Tobi Lütke on AI Agents, Entrepreneurship and Shopify’s Future

Alex Heath opens by asking Tobi Lütke about Amazon’s decision to block Meta’s Muse shopping agent. Lütke says he learned about it after publicly welcoming Muse to Shopify, which made the contrast between the two companies look especially pointed.

Lütke sees the difference as a continuation of two longstanding approaches to commerce rather than proof that one philosophy is inherently correct. He describes Amazon as an everything store whose strength is breadth of inventory, and Shopify as serving independent merchants who distinguish themselves through products and storytelling.

Heath asks whether large platforms’ advertising businesses help explain their resistance to agents. Lütke says he does not know Amazon’s reasoning. Shopify deliberately built its business around merchants, he says, rather than creating a traditional marketplace that must balance its own interests with those of buyers and sellers.

It is too early, Lütke says, to know definitively whether an agent-driven web favors Shopify’s model. He understands why aggregators want control: some agents use browser automation and impersonation in ways that can be unsettling, even if the approach makes new things possible.

Shopify’s immediate goal is to make checkout work whether a person or an agent acts on the customer’s behalf. Lütke says the company is building APIs and systems for that purpose and intends to support different ways of buying.

Asked which agents Shopify wants to support, Lütke points out that merchants already sell through online stores, social channels, point-of-sale systems and craft fairs. Their own store remains a place for design and storytelling, but he believes making their products available in more places makes Shopify more useful.

Heath asks what early shopping behavior Shopify can see. Lütke says agent-led purchasing is increasing from a low base, so high percentage growth should not be overinterpreted. What interests him more is the character of the purchases, particularly considered, project-based orders.

One example is setting up a home gym: a user may ask an agent to assemble everything needed, perhaps after discussing the room or a floor plan. Lütke imagines that some of the planning happens outside Shopify’s view, while the resulting purchases can have a relatively high order value.

Another example is a weekend project with a child and a 3D printer. An agent might identify the electronics, magnets and filament needed and arrange orders so the supplies arrive in time. Lütke says the uses are evolving quickly and that people are sharing their discoveries with one another.

Lütke says he enjoys how futuristic these examples feel. A science-fiction reader since childhood, he likes comparing authors’ imagined futures with current capabilities, while recognizing how difficult it is to predict details many decades ahead.

He describes connecting an agent to a home network and, if the user is comfortable with the access, to a 3D printer. In his example, it could reorder filament. He sees today’s agents largely helping people do existing activities better, with genuinely new activities potentially emerging later.

Asked which fiction anticipated the present, Lütke mentions rereading Neal Stephenson’s Cryptonomicon and seeing ideas that now feel familiar, including digital money and fractional ownership. For more distant futures, he says, individual predictions are often wrong even when pieces of the larger picture resonate.

He also finds that science fiction reveals the period in which it was written. He uses the changing origins of Spider-Man as an example of contemporary anxieties entering stories. In his view, popular near-future fiction spent a long stretch favoring dystopian settings over optimistic ones.

Lütke welcomes what he sees as a return of optimism in technology. He finds it remarkable that people now train software and then test it to discover its capabilities, and he describes patient digital assistants as a way to make everyday tasks easier.

Heath asks whether invention itself produces optimism. Lütke agrees and argues that people benefit from having a frontier of opportunity. Where there is visible growth, he says, people can pursue new ideas instead of treating every gain as a contest over a fixed pool.

He compares the feeling to joining a startup that is succeeding: growth makes optimism easier. He thinks pessimism often sounds more sophisticated, but can discourage people from taking opportunities, reinforcing a cycle of limited growth and zero-sum competition.

Lütke characterizes AI as a new frontier that gives companies a chance to reconsider established assumptions and rebuild products. This is an optimistic interpretation of the opportunity, not a guarantee that every application will succeed.

Heath connects the argument to abundance and asks why zero-sum thinking is harmful. Lütke says it implies that each winner needs a loser, whereas new knowledge and products can create value that did not previously exist.

As examples, Lütke points to computation and to knowledge about nuclear energy. He argues that useful resources often become economically valuable only after people discover how to use them. His larger point is that a fixed view of available value can obscure opportunities for invention.

He brings the argument back to Shopify merchants: someone might decide during a lunch break to leave accounting and build a product that finds customers. He says seeing such businesses form gives him a favorable view of entrepreneurship, while acknowledging that not everyone has the same vantage point.

Heath asks how that optimism fits with growing negative sentiment about AI, which may vary by geography and economic circumstance. Lütke cautions that he may not be representative, but says entrepreneurs are particularly well placed to benefit because they must understand many different jobs.

Lütke recalls a survey in which many Shopify users said they could reach someone who would answer an entrepreneurial question quickly. He interprets the result partly as survivorship bias: people able to build businesses may already have had such access. He believes AI can make some of that advice easier to obtain.

He invokes the myth of Prometheus to discuss the intuition that powerful advances carry equally powerful negative consequences. Lütke acknowledges that AI can be used for good or harm and can affect individuals differently, but says he expects the positive effects to outweigh the negative ones.

He states that he personally assigns essentially no probability to an AI-doom scenario, a subjective assessment rather than a demonstrated fact. Heath agrees that AI can be understood as leverage, while suggesting that some public anxiety comes from leaders describing it as an entity people may not be able to control.

Heath asks whether the political and public reaction could worsen before it improves. Lütke says people have a right to pay close attention and that the public discourse can make it hard to form an opinion. He expects direct experience with useful tools to make attitudes more positive over time.

He says customers are beginning to experience computers as more capable, closer to how programmers once described them. When he was young, he could write software to give a computer new abilities, but that skill required years of practice and did not spread easily to everyone.

Lütke believes newer tools can narrow that gap by helping people express what they want and collaborate with software. He describes the ambition as giving people more assistance in pursuing their goals, rather than requiring every merchant to become a programmer.

Asked about new business formation, he says Shopify is seeing strong activity and younger entrepreneurs. He credits both AI and the entrepreneurial elements of online creator culture, while acknowledging that parents and children may disagree about particular media.

He uses the visible progression of large creators as an example: audiences can watch resources being gathered and projects grow in scale. Lütke thinks following such a journey can make the effort and possibilities of entrepreneurship more concrete to young people.

Lütke says adoption also varies by region. He characterizes the United States as especially quick to adopt new tools, Europe as more skeptical and Canada as somewhere in between. These are his observations and generalizations, not a controlled comparison.

Discussing Canada, he points to Toronto’s research history and Geoffrey Hinton’s work as sources of local pride. He says the conversation there includes how AI could improve companies, government services and productivity, even if enthusiasm is less unrestrained than in Silicon Valley.

Heath asks whether that energy makes Lütke want to move Shopify to California. Lütke says no: he enjoys Silicon Valley but values observing it from outside, filtering its ideas and making decisions with some distance from its rapid swings.

Lütke criticizes what he sees as convergence in company cultures, including widespread reliance on similar planning systems. He enjoys building organizations that make different choices and believes people benefit when companies have distinct cultures rather than all trying to appeal to everyone.

Heath asks about Lütke’s experience on Meta’s board and working with Mark Zuckerberg. Lütke praises Zuckerberg’s ability to lead and change a large company, and says Meta’s release of Llama models helped establish a norm that other companies could follow.

Lütke says the availability of open models contributes to his optimism because it supports competition. He discusses what he describes as a setback with Llama 4 and praises Meta’s response and its effort to assemble a new training team. These are the impressions of a Meta board member.

From that board vantage point, he says, he is grateful Shopify is smaller than Meta. He admires the challenge of managing a much larger organization and does not assume that the leadership skills required at one scale automatically transfer to another.

After a sponsor break, Heath asks whether there were moments when Shopify particularly benefited from having its founder as CEO. Lütke argues that a founder can sometimes cut through bureaucracy and act on a conviction faster than a hired manager constrained by a more formal process.

He gives his memo calling for reflexive AI use as an example. Once he had formed a view about where the technology was going, he could tell the company that teams should experiment with AI and ask how those experiments had failed before requesting additional resources.

The memo attracted controversy outside the company. Lütke says Shopify employees largely recognized it as consistent with what they had already heard from him, and that the resulting press attention prompted some internal conversations about work already under way.

He describes present-day reviews as full of experiments that surprise him, and says he sees a company of people who enjoy working with computers. He presents his enthusiasm as an account of his own experience, not as an independent measurement of every employee’s experience.

Lütke says he does not operate from a fixed annual plan. Instead, he tries to teach people a way of evaluating current inputs and choosing the next useful step, with teams moving in a shared direction while remaining loosely coupled. He thinks that arrangement can support creativity.

Asked whether the company needed him in particular, he says his long-term ambition is for Shopify to endure without depending on him. While he remains CEO, he wants to contribute as much as he can; he recalls expressing the longer-term ambition in a letter around the company’s public listing.

Heath returns to two debated elements of the 2025 memo: AI in performance reviews and asking whether AI can address a need before adding headcount. Lütke says the habits now happen organically, though he is unsure whether the specific peer-review question remains in the current questionnaire.

He explains that including a question in a review can signal what a company values. He also criticizes projects that reject AI-generated contributions solely because of how the code was produced, arguing that the code itself should be assessed. That is his position in a broader disagreement about contribution policies.

Lütke links explicit expectations to his preference for distinct company cultures. He says an AI skeptic may prefer a workplace whose practices match their beliefs, rather than one where colleagues rely heavily on agents. His argument is that making such differences visible helps people decide where they can thrive.

Heath asks how personal agents will make money. He hears many founders discussing a small share of transactions, but questions where that share would come from when existing commerce fees already have to be divided among multiple parties.

Lütke replies that payment networks receive much of the roughly three-percent figure Heath mentions for Shopify. He is skeptical that every agent will receive a two-percent transaction fee, but believes products that provide enough value can find ways to be paid.

He suggests thinking about how much someone would need to be compensated to give up a useful agent, while recognizing that the answer will vary. The business model could depend on unique capabilities, distribution, onboarding or direct user payments rather than a single universal transaction fee.

Lütke observes that large consumer companies have sometimes built an audience before optimizing monetization, provided they can finance that period. Heath notes that AI startups face substantial compute expenses. Lütke agrees that financing conditions matter and recalls the infrastructure costs of an earlier internet era.

He adds that his own situation is unlike that of a new founder raising money today. He remembers asking his father-in-law for help meeting payroll in Shopify’s early days, whereas the company’s present scale gives him more room to experiment without the same immediate financing pressure.

Heath describes current startup valuations and the impression that capital is being deployed rapidly, much of it toward GPUs. Lütke jokes that agents expecting large transaction percentages will meet the world of basis points, while maintaining that useful services can create monetizable value.

Lütke says he uses substantial amounts of AI both at work and privately and can afford the expense, while recognizing that this puts him in a different position from many users. He expects costs to fall and pricing or metering models to evolve, but does not claim to know which model will win.

Possible mechanisms, he says, include metered use, transaction flows or other arrangements. He also thinks agents might help people identify unnecessary expenses, such as excessive financial fees, and that saving users money could itself create value.

Heath asks whether Shopify’s AI tools could let the company take a larger share of merchant activity. Lütke says he would prefer a lower percentage on a larger volume: he focuses on growth in gross-profit dollars rather than maximizing the percentage taken from each transaction.

He ties that choice to Shopify’s mission of making entrepreneurship more common and easier to attain. He says many people start a Shopify business around their local lunch break, perhaps when considering an alternative career path or testing an idea alongside a job.

A small venture need not become enormous to be worthwhile, he says; it can become the kind of independent business its creator wanted. His goal is to reduce obstacles at the beginning, including making it easier for AI systems to help people sign up.

Lütke says Shopify can spread engineering costs across businesses at many scales, including large merchants, while investing in tools for new founders. He sees this as a way to make the first steps easier even though beginners are not the source of most revenue.

Asked about risk to Shopify, he answers that someone might build a better version. He says he regularly imagines how he would design a competing product and uses that exercise to identify what Shopify should improve.

The challenge, he says, is serving both someone opening a first store during lunch and a business operating at very large scale. Existing customers also create obligations that can make evolution harder. He wants the company to move quickly enough to compete with a small team using agents.

Lütke compares the situation to Shopify’s beginnings, when a small company could challenge expensive incumbent software with a new web-based approach. He acknowledges that Shopify is now an incumbent itself, and wants it to remain responsive to challengers using AI.

Heath asks whether that requires deliberately disrupting existing products. Lütke agrees that sunk costs can prevent change. He suggests imagining the best version of Shopify if it were built from scratch today, then comparing that picture with the existing product to identify the gaps.

Heath says the underlying mission of supporting entrepreneurship seems durable. Lütke responds that trade and merchant activity have deep historical roots. He sees Shopify as serving one part of a much older human practice, even as the tools used to trade change.

Asked to summarize his advice for agent startups, Lütke says the immediate focus should be creating enough value to be hard to ignore. He extends the idea to careers: build skills and capabilities that people find genuinely useful, while recognizing that customers have limited resources.

He and Heath compare consumer subscriptions with everyday spending such as a cup of coffee. Lütke’s point is that even a small recurring expense competes with something people already value; a product has to earn its place in that budget.

Heath asks how advertising-supported aggregators should respond to agents. Lütke advises returning to the underlying problem a company solves, rather than treating its existing ad system or subscription fee as permanent. A different solution may eventually support a different way of making money.

Lütke says he is not pessimistic about advertising itself. He expects discovery and growth mechanisms to remain necessary so that smaller companies can reach customers, though the form of the market may change. Heath says the unresolved question is how that will work when an agent acts for an individual.

The trust issue is direct: if an assistant is paid to favor a product while presenting itself as acting for its user, the arrangement could feel like a betrayal. Lütke says companies need to find acceptable ways to connect merchants and customers without undermining that relationship.

He worries that an AI-driven shift might concentrate commerce among the largest businesses. In his account, the internet helped smaller merchants reach audiences for distinctive products rather than relying only on the shelf space available in large stores.

Lütke invokes the idea of finding a thousand true fans. A product that some people love can overcome indifference even if it is not designed for everyone. That requires a sufficiently open discovery market in which a less familiar company can present its product to potential buyers.

He credits advertising systems, especially Meta’s, with helping small businesses find customers, while also acknowledging harmful aspects of advertising. He says figuring out how discovery and matchmaking work for agents is both important for merchants and a potential business opportunity.

Heath turns to Lütke’s post about replacing a wine cellar with a server rack. Lütke says the room had been underused and describes the project as an expression of his long-standing enthusiasm for computers, with equipment procurement and cable management becoming part of the fun.

He says the rack includes a Dell GB300 system and can run models locally. As a public-company CEO, he explains, there is information he cannot simply send to an outside laboratory. A local setup lets him experiment with an assistant while keeping that work in his house.

Lütke describes himself as a digital-archiving maximalist: he has kept a great deal of what he has written and many logs of his computer activity. He also notes that public information about him may already be present in model training data, making his experience with agents unusual.

He recounts asking a new assistant to guess his identity during setup. He says it inferred who he was from contextual clues such as the local username and location. He presents the story as an example of his own setup, not a general claim about what an agent should infer.

His assistant performs research and prepares briefings, he says, drawing on information that can include his calendar, location and Fitbit data. He treats the ability to combine personal records as part of what makes the tool useful to him.

For a personal experiment, Lütke says he compared heart-rate data with meetings to identify which interactions were most stressful, and told the people involved. Heath reacts with surprise. The anecdote describes Lütke’s own use of sensitive information, not a recommendation for workplace monitoring.

Asked whether he could live without the assistant, Lütke says he could, but losing the capability would feel like a significant regression. He reiterates his optimism that people can benefit from technological progress, while Heath closes the interview.