Shopify has gone all-in on AI, and its head of engineering, Farhan Thawar, explains how the company is reshaping its engineering culture, tools, and hiring around AI-first workflows.
Shopify was the first company outside GitHub to adopt GitHub Copilot in 2021, a year before ChatGPT launched, after Farhan personally emailed GitHub’s new CEO to demand access.
The company now uses multiple AI coding tools including Cursor, VS Code with Copilot, Claude Code, and previously experimented with Devon, reflecting a shift from its usual “one tool” philosophy to exploring many options during a period of rapid AI proliferation.
Non-engineering teams like sales, finance, and support are among the fastest-growing users of Cursor at Shopify, building MCP servers to connect services like Salesforce, Google Calendar, and Gmail to create personalized dashboards without engineering help.
Shopify built an internal LLM proxy on top of the open-source LibreChat to let all employees use AI models safely without leaking private data, with token tracking, cost visibility, and a leaderboard that celebrates high-usage teams doing meaningful work.
The company has no cost limit on AI token spending, and Farhan argues that $1,000/month per engineer is “too cheap” if it delivers even a 10% productivity gain, pushing back against leaders who see AI tool costs as a line item to cut.
Shopify is hiring 1,000 interns per year (about 350 per term) specifically because they expect this generation to be more AI-reflexive, and the company views internships as a way to learn from younger employees rather than as charity.
Engineering directors and above now go through coding interviews where they pair with Farhan using AI tools like Copilot, testing whether they can evaluate AI-generated code critically rather than just accepting it.
Farhan’s core advice for other companies wanting to become AI-first is role modeling: leaders must use AI tools visibly, share prompts and workflows openly, and create internal prompt libraries so others can learn from real examples.
How Shopify works with AI labs
Farhan pairs directly with engineers at companies like Anthropic to understand how they use their own tools internally.
He spent an hour pairing with an Anthropic applied AI engineer on Claude Code, building things together and comparing how each company uses the tool.
These sessions happen through shared Slack channels and are true collaborations, not vendor-customer negotiations.
The goal is to learn quickly and bring insights back to Shopify’s own teams.
The recent Code Red at Shopify
From November 2024 through mid-2025, Shopify ran a seven-month “code red” to address growing technical debt that was producing dangerous signals like seg faults and rising exception counts.
Between 30–50% of engineering was redirected to reliability work with no fixed headcount target, just clear metrics: unique exception counts had to shrink and seg faults had to reach zero.
The effort ended when all reliability surfaces (storefronts, merchant admin, point of sale) showed green on 28-day rolling averages.
Shopify’s deep stack involvement—they are core contributors to Ruby and operate one of the world’s largest MySQL fleets—means they see low-level failures that other companies might miss.
Shopify’s internal AI infrastructure
The LLM proxy gives every employee access to enterprise-grade AI models without risking data leaks.
It sits on top of LibreChat, which Shopify contributes to as an open-source project.
Employees request tokens, and the system tracks usage by team and person, enabling cost visibility and identification of high-impact users.
Shopify has about two dozen MCP servers and growing, putting them in front of every internal system.
Their internal wiki called “The Vault” is accessible via MCP, letting employees query project histories, internal talks, and board letters through a chat interface that works like an internal Perplexity.
New systems get an MCP endpoint as a matter of practice, making data immediately accessible to agents and chat tools.
How AI is changing engineering work at Shopify
The best engineers are using AI tools to tackle tech debt reduction, refactoring, and code readability improvements they never had time for before.
Some run parallel AI agents on large refactoring tasks, reviewing the resulting PRs as if they had a temporary team.
Others experiment boldly, letting tools like Devon run for 24 hours and deleting the results if they fail, treating it as a worthwhile experiment.
Shopify built its own project management tool called GSD (Get Sh*t Done) instead of using Jira or Linear.
Every project has weekly updates, and an AI tool now drafts those updates by pulling from PRs and Slack conversations, though engineers are expected to review and take responsibility for what gets published.
Every six weeks, leadership reviews every project in the company for direction, resourcing, and whether it should continue.
Hiring and culture changes
All entry-level engineering hires at Shopify now come exclusively through the intern program.
Interns are brought into the office as a cohort despite Shopify being a remote company, because younger employees benefit from working alongside peers.
The company pays interns well, including housing and travel costs, after seeing candidates reject low offers elsewhere.
For director-level and above hires, Shopify added a coding interview where candidates pair with Farhan using AI tools.
Candidates who don’t use Copilot or similar tools tend to be outperformed by those who do.
Farhan evaluates whether candidates can critically assess AI-generated code, spot errors, and make manual fixes rather than endlessly prompting.
The goal is to confirm that senior leaders still love technology and aren’t running away from code.
AI’s impact on SaaS and software demand
Farhan is not worried about AI killing SaaS, arguing that the world needs 10,000 to 100,000 times more software than exists today.
He sees parallels with how smartphones created YouTube and TikTok alongside a thriving professional film industry, invoking Jevons’ paradox: the more accessible software creation becomes, the more demand there will be.
He believes AI tools may benefit expert engineers more than mediocre ones, similar to how iPhone cameras benefited skilled filmmakers most.
Today’s vibe coding works for personal tools and prototypes, but architecturally sound production systems still require deep understanding.
Farhan’s personal AI workflow and tool preferences
Farhan uses a range of tools depending on the task, emphasizing that different tools operate at different layers of the stack.
Claude Code for coding tasks, Cursor for personal workflow automation, and Gumloop for browser-based automation and web scraping.
He cites an example where Claude Code took an hour and failed to check if a SanDisk 8TB drive was in stock, while Gumloop solved it in 2 minutes because it already knows how to scrape websites.
He encourages engineers to move away from default small models in Cursor and use more capable models like o1 Pro or o3 Pro, even at higher cost.
Shopify’s CTO Miki personally advocates paying $200/month for top-tier models like Claude Advanced or Gemini Ultra, arguing that sticking to $10/month plans means missing the real progress in LLMs.
Advice for companies wanting to become AI-first
Role modeling is the single most effective driver of AI adoption.
Leaders must use AI tools visibly, share their workflows, post in internal channels about what worked and what failed, and maintain prompt libraries others can reuse.
Shopify ran a hackathon at its recent Summit focused heavily on AI, with senior engineers trying Claude Code for the first time and discovering they could be 95% AI-driven while still making manual adjustments.
No company has AI adoption figured out yet, so sharing experiments openly—including failures—is essential.
