What Software Is Worth After AI Takes Over the Interface: Dharmesh Shah’s HubSpot Argument

1. Guest Background

Dharmesh Shah approaches the software industry’s AI anxiety from inside a company that has already changed its product boundaries substantially. He is HubSpot’s co-founder and chief technology officer. The supplied guest background describes HubSpot as a software company serving small and midsize businesses, initially integrating marketing applications to reduce the complexity of fragmented tools and help customers grow. That experience gives the conversation its recurring subject: how technical capability becomes something a customer can actually use.

Shah says he has spent more than 30 years building commercial software and is enjoying building more than ever. He sees a renewed version of HubSpot’s original mission. Small businesses once needed help benefiting from the internet because its tools felt overwhelming; now the same customer group faces an intimidating new technology in AI. His enthusiasm therefore concerns both new products and the adaptation of an existing business.

The approximately 59-minute interview examines what deserves to be built, what established software companies must change, and what happens when customers reach applications through AI. Shah makes his interests explicit: he speaks as HubSpot’s co-founder and discloses an investment in OpenAI. His account is valuable as a participant’s reasoning, while his competitive forecasts remain judgments rather than independent evidence of how the market will develop.

2. What the Episode Covers

Shah places AI within software’s long movement toward higher abstraction. Better tools make it easier to express an idea and widen participation in building, while the business problems underneath remain. What excites him is the shorter learning cycle: release something, put it in front of actual customers, and discover sooner what might previously have taken months or years to learn. The interview supplies no measured productivity multiplier. Its claim concerns the possibility of earlier feedback.

That distinction shapes his advice about ideas. An elaborate spreadsheet cannot establish that a product deserves to exist, and counting small law firms does not demonstrate demand. He asks founders to narrow their initial focus and explain what they will do in the first 30 days. Distribution remains a separate difficulty. In his view, easier production increases the number of competing options and can make reaching a customer harder.

HubSpot’s origin provides a concrete example of revising the question. Shah recalls choosing SMB customers because they seemed to combine willingness to pay with a broad potential customer base. Enterprise software brought long sales cycles, concentrated revenue, and customers capable of redirecting the roadmap; consumer software posed different monetization problems. The original product concept was an integrated ERP for small businesses. A business-plan competition narrowed it to law firms under the Legal Spot name. Conversations then revealed greater interest in websites, marketing, and finding suitable clients than in back-office functions. The marketing direction followed those discussions.

CRM was another discovery rather than a founding blueprint. Shah explicitly rejects the idea that the company began with a planned destination as a CRM platform. He recalls a survey around its fifth or sixth year: 20% of customers used Salesforce, 20% another CRM, and 60% none. Those figures describe his account of that customer survey, not a universal adoption rate. The gap mattered because customers were already paying HubSpot to generate leads but lacked a system for managing what followed.

Several considerations converged: customers feared complexity and cost, sales was adjacent to marketing, HubSpot needed another product, and CRM vendors might move into marketing themselves. Shah says the response was a particularly easy, free CRM, explicitly distinguishing free from freemium. An existing marketing business could subsidize it. The platform direction emerged through subsequent iteration. Likewise, the earlier suite’s breadth across content management, SEO, landing pages, and forms served a defined problem: fragmentation. Focusing on a customer group did not require reducing its problem to one feature.

The interview connects that history to HubSpot Next, an experimental mechanism allowing projects outside the existing stack and brand. Shah’s current example is USpot, a CRM for solopreneurs that was only 30 days old at recording. It demonstrates continued experimentation with a narrow audience; its age also limits what the episode can establish about its eventual business.

3. Core Views: Reasoning, Examples, and Limits

Shah’s argument for independent AI software begins by taking the frontier labs seriously. A startup should ask whether adding “yet” to a missing capability leaves its pitch intact. It should also distinguish what ChatGPT currently does from what OpenAI as a company could build. A temporary product omission is weak evidence of durable opportunity. He suggests examining structural constraints, including regulation, and incentive constraints: places a model company could enter but may have insufficient reason to pursue.

His CRM example exposes both the usefulness and the uncertainty of that framework. Citing HubSpot’s revenue as more than $3 billion, Shah argues that frontier companies have the ability to target the business but ambitions that may make such a focused campaign unattractive. This is an allocation-of-attention thesis. It does not establish that they will avoid CRM, that incentives will stay fixed, or that a different product could not reach the same work. The framework helps identify assumptions; it does not confer protection.

He also sees room for commercial coexistence. Model companies can sell inference directly to consumers and large enterprises while reaching other customers through software businesses with domain knowledge and distribution. His cloud-infrastructure analogy explains why those channels need not be mutually exclusive. He reinforces this argument by revising his own earlier view: in 2023 he thought leading model companies might be too far ahead to catch. Commercial alternatives, open-source models, and task-specific routing now make him less worried about one supplier controlling the crucial input. He stops short of calling models commodities. More choice does not itself demonstrate costless substitution or equal performance on every task.

A second distinction separates the customer’s entry point from the work performed underneath. Shah reports HubSpot connectors for ChatGPT and Anthropic, then uses delivery platforms to illustrate value beyond a visible app: partnerships and fulfillment capabilities remain important even if the interaction changes. He explicitly excludes the inference that business software therefore possesses the same network effects. The comparison supports a narrower claim about operational value. He also acknowledges a spectrum: some products depend much more heavily on their interfaces. His rejection of an almost entirely interface-based valuation of HubSpot cannot settle where every application belongs.

Thinking of AI interfaces as browsers or operating systems redirects product decisions toward outcomes, cost, latency, and friction. Shah’s strongest architectural example is deliberately ordinary. If someone asks how many customers are in Ireland or how much revenue they generate, a database query should answer. Inferring the result from messages is an unnecessary replacement of a deterministic operation. His references to millions of messages and a supposedly extraordinary model IQ are illustrative rhetoric, not benchmark results.

This is why his proposed architecture retains a system of record alongside a system of context. Known facts need persistent storage; unstructured information can supply context for inference and agent execution, potentially through a graph. A “second brain” complements structured records. The interview does not compare graph implementations or prove that one representation is best, so this remains an architectural direction rather than a demonstrated universal design.

Existing software can contribute reusable capabilities to that design. Shah describes marketing automation primitives—send an email, wait three days, wait for a click—being recomposed for sales automation. The example locates accumulated value in working capabilities, not simply in old screens. It also suggests a test for incumbents: identify which parts of their accumulated software can serve new work, rather than assuming all historical complexity is an asset.

His broader rejection of the SaaS apocalypse follows from an expanding problem set. He recalls arguing that the internet would make marketing more amenable to software; AI may similarly make previously inaccessible work tractable. The interviewer adds that vertical software could create more value for small customers beyond earlier combinations of low contract values and payments income. Neither argument demonstrates that new revenue will exceed new costs for a particular company. Shah expressly allows that companies will fail. His optimism concerns software’s aggregate opportunity, leaving individual businesses responsible for earning their place within it.

4. Learning and Application

For a founder, the most useful application is to turn a market thesis into observable customer behavior. Following Shah’s first-30-days question, define the customer, the work to complete, how to reach someone willing to try the product, and what response would justify changing direction. Legal Spot illustrates why these questions matter: conversations shifted the priority from the proposed back office to customer acquisition. The transferable lesson is permission to revise the problem. Cheaper development produces cheaper learning only when the experiment reaches real users.

Platform teams can apply the same discipline to developers. Shah invokes Twitter’s retreat from its early open ecosystem as a warning about attracting builders and subsequently constraining them. His description of HubSpot emphasizes communicating direction and acknowledging what it cannot promise. Treat a developer’s first successful API call and first earned dollar as different milestones, then investigate the obstacles between them. Technical documentation may solve the first problem while leaving commercialization untouched. His “time to joy” framing makes both part of the product experience, although better communication cannot eliminate future competition from the platform itself.

Engineering teams should read the architecture discussion as an invitation to controlled experiments. Select a bounded behavior, compare an alternative implementation against existing tests and evaluations, and examine quality, latency, and economics together. Shah’s $200,000 token budget, Rust rewrite, and roughly three-percentage-point quality concession are hypothetical examples. They are not reported HubSpot outcomes or acceptable thresholds to copy. This approach depends on evaluations covering the work that matters; passing an incomplete test suite cannot establish that a broad replacement preserves everything customers need.

For memory or second-brain products, specify the audience before the storage mechanism. Shah advises against merely repairing a current weakness in general-purpose AI memory. Ask which information that customer needs to ingest, what context it supplies, and which task an agent should accomplish with it. If these remain vague, a memory layer has not established demand. The useful boundary is a customer problem whose value can be tested even as general platforms improve.

An established company also needs a route from experimentation into operations. Shah says Agent AI’s Agent Builder was absorbed into HubSpot. USpot might be absorbed, remain independent, or contribute only reusable primitives. A practical adaptation is to grant early freedom over stack and brand, then decide what deserves access to distribution, customer relationships, evaluations, and accumulated code. An experiment need not justify importing its entire product to produce something useful.

Pricing should generate evidence as well. Shah says USpot charged early to learn how users valued it and reports unusually strong willingness to provide credit cards for consumption billing. No conversion rate or revenue is supplied. Teams can therefore treat willingness to start paying, continued use under actual bills, and acceptance of the charging mechanism as separate questions to investigate. These are applications of his learning principle, not results reported in the episode.

The same caution applies to market sizing. Shah cites about 5.6 million US business formations in 2025, more than 4 million by August of the recording year, and a predicted non-employer share above 70%. The supplied evidence does not independently verify these figures. A prediction that a business will not hire also does not establish its demand for a particular CRM. A solopreneur thesis should separately test the market narrative, the early payment signal, and durable customer need.

Source

• Original episode: Why HubSpot’s Co-Founder Is Betting Against the SaaSpocalypse