Author: Lincoln Wang | Founder & CEO, MindsLeap | Partner and CEO, Founders Space China | Founder, MindsLeap Founders AI Club
This article is Lincoln's interpretation of Inside Shopify's AI strategy with CTO Mikhail Parakhin, published by the official Claude channel on October 8, 2026.
A problem that had resisted seven years of work finally became tractable. In his conversation with Boris Cherny, Shopify CTO Mikhail Parakhin described using a model to tackle a problem involving a Wasserstein loss function. He did not provide enough detail for outsiders to reproduce the work. But the experience supported a question worth considering: some tasks may remain out of reach even when a company adds people and time.
Parakhin described a relationship of mutual dependence: he needed the model, and the model still needed him.
The conversation explored how Shopify uses AI. Its most useful thread was how Parakhin connects model capability, engineering management, and business growth. It moves the familiar question of how much money AI saves toward a harder one: what new capabilities does it give us?
A Different Measure of Value
Parakhin distinguishes between raising the floor and raising the ceiling. When a model helps engineers complete work they already know how to do, the team can ship faster. That raises the floor. When it helps them do something previously impossible, or something that would never have made the project list, it raises the ceiling.
The first benefit is comparatively easy to estimate: fewer people, fewer weeks, less effort. The second is harder to put in the same spreadsheet. The comparison is not with a less efficient project. It is with a project that never happened.
In his view, raising the ceiling has the larger impact. That is his judgment, not a financial conclusion that automatically applies to every business. Still, a company that measures only faster coding and shorter report-writing time may miss part of AI's value. Options that were too expensive, too slow, or technically inaccessible may become feasible.
That is why the seven-year problem matters. It gives capability expansion a concrete reference point, beyond the abstract observation that models keep improving.
Test a Company's Digital Twin Before Changing the Business
Parakhin offered a more operational example: representing merchant behavior as a sequence of actions. Opening a credit line, shipping an order, accepting credit cards, applying for a loan, and changing business practices can all form part of that sequence.
He described building digital twins of businesses and exploring interventions within the model. What might happen after more advertising, shipping a day earlier, or providing a loan? After identifying potentially useful actions, the team can offer financing or recommendations in the real business.
Here, a digital twin is a model used for prediction and simulation. Its value is to produce more specific hypotheses that can then be tested in practice. There is an important limit: the interview did not disclose independent evaluations, prediction errors, or methods for identifying causality. A simulated improvement does not prove that the real business will improve.
Even with that caveat, the direction is useful. Ecommerce businesses hold substantial order, fulfillment, and customer data. The difficult step is turning that information into the next action. If models help operators compare alternatives, AI begins to move from producing content into supporting business decisions. Accumulated data acquires another use.
An Open Token Budget Has a Second Half
Shopify gives employees room to use models extensively. Parakhin said the company does not impose individual token limits and is willing to invest in stronger models during development. His reasoning is that without trying a more capable model, people cannot know what possibilities they are missing.
That can sound like a recommendation to buy the most expensive model. His full argument is more specific: use capable models for research, testing, and exploration; once a task enters production, select an appropriate model and optimize it, including fine-tuning where needed, to balance quality, throughput, and cost. He described this as a barbell strategy.
During exploration, neither the problem nor the solution is fully settled. More capability may reveal another approach. In production, the task is usually clearer, while the number of executions can be large. Unit cost therefore matters much more. The two budgets serve different purposes.
He also emphasized circuit breakers to prevent runaway processes from consuming tokens indefinitely. Unrestricted use still needs engineering controls and monitoring. For enterprise leaders, the transferable lesson is to distinguish spending that discovers new capabilities from spending on repeated execution. That makes it easier to decide where to invest and where to optimize.
Engineers Are Managing Another Layer of Execution
Parakhin also described changing his mind about management. He initially doubted that large language models would help him much in managing teams. Later, he built a system to analyze project progress and try to detect schedule risks and team issues earlier.
These are his reported experiences. They do not establish that a system can reliably assess every employee's emotions. But they point to a change in engineering management: leaders manage engineers, while engineers increasingly supervise models and AI agents.
How should tasks be divided? Could agents' work conflict? Who checks the results? When a model can generate an implementation quickly, technical leaders need to pay even closer attention to whether the problem has been defined correctly. Parakhin warned that models readily follow the questioner's intent. A flawed requirement can therefore become a flawed implementation faster.
An ambiguous idea previously had to pass through discussion and coordination before development. Faster implementation means organizations must deliberately preserve judgment and review. As execution capacity expands, deciding what to do and how to verify it deserves more attention, not less.
Reopen the Project List
For me, the most useful implication is a reason to revisit projects that have sat on the shelf. Which ideas were excluded because talent was scarce, development was expensive, or analytical capability was insufficient? Is there now a small, clearly defined part that the team could try again?
This does not mean every company should copy Shopify's digital twins. Different industry data, technical teams, and operating problems require different starting points. But when evaluating AI, we can ask two questions at once: how much have existing workflows improved, and has the team accomplished something it could not do before?
The first helps us manage efficiency. The second helps us see new room for growth. Considering both creates a path from buying AI tools to rebuilding organizational capability.
Based on the Claude channel's Shopify CTO interview and accompanying transcript. Descriptions of digital twins and management systems reflect the interviewee's account, not independent validation. Quotations in the Chinese version are translations; this English edition paraphrases those remarks.
About MindsLeap
MindsLeap is an enterprise AI transformation and AI-native startup acceleration platform. It supports traditional enterprises adopting AI as well as AI-native ventures, one-person companies, and technology founders, connecting industry use cases with capital, Silicon Valley resources, and global markets.
MindsLeap is a global partner of Founders Space. Through the MindsLeap Founders AI Club, AI training and strategic advisory, Forward Deployed Engineering (FDE), and startup acceleration, it helps organizations move from understanding AI to implementing it in real business workflows.
Its ecosystem connects entrepreneurs, AI engineers, industry experts, investors, and global innovation resources to turn AI into practical organizational capabilities, product innovation, and growth.