a "this could be a hundred billion dollar business" story.
- Ambition is fine, but the real question is what you will focus on in the first 30 days. Joe sums it up as earning the first dollar of revenue first.
- Dharmesh still applies this inside HubSpot: new ideas start with a deliberately narrow field of view, because the teams with the shortest iteration time and cheapest learning win on average.
3. Pick markets investors underrate
Dharmesh and his co-founder picked small and mid-sized businesses (SMB) by comparing three markets:
- Enterprise: you can charge for solving a problem, but sales cycles are long, revenue is concentrated, and a small set of customers ends up steering your roadmap.
- Consumer: outcomes are bimodal, and many products cannot charge directly, so they rely on advertising or other subsidies.
- SMB: you can charge money like in enterprise, but you can have tens of thousands to millions of customers like in consumer, so no single customer dominates.
VCs pushed back, saying no software company other than Intuit had built a multi-billion dollar SMB business. HubSpot's view was that the market was underserved and big enough, even with its painful parts. Dharmesh now sees a similar overlooked market in one-person companies:
- 2025 had a record number of new US companies, about 5.6 million, "closing on 6 million," according to US Census data he cites.
- By August of the current year, there were already over 4 million, so it is on track for another record.
- The Census Bureau predicts whether each new company will hire an employee within two years, and the share not expected to hire is now over 70% and rising.
- HubSpot Next, the company's innovation group, is building a CRM for solopreneurs aimed at this market.
4. Customer signals beat the business plan
Both of HubSpot's big product turns came from customers, not from the original plan:
- The first pivot: HubSpot's business plan for the MIT 50K competition was ERP software for small business, an "Oracle for small business," narrowed to ERP for law firms purely to test the market early. The lawyers cared little about time tracking. What interested them most was marketing and the website: could the founders help them find more ideal clients? That pushed the company to marketing software for SMB. Dharmesh also learned lawyers are the second hardest software market to sell into, after software developers.
- Product two: HubSpot never planned a CRM; it "was not in any of our powerpoint decks." About five or six years in, a survey of paying customers found 20% used Salesforce, 20% used another CRM, and 60% used no CRM at all, even though they were paying HubSpot to generate leads.
- Why they had no CRM: HubSpot concluded customers found CRM overwhelming and intimidating, and thought it would cost too much to start.
- The response: build the easiest CRM they could and make it free, "Not freemium, free," subsidized by the existing marketing software business.
- The strategic reason: existing CRM vendors were likely to move into marketing, so staying only in marketing software would not last forever. Once in CRM, HubSpot learned that winning required becoming a platform.
5. Solve the whole customer problem
At launch, HubSpot offered a broad set of marketing tools, including content management, SEO, landing pages and forms. That went against the standard advice to do one thing really well, and each category already had strong standalone products. Dharmesh says they did it because the customer's real problem was feeling intimidated by fragmented tools, so the only fix was one combined offering. His advice to founders is to define the actual problem, wherever you draw the circle, and solve all of it.
6. Build off the labs' road
Dharmesh borrows Joe's "Yellow Brick Road" image: a shiny, empty looking road toward the Emerald City, with a fleet of tanks, the frontier AI labs, coming down it and getting faster. He advises staying off the road and a little away from its edges, since the road keeps widening in somewhat predictable ways. Off the road, he says building on AI is less scary than in 2023, when one lab seemed to have a magic elixir nobody else could match:
- Models are trending toward infrastructure, like AWS and cloud computing 15 years ago.
- Tools like OpenRouter exist because swapping models is easy, and HubSpot and most SaaS companies break use cases apart and route each task to the right model.
- With open source and many strong commercial models, no single supplier controls the one key ingredient, so a startup is less exposed to having its margins squeezed.
- The labs must sell tokens to recover their huge spending, and they can do it both directly and through software companies that know a domain and act as resellers of inference.
Mental Models & Frameworks
Structural or incentive constraints
To find ideas the frontier labs will not crush, Dharmesh looks for two kinds of protection:
- Structural constraint: the labs simply cannot do it, for example for regulatory reasons.
- Incentive constraint: the labs could do it, but the odds are very low. His example is HubSpot itself: OpenAI is smart and ambitious enough to chase HubSpot's 3 billion plus dollars of CRM revenue, but that same ambition makes such a small target unlikely.
Use it as a filter when scoping an AI product: if neither constraint applies, assume the labs will eventually build it.
The interface value spectrum
Asked whether customers using HubSpot through ChatGPT or Claude Cowork erodes HubSpot's value, Dharmesh places products on a line. At one end are products like Uber Eats or DoorDash, where the app is trivial and the real value is the restaurant network and delivery logistics. At the other are products where the interface is the whole product. He argues most business applications sit near the Uber Eats end, because the value is the work getting done. He compares ChatGPT and Claude Cowork to a new operating system or browser, and notes the web browser did not let Google own the whole application stack.
Use it when deciding how much to invest in your own interface versus connectors into AI assistants. His advice for B2B builders is not to overthink the interface and instead find the best, cheapest, lowest latency and lowest friction way to deliver the customer's outcome.
Query before inference
Dharmesh's rule is that anything that can be solved deterministically with a structured query should be. His example: to answer how much revenue comes from Ireland, you should not search 8 million Slack messages and emails with a powerful model, even if the data is in there. It should be a database query. Use it when designing AI features, to decide which parts need a model and which need plain data.
Reusable primitives compound
Get the core building blocks of the software right, then remix them as the market changes. HubSpot built marketing automation as a primitive (send an email, wait three days, check for a click). When it moved into sales, that same primitive meant its sales automation had more on day one than most rivals. HubSpot is now rethinking its primitives for AI:
- System of record: a structured store for known facts, which Dharmesh insists is still needed.
- System of context: unstructured data, likely organized as a graph (a "context graph" or "second brain"), used with the structured store to drive agents and get work done.
Time to joy for developers
Dharmesh says building an ecosystem is like building a product. Just as HubSpot measures time to value for customers, it measures "time to joy" for developers: how fast someone goes from first getting something working with the HubSpot APIs to making their first dollar in the ecosystem. The goal is to compress that time.
Startup within a startup
HubSpot Next projects are cordoned off from the main company: they do not have to use HubSpot's tech stack or brand, and they are told to just solve the problem. Each project pursues a new product category, product idea, market or technology. HubSpot built its CRM this way too. When a project works, the core team picks what it wants, which Dharmesh likens to an acqui-hire. He says those degrees of freedom are what let the team run experiments at the rate it wants.
The reasoning: Dharmesh says large companies do not lack anything startups have; their problem is organizational baggage and weight. If a company can shed that, it keeps real assets: distribution, customer relationships, evals, and 20 years of code that can inform the next product. HubSpot Next acts like a startup without those liabilities, then leans on the HubSpot brand or distribution advice when useful.
Decision Principles
Principle: Create value before extracting it
- When: customers are not adopting a category because it feels too expensive or too hard to start.
- Why: HubSpot's philosophy is to get something useful into customers' hands first. That is why its CRM launched free, not freemium, paid for by the marketing business.
Principle: Skip gaps labs will close
- When: an idea rests on a current gap in ChatGPT or Claude, such as weak memory.
- Why: Dharmesh says those gaps will close. Instead of solving the horizontal problem, solve it for a specific class of customer, deciding what data to ingest and what agents and work it should produce.
Principle: Judge the company, not product
- When: judging whether a frontier lab will compete with you.
- Why: ChatGPT the product may not cover your use case, but OpenAI the company can launch a new product line, so the risk is wider than today's feature list.
Principle: Monetize early when testing pricing
- When: you need to learn what value people place on a new product.
- Why: on his one-person CRM project, Dharmesh charged very early specifically to learn pricing lessons. He was surprised by how willing people were to put a credit card down on a consumption-based model, more than on any project he had launched at HubSpot.
Trade-offs & Nuance
Focus versus solving everything
Dharmesh gives two pieces of advice that seem to conflict: narrow the aperture, yet solve the whole customer problem. HubSpot resolved it by narrowing the customer (SMB) while going broad on the problem (all the marketing tools in one place), because the customer's pain was fragmentation itself. The narrowing applies to who you serve; the breadth applies to how fully you solve their problem.
Faster building, harder distribution
AI shortens the learning loop, which favors new teams. The same ease of building floods the market with competitors, which makes reaching customers harder. Faster shipping alone is not an edge if nobody can find the product.
Rewriting old architecture with agents
Agentic coding lets incumbents attempt projects they would have skipped, such as replaying an old architecture choice and spending, in his example, 200,000 dollars on tokens to redo a body of code. The limit is proof: it only works because HubSpot has tests and evals to check the result still works. HubSpot already runs its evals on new models and weighs trade-offs such as losing about 3% in quality for gains in latency or cost. He suggests testing these rewrites on small bodies of work first.
Some software companies will die
Dharmesh rejects the "SaaS apocalypse" view but does not deny losses. He expects some software companies to die, just as some failed to move to SaaS and the cloud. His argument is that AI, like the internet did for marketing, turns a new class of problems into ones software can solve, so the total opportunity is bigger than before.
Common Mistakes
Mistake: Treating ideas as spreadsheet problems
Founders build detailed market models, often from census counts, and expect them to reveal the best idea. Dharmesh says idea quality cannot be known in advance, so the better approach is to ship and test quickly.
Mistake: Leading with a giant TAM
Founders pitch a hundred billion dollar market before knowing what they will do in the first 30 days. A broad target slows learning and makes distribution harder. Start narrow and earn the first revenue.
Mistake: Filling a temporary lab gap
Founders say "ChatGPT or Claude doesn't do this, so we will." If the claim still holds with "yet" added, the frontier labs will likely close the gap and run the startup over.
Mistake: Using AI where queries work
Teams throw a powerful model at questions a database can answer exactly. This costs more, adds latency, and is less reliable than a deterministic query.
Practical Application
Run the "yet" test on ideas
- Do: for each AI feature or product idea, write the gap it fills as "AI assistants can't do X," then add "yet."
- Then: if it still sounds true, look for a structural constraint (such as regulation) or an incentive constraint (a market too small or awkward for the labs) before investing.
- Why it works: it filters out ideas that depend on the frontier labs staying behind.
Survey customers about adjacent tools
Ask paying customers which tools they use next to yours, as HubSpot did when it found 60% of its marketing customers used no CRM. Look for a large group using nothing at all. Then ask why they have not adopted anything, since HubSpot's answer (too intimidating, too expensive) directly shaped its free, simple CRM.
Define your first 30 days
Before writing a strategy deck about market size, write down the one narrow customer and the one problem you will focus on for the next 30 days. Aim for the first dollar of revenue or the first real usage signal, not a market model.
Split model work from query work
Audit your AI feature list and mark every question that a structured database query could answer exactly, such as revenue by country. Route those to queries and save model inference for unstructured work that needs it.
Measure developer time to joy
If your product has APIs or partners, measure how long it takes a new developer to get something working, and then to earn their first dollar in your ecosystem. Treat shortening that time as a product goal, the way HubSpot does.
Questions to Consider
- If an AI assistant like ChatGPT or Claude became the main way customers used our product, how much of our value would remain beyond the interface itself?
- Which of our current ideas rely on a frontier AI lab being unable to do something "yet," rather than on a structural or incentive reason the labs will stay away?
- Are we narrowing the customer we serve while still solving that customer's whole problem, or are we doing the reverse?
- Which questions in our AI features could be answered by a deterministic database query instead of a model?
- What is the fastest and cheapest test we could run this month to learn whether our next product idea is worth pursuing?
Bottom Line
Dharmesh Shah's lesson is that good ideas are found by testing, not analysis, so build the shortest and cheapest learning loop you can and point it at a narrow market. Then solve that customer's whole problem in a place the frontier AI labs have structural or incentive reasons to avoid.
Resources Mentioned
| Resource | Type | Why it was mentioned |
|---|---|---|
| "Avoiding Death on the Yellow Brick Road" by Joe Schmidt | Essay | The a16z essay that started the conversation, arguing startups should build outside the path the frontier AI labs are racing down. |
| Joe Schmidt's essay on becoming AI native | Essay | Joe says he wrote it around the same time, using Workday as an example of a company that must rethink itself to become AI native. |
| US Census Bureau business formation data | Data | Dharmesh uses it to show record new company formation and that over 70% of new companies are not expected to hire within two years. |
Case Studies Mentioned
Twitter's closed developer platform
Dharmesh recalls that early Twitter was an open ecosystem where many clients and companies were built on its APIs. It later closed down and became, in his words, developer unfriendly to almost developer hostile. HubSpot took the lesson to heart and tries to be open with partners about where it is unlikely to compete.
Agent.ai split into two
Agent.ai was a HubSpot Next project with two parts: a marketplace for AI agents and a builder that let regular people create agents mixing deterministic and AI steps. Dharmesh says it was too horizontal, and he jokes it was a time he accidentally competed with Sam Altman. The agent builder turned out to be the right half, but it needed the full HubSpot team, so it was moved into HubSpot as a product. The lesson is that an experiment can partly succeed, and the valuable part can be absorbed into the core business.
Marketing as a software problem
When HubSpot raised money, VCs said there had never been a billion dollar marketing software company. Dharmesh argued that every line on a business's profit and loss statement had a 100 billion dollar plus software company attached, except marketing, because marketing had been an arts and crafts discipline. The internet made marketing look like a problem software could solve. He makes the same argument now about AI creating new classes of problems software can solve.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| HubSpot | CRM and marketing software platform for small and mid-sized businesses | The main case study for choosing a market, adding a second product, and building an ecosystem. |
| ChatGPT and Claude Cowork | AI assistants from OpenAI and Anthropic | Dharmesh calls them the new operating system or browser for software; HubSpot was the first CRM to ship a connector to ChatGPT, over a year before the recording. |
| OpenRouter | Service for routing requests across many AI models | His evidence that models are becoming swappable infrastructure. |
| Agent.ai | Marketplace for AI agents plus an agent builder | A HubSpot Next experiment whose agent builder became a HubSpot product. |
People to Follow
Dharmesh Shah
Co-founder and CTO of HubSpot, who says he has built commercial software every day for over 30 years and has never had direct reports at HubSpot. He leads HubSpot Next experiments, including a new CRM for one-person businesses, and is on X as @dharmesh.
Joe Schmidt
Partner at a16z and author of the "Avoiding Death on the Yellow Brick Road" essay about where startups can build alongside the frontier AI labs. He is on X as @joeschmidtiv.
Experiments to Try
Experiment: Charge from day one
- Hypothesis: charging for a new product early teaches more about its value than a long free period.
- Change: launch the next small product or feature with a paid, usage-based price from the start, as Dharmesh did with HubSpot's one-person CRM.
- Test: compare sign-up and paid conversion with a past free launch of similar size.
- Measure: share of users who add a credit card, usage per paying user, and what they say they are paying for.
Experiment: Evals on cheaper models
- Hypothesis: some AI tasks can move to a cheaper or faster model with little quality loss.
- Change: run your existing eval set across several models for each task, as HubSpot does when new models come out.
- Test: route a small share of live traffic for one task to the best-value model.
- Measure: quality score change (HubSpot cites trade-offs like losing about 3% of quality), latency, and cost per task.
Notable Quotes
"I think the people that win are the ones that have the shortest iteration time and the cheapest learning time." (Dharmesh Shah)
"Don't use inference to solve something that a query can solve." (Dharmesh Shah)
"Because it's so much easier to build software now, many more people will be doing it." (Dharmesh Shah)
"The interface is a means to an end." (Dharmesh Shah)
