Shopify CEO on AI: Cheaper Output, Better Judgment

Original source: Shopify CEO: Most Businesses Are Using AI Wrong

Program: The Knowledge Project Podcast. Guest: Tobi Lütke, Shopify founder and CEO. Host: Shane Parrish. Video published on 2026-09-15. Duration: 01:04:49. Official episode page.

This article draws on the public interview. Figures about internal adoption, working practices, and personal efficiency are Lütke's descriptions and estimates.

A Shopify team discussing a feature in Slack can bring River into the conversation and ask it to summarize the discussion, investigate relevant research, or build a prototype.

River is an internal AI colleague. Lütke estimates that as many as roughly 50% of Shopify's pull requests now originate in conversations of this kind. That figure concerns proposed code changes. It cannot be read directly as a share of merged code, successful deliveries, or labor saved.

In the same interview, he describes a problem that AI has made worse: people generate substantial amounts of work, barely understand it, and hand it to colleagues to review.

As producing work becomes easier, someone still has to decide whether it deserves to exist, whether it is correct, and who owns the consequences. That concern connects River's design, Lütke's personal decision process, and his ideas about building an enduring company.

How River turns conversation into work

In his description of Shopify's internal AI, Lütke presents an agent embedded in the place where work already happens. He describes a Slack environment with 7,000 people and 10,000 channels. River can be invited into a channel, use the relevant context, access code and tools, and propose a change.

He says its operations are sandboxed, but the interview does not explain the permission architecture. The account illustrates a working arrangement; it does not establish that any particular access configuration has been independently shown to be secure.

River has a name, a profile image, memories, and a personality. It can make an occasional joke and push back on an unreasonable request. The CEO is eligible for that pushback too.

Lütke relates this design to familiar metaphors in computing. Files, folders, and desktops let people apply existing knowledge to software. A colleague who can discuss a problem and accept a task offers another understandable model. He credits the personality of Microsoft's early Bing chatbot, Sydney, as an inspiration.

Agents also figure heavily in engineering. He says that few Shopify developers he knows now write code entirely by hand, and describes people coordinating 10 to 50 agent instances. Difficult edge cases, state management, and code review still demand close human attention. Faster production has not removed the need for engineering judgment.

Visible use helps a team learn

River has a deliberate constraint: it works in open channels. Open in this context means an observable company workspace, not company information published on the internet.

Lütke wanted to recover something that happens naturally in an office. A junior colleague sitting near an experienced practitioner can watch how questions are framed, problems investigated, and decisions made. Remote work can preserve the task while hiding much of that process.

With River participating in a channel, colleagues can see how someone supplies context, delegates a task, follows up on an answer, or decides that the result still needs work. Habits spread through actual work rather than through isolated demonstrations.

The examples are ordinary: summarize a discussion, create a ticket, draw a diagram, investigate papers, or make a prototype. The agent is available where the problem emerges, so employees do not first have to leave the discussion and reconstruct its background elsewhere.

During quieter periods, River also reviews conversations, identifies mistakes, and revises skill and instruction files. The team calls this dreaming. Lütke specifies that the outputs are text files. His description concerns an evolving working memory and operating instructions; it does not establish that the underlying model retrains its weights overnight.

Visible examples and an environment that retains useful experience can help turn individual AI techniques into a shared team capability.

An AI council prepares the decision

Lütke also uses a personal AI chief of staff. For an important strategic question, he has it assemble a council of roles and models.

The process assigns 5 to 6 perspectives, including data analysis, research, business, and engineering. Different models examine the question through those roles, and their findings pass through additional synthesis. The model mix changes as capabilities change. The final briefing can arrive as audio for him to hear during a morning workout.

He estimates that a demanding run costs about $15 to $20 in model usage and returns in roughly 30 minutes. Preparing comparable material himself might take a month. Those are estimates for his personal workflow, not a general performance guarantee.

The purpose is to improve the preparation that precedes judgment. Where did an argument begin? Does the evidence support the conclusion? Which perspectives have been left out? These questions become easier to investigate.

The choice remains human. Lütke treats accountability as central to management: a machine can supply information, suggest a course of action, and perform tasks, but it cannot assume the decision maker's responsibility for the outcome.

That also defines the limits of the council. Agreement among several answers does not settle whether their evidence is sound, whether their assumptions fit the business, or whether failure would be acceptable. More answers do not complete those judgments on their own.

Parrish presses the argument further, asking how people could expect to control intelligence greater than their own. Lütke compares it with a city such as Toronto, where specialization, rules, and checks coordinate abilities far beyond those of any individual. He takes an optimistic view of applying a similar institutional approach to AI. The comparison explains his position; it does not demonstrate that artificial superintelligence will necessarily be controllable.

Cheap generation can create expensive review

Asked what AI has made worse, Lütke describes what colleagues call slop grenades: packages of low-quality output tossed to somebody else.

An employee asks an agent to make extensive code changes, gives the result a cursory look, and submits it. Understanding the change, checking it, and repairing mistakes become the reviewer's problem. The submitter appears fast, while the team's total effort may not have fallen.

Email can follow the same pattern. A sender expands a simple point into a lengthy message with AI. The recipient then uses AI to compress it again. The extra words create a round trip without necessarily improving communication.

Generation has become cheaper, while understanding and review still consume time. Undigested output transfers that cost to someone else.

Lütke prefers using AI to make a point concise and clear. Whoever submits a document or a code change should still understand it, explain its purpose, and own its quality. Output volume cannot discharge that responsibility.

Software can adapt around its user

Lütke's expectations for software are shaped by his own computer. He uses Omarchy, the Linux system created by David Heinemeier Hansson, and treats an agent as a way to reshape it. He describes the change he wants instead of making every adjustment himself.

During a design discussion, he needed a way to annotate screenshots. He gave an agent a spoken request, supplied references, and followed up with 3 steering messages. According to his account, he released the tool that evening and found 6 community pull requests the following morning. He also said it was going into a subsequent Omarchy release.

The significance of the example is the ability to express a preference, try the result, and keep adjusting the tool around a particular way of working. His claim that it was the best tool was explicitly about his own preferences.

He also describes his home AI assistant using Wake-on-LAN to start a machine so that a task could continue, and dealing with a network failure. These are personal reports of agents working in a concrete operating environment.

That experience informs his direction for Shopify: a merchant describes how a business works, and the software adapts around it. He sees malleable, collaborative software as the future. This is a product vision, not an announcement that every capability implied by it is already generally available.

Choosing among good options is the hard part

When prototypes, research, and comparisons become easier to produce, a decision maker faces more plausible options. Lütke expects taste and judgment to become more valuable under those conditions.

He describes taste as something developed through repeated practice and the study of excellent work and durable institutions. Why does a design succeed? Why does a system endure? How do human habits and limitations shape the outcome? Understanding these relationships creates a better basis for making tradeoffs.

In the discussion of intuition and feedback, Parrish invokes Kahneman's emphasis on repetition, a stable environment, and rapid feedback. Lütke responds that consequential business choices often lack quick feedback. Rebuilding a product, entering a market, or declining expansion to improve the existing business may take a long time to evaluate.

The exchange involves related but distinct questions: how reliable judgment develops, and how decisions can be made when feedback arrives slowly. Lütke accepts that a review must eventually establish whether the judgment was right. He does not argue that feedback is unnecessary forever.

Quarterly incentives complicate the choice. Several options may be sound, yet the one that produces visible progress this quarter is easiest to defend. A rebuilding effort with greater long-term value may never begin.

He illustrates the tension with his original snowboard business. It was profitable, so continuing to operate it was reasonable. Turning toward an ecommerce platform opened another possibility. The difficulty was comparing the long-term consequences of viable choices.

Fast feedback is a property of an option, not proof that it is the best option to pursue. Lütke also acknowledges exceptions to his suspicion of conventional answers. In regulated activities such as payments, the orthodox approach may be correct or required.

A metric can pull a company away from its purpose

Lütke returns repeatedly to Goodhart's law: once a measure becomes the objective, it can stop being a useful indicator of the underlying goal.

His example is merchant churn at Shopify. A closed customer account is usually treated as bad news for a software company. Shopify, however, participates in an entrepreneurial process that includes many experiments.

A merchant may try a product, find insufficient demand, and close the store. That experiment has ended. The entrepreneur's journey may only have paused, and a later attempt could bring the same person back to Shopify.

If a team focuses only on preventing account closures, it can treat ordinary entrepreneurial experimentation as something to eliminate. Improving the number would then be a poor substitute for helping more people build businesses.

This does not make all churn beneficial. An unusable product, poor service, and a business experiment that did not find demand have different causes. A metric creates a reason to investigate; it cannot replace understanding the circumstances behind it.

Progress sometimes requires removing what exists

Lütke extends craftsmanship to things users may never see: architecture, readable code, and the ease with which colleagues can understand a system. He values Ruby's expressiveness because well-written code can explain a system to people while also being executable by a machine.

Carefully made work can still need rebuilding. When discussing SpaceX's Raptor engines and organizational renewal, he focuses on subtraction across successive iterations. A structure that was necessary under earlier constraints can be reconsidered when those constraints change.

Companies accumulate structures in much the same way. A problem produces a new process; another produces an additional management layer. Eventually, the original purpose is buried beneath arrangements the organization continues to maintain.

He calls a deliberate return to the purpose a refounding event: an opportunity to rebuild a department, a product, or a company around what now makes sense. The fact that an existing solution works does not establish that adding to it indefinitely is the best approach.

As AI makes adding features easier, organizations need to become more deliberate about ending things. Stopping a failed experiment can also free a creative person to pursue something more useful.

The argument has a boundary. Lütke distinguishes losses involving replaceable resources from catastrophic failure, specifically mentioning crewed spaceflight. Tolerance for a failed software experiment cannot simply be transferred to an irreversible, high-consequence setting.

The work of developing judgment remains

Later in the conversation, Lütke describes a family habit. When a child says something cannot be done, the family adds yet. He wants ability to be understood as a state that can change.

He has also used affirmations to change his own behavior. When public speaking frightened him, he spent about a week writing, for 5 minutes a day, that he enjoyed speaking about subjects that interested him. He believes the exercise helped, while acknowledging uncertainty about whether it caused the change. It is a personal experience, not evidence of a universally effective training method.

Among books with a lasting influence on him, he names Parkinson's Law, Will and Ariel Durant's The Lessons of History, James Burnham's The Managerial Revolution and The Machiavellians, and Meditations. His fiction examples include Foundation and The Three-Body Problem.

AI provides additional ways to acquire knowledge, make tools, and try ideas. Works that have survived the passage of time still help him think about people, institutions, and choices.

His definition of success is similarly concrete: become capable of more things and create products, toys, or other work that improves someone's day, or gives someone more ability and willingness to act.

That supplies a measure for the enthusiasm running through the interview. The amount of content a system can generate describes a tool's output. What matters to an organization is the useful work that remains, the choices it makes, and whether someone takes responsibility for those choices.