Shopify isn’t just using AI. It’s rebuilding its entire engineering operation around it. In a new interview on the Latent Space podcast, Shopify CTO Mikhail Parakhin laid out exactly what happens when a $200 billion company goes all-in on AI internally. The punchline? Writing code is the easy part now. The real problems are everything that comes after: reviewing it, testing it, and deploying it without breaking things.
If you run any kind of service business and you’re experimenting with AI-assisted development or AI workflows, Shopify’s experience is a preview of what’s coming for you. The bottlenecks shift fast once AI starts writing code at machine speed.
What happened
- Shopify CTO Mikhail Parakhin (formerly CEO of Microsoft’s Windows/Edge/Bing business unit) went deep on the company’s internal AI adoption in a Latent Space podcast interview.
- AI tool usage inside Shopify has hit close to 100% of daily active employees. It’s now hard to do your job without interacting deeply with at least one AI tool.
- A December 2024 phase transition in model quality caused adoption to explode. Small improvements accumulated into a massive inflection point.
- Shopify funds unlimited tokens for every employee and sets a model quality floor (nothing less than Opus 4.6).
- The company built three major internal AI systems: Tangle (reproducible ML workflows), Tangent (automated research and optimization loops), and SimGym (customer behavior simulation using real historical data).
- CLI-based AI tools (Claude Code, Codex, internal agents) are growing faster than IDE-based tools like Cursor or GitHub Copilot.
The numbers
- Shopify is a $200 billion company that’s been around for 20 years.
- Internal AI tool usage is approaching 100% daily active across all employees.
- The company enforces a minimum model quality of Opus 4.6 for all staff. Some use GPT 5.4 Extra High.
- Employees get unlimited token budgets, with no cap on AI usage.
- The December inflection point triggered a visible hockey-stick growth curve in internal adoption charts.
5 lessons from Shopify’s AI infrastructure that apply to every service business
- Generation is solved. Review is the new bottleneck. AI can write code faster than any human. But Shopify found that PR volume, test failures, and deployment rollbacks are now the constraints. Their CTO says most off-the-shelf code review tools miss the point, which is why they built their own. If your team uses AI to produce work, you need equally strong systems to check that work.
- More AI code can mean more bugs. Even though models write cleaner code on average than humans, AI-written code can still increase bugs in production. Speed without quality control is just faster failure. You need critique loops, not just generation loops.
- Reproducibility matters more than you think. Tangle, Shopify’s internal workflow engine, makes ML and data workflows reproducible, collaborative, and production-ready from the start. It uses content-addressed caching that creates network effects across teams. Translation: when one team solves a problem, every other team benefits automatically. That’s the kind of infrastructure that compounds.
- Experimentation should be for everyone, not just engineers. Tangent, their auto-research tool, lets PMs and domain experts run optimization experiments without being ML engineers. It handles search tuning, prompt compression, storage optimization, and more. The takeaway: if only your technical staff can experiment with AI, you’re leaving value on the table.
- Simulation beats guessing. SimGym simulates customer behavior using real historical data. It evolved from comparing A/B test variants to telling merchants exactly what to change on a live storefront to raise conversions. The key insight: simulated customers only work if you have real behavioral data to train them on. Your existing customer data is a moat. Use it.
The hot take
Jensen Huang was right about token budgets being directionally important, but Shopify’s experience proves that raw token count is a terrible way to measure engineering output. Spending more tokens on review and critique is worth ten times more than spending them on generation. The companies that win in 2026 won’t be the ones generating the most AI code. They’ll be the ones with the best systems for catching bad AI code before it ships. Git, pull requests, and CI/CD pipelines were designed for human-speed development. They’re cracking under machine-speed output. The entire deployment stack needs rethinking, and most teams haven’t even started.
The Agency OS play
You don’t need to be a $200 billion company to steal from this playbook. Start with the bottleneck Shopify identified first: code review and deployment stability. If your team is using AI to write code, build automations, or generate content, audit how much time you spend reviewing that output versus producing it. If the ratio is less than 1:1, you’re probably shipping mistakes. Set up a dedicated review step for every AI-generated deliverable, whether that’s code, copy, or client work.
Next, look at reproducibility. Whatever AI workflows your team runs today, document them well enough that someone else could rerun them and get the same result. That means saving prompts, model versions, and input data together. It sounds boring. It will save you hundreds of hours when something breaks at 2 AM and nobody remembers how the workflow was set up. Tools like DVC, MLflow, or even a well-organized SQLite database can get you 80% of what Tangle does.
Finally, think about who on your team gets to experiment. If AI experimentation is locked behind your most technical people, open it up. Give your project managers, your account leads, your domain experts access to AI tools and a sandbox to test ideas. Shopify found that democratizing experimentation (through Tangent) produced insights that pure engineering teams missed. You probably have people sitting on great ideas who just need permission and a low-risk way to try them. Give them that this week.