Des Traynor on AI Products, Competition and Building Intercom

Des Traynor, co-founder of Intercom, tells Harry Stebbings that he once wanted to write about sport. Intercom emerged from an earlier software product when its team added an in-product message bubble to talk to customers. A Dublin coffee shop’s personal conversations with regulars helped inspire a more direct connection between software businesses and their users.

Traynor is skeptical of “AI-first” labels when a product’s core job remains unchanged and AI is merely added as a small feature. He thinks customer support is more fundamentally affected because summarization, knowledge retrieval and conversation are central workflows. Stebbings argues project management could also change substantially; Traynor agrees in principle if the capabilities are actually built.

He says Intercom’s existing messaging interface gave it a useful starting point for AI customer service, although it also supports tickets. He traces a progression from rules-based bots to fuzzy intent matching and then language-model-based responses. The company’s Fin product uses that newer approach, according to his account.

Traynor distinguishes a simple model wrapper from a product that supports an entire customer workflow: knowledge ingestion, refreshes, reporting, permissions, customer context and integrations. He argues that depth can differentiate an application even when its underlying model comes from another company. As an Intercom co-founder promoting Fin, he has a direct commercial interest in this argument.

Stebbings points out that model-provider releases can erase startup features. Traynor agrees a narrow bot may be vulnerable, but argues a product that integrates deeply with a specific workflow may have room to survive. His wealth-management and creative-tool examples are hypothetical, not guarantees of defensibility.

Model commoditization would help Intercom by lowering its costs, Traynor says; alternatively a specialized model trained on support conversations might confer an advantage. He declines to say whether the company is training its own. This is an explicit description of the incentives behind his views, not evidence that commoditization has occurred.

Traynor says Fin was producing high resolution rates for some customers and was then using OpenAI’s GPT-4. His team assessed conversation quality, staying on topic and hallucination behavior, but he supplies no comparative numerical benchmark. Reported customer outcomes and product reliability are company claims, not independently verified results.

He says model quality was close and changing quickly, and that the team had not tested every provider’s newest release. Switching models could be useful when one improves, although he says provider flexibility alone is not enough to make a company win. Speed of upstream changes also forces product teams to reassess their plans.

Traynor likens the AI shift to the early internet and believes customer-service vendors cannot simply wait while the market changes. His confidence in Intercom’s opportunity is a 2023 prediction. He describes a central AI team working on Fin, model selection and riskier research, then partnering with product teams to apply new capabilities.

He estimates Intercom had roughly 350–400 people in research and development out of about 900 employees, and anticipated an AI team of around 50. These are approximate headcount statements from the interview, not a current staffing report.

Asked whether unusually many AI startups will fail, Traynor expects higher mortality from thin products overtaken by platforms and from model costs compressing margins. He also thinks established products could be disrupted. Stebbings’ baseline failure percentages and Traynor’s prediction are conversation estimates, not verified industry outcomes.

Traynor imagines an advertising optimization workflow where AI might generate variants, run tests and report results. Stebbings asks whether an end-to-end workflow is supposed to be defensible. Traynor replies that even an entire workflow may lose value if its steps can be automated and human judgment is not needed. This is a scenario, not evidence that such systems reliably automate every step.

His test for disruption is whether a new entrant would rebuild an incumbent’s product very differently, or whether it would still need the incumbent’s infrastructure and distribution. A startup adding AI writing assistance to email marketing must reproduce sending, integrations and reputation, whereas the incumbent may only need to add the writing feature. He presents this as a framework, with exceptions possible.

Stebbings cautions that large enterprises can adopt slowly and experimental budgets can disappear. Traynor says investors must consider multi-year adoption, not only the next year, and claims Fin’s value was visible when an OpenAI outage shifted support back to people. That anecdote is not an independently measured return on investment.

Traynor predicts more AI pricing will be tied to work or consumption rather than employee seats. He says some Intercom customers were retaining staff and using automation to improve service, while Stebbings asks whether that reflects caution rather than a lasting preference. Both the labor effect and pricing shift remain uncertain.

He expects a spectrum of models and chips with different prices, speed and quality, and competition to reduce costs over time. Asked whether he would invest in OpenAI at a hypothetical $90 billion valuation, he says he might make a small speculative investment but would hesitate to commit heavily, partly because of competition and the return multiple required.

Traynor says he was underwhelmed by Google’s perceived leadership response to AI and believes the company faces a difficult choice about disrupting its own search business. As a hypothetical CEO, he considers sponsored content in model answers, explicitly describing it as paid material. These are speculative business ideas, not claims that Google had adopted them.

He predicts Apple could use its consumer devices and privacy-oriented design to make assistants more useful, potentially mediating which services users reach. He also speculates about possible Meta uses in messaging, social media and virtual reality. The discussion is hypothetical, and it does not establish the capabilities or data practices of any product.

Traynor sees an opportunity for Adobe to broaden access to creative tools if AI helps people express ideas they cannot draw or edit themselves. He bases the view on demonstrations he has seen. He also expects infrastructure competition to shift value toward differentiated products, while acknowledging cloud providers will remain important.

Stebbings notes that a better standalone product may lose to an adequate feature bundled with an incumbent’s existing contract. Traynor agrees a new tool must overcome integration, procurement and distribution advantages. He thinks incumbents may benefit substantially from AI, while startups can create new categories that were not previously possible.

They debate AI video creation and the extent to which existing editing platforms might absorb it. Stebbings discloses an investment in Synthesia; Traynor says he previously met its founders. Their discussion of the category’s prospects and competitive advantage consists of their opinions and forecasts, not verified future market outcomes.

As an angel investor, Traynor says he wants founders of B2B companies to have considered where AI could change the product, even if they decide not to use it. He says recent checks had leaned toward AI-related companies. His own investments and role at Intercom shape the perspective.

He criticizes AI features added solely to signal trendiness. His practical test is whether something previously impossible is now possible, or whether a user can avoid work they previously had to do. This is his product criterion, not a claim that an AI feature automatically delivers a net benefit.

For B2B marketing, Traynor says a message should be distinctive, useful and simple enough to understand. He favors describing what the product does and why it differs from alternatives, without losing a human voice. A company’s origin and broader mission can matter in another context, but should not obscure the immediate product proposition.

He regrets growing Intercom’s content audience without focusing early enough on people likely to buy customer-support software. Stebbings argues a broad audience can still create reach and word of mouth; Traynor agrees both approaches can work, but would choose greater early focus. He now values engagement from the relevant audience more than a viral post alone.

If he could revisit Intercom’s product strategy, he would invest earlier in integrations with the rest of a customer’s stack. He also warns that startup tactics can expire as markets change: advice to move upmarket through sales has coexisted with successful self-service businesses. Founders should separate durable leadership skills from time-sensitive playbooks.

He says experience can teach a founder how to hire executives, build culture, identify gaps and face unwelcome facts, even when specific channels and tactics change. Stebbings asks whether culture itself changes as work preferences evolve. Traynor agrees management must adapt and says a growing company needs ways to develop people at different experience levels.

His leadership mistake was softening difficult feedback so much that the recipient could not see the seriousness of the issue. He now tries to be clearer about the skills a person must develop. On layoffs, he says hiring discipline matters because reductions are disruptive and emotionally costly; he also acknowledges the temptation to spend available venture capital on additional people.

Traynor says having children made him more conscious of using working time efficiently, although he cannot isolate parenting from other changes such as the pandemic. He describes the challenge of preserving energy for family after work rather than claiming a general causal rule about parents.

In quick-fire discussion, he says he would like to improve collaboration across Google’s meeting and document tools, and imagines a single destination for live sports while recognizing rights as a challenge. These are product ideas, not announcements or evidence that such a service is easy to build.

He recommends understanding competitors with respect: a rival may win through distribution even if its product appears weaker. A team should make sure beating a competitor serves a worthwhile goal rather than becoming the goal itself. He describes Intercom’s renewed focus and a 52-week company-wide effort to rebuild several functions after its CEO returned.

Traynor says he asked an Equals founder to publish regularly because companies can fail to explain their product, vision and reasons to use it. He encourages founders to communicate repeatedly and allow their thinking to evolve. He notes his own early role writing Intercom’s blog; his investment in Equals is part of the context.

Discussing Hopin as an outside observer, he suggests the business needed to recognize the return of in-person events and adapt quickly. He and Stebbings also reflect on selling some shares when private valuations become very high. These are retrospective opinions, not a complete account of Hopin or investment advice.

Looking ten years ahead, Traynor hopes Intercom can become a primary support platform for internet businesses. He also hopes his role will require less intensity as he ages and spends time with family. Both are aspirations from the 2023 interview, not predictions of realized company performance.