How Production Software Uses AI: Real Implementations from Stripe, GitHub, and Notion
AI in Production Looks Different from AI in Demos
Most of the AI content on the internet is about demos. Someone gets ChatGPT to write a poem. Someone builds an agent that books a flight. The real story — the one that matters for builders — is what production teams have actually shipped to paying users at scale. The pattern is consistent: AI is rarely the headline feature. It is plumbing that quietly makes existing features better, often without the user even noticing.
We have studied a lot of production deployments. The companies that get AI right do not treat it as a separate product. They treat it as an ingredient. The user comes for the original value proposition. The AI is what makes the experience feel ten times better than the same product without it.
GitHub Copilot — The Inline Assistant Pattern
Copilot is the most-cited AI integration in production for a reason. It pioneered the inline assistant pattern — AI that appears exactly at the moment of user intent, suggests the most likely next action, and gets out of the way if the user keeps typing. The genius is the lack of friction. You don't ask Copilot to do anything. It is just there, predicting, occasionally correct, never demanding attention.
The architecture lessons are worth studying. Copilot batches requests aggressively, caches recent suggestions, and runs a smaller routing model that decides whether the larger generation model should be called at all. This is why latency feels instant even though the underlying model is doing nontrivial work. Most builders trying to copy this pattern make the mistake of calling the model on every keystroke. That is expensive and pointless. The lesson: the smartest part of the integration is knowing when not to call the model.
Stripe — AI for Fraud, Reconciliation, and Developer Docs
Stripe's AI integration is mostly invisible to the end user, which is exactly why it works. Their fraud detection layer, Radar, has used machine learning models for years, but the post-LLM era saw them add language-model-powered analysis on top — explaining why a transaction was flagged in plain English to merchants, summarizing dispute evidence, and helping merchants write better dispute responses.
The other place Stripe leans on LLMs is documentation. Stripe's docs site has an AI assistant that does not just retrieve relevant pages — it composes complete code samples in your specific language and integration context, references your actual API key environment, and reasons about which integration path is right for your use case. The lesson: AI in documentation is one of the highest-ROI integrations a SaaS product can build. It deflects support load while making the docs feel personal.
Notion AI — Inside the Document, Not Beside It
Notion's AI integration succeeded because it lives inside the document. Selecting text and pressing space brings up a context-aware AI that operates on what you have selected, with the surrounding document as context. There is no separate chat panel. There is no copy-paste back and forth. The AI is just another way to interact with your content.
The technical implementation is more interesting than it looks. Notion uses retrieval-augmented generation across the entire workspace — the AI can reference other pages, databases, and historical edits without the user explicitly attaching them. The relevance ranking is the secret. A naive implementation would dump every page into the prompt. Notion's system uses embedding-based retrieval to fetch only the most relevant context, keeping latency reasonable and prompt tokens affordable.
Linear — AI as Triage, Not Author
Linear's AI features are restrained, which is exactly why they work. The AI does not write your tickets for you. It triages — auto-tagging, suggesting linked issues, summarizing long discussions, drafting status updates from completed work. Each individual feature is small. Together they reduce the manual overhead of project management by maybe 30 percent. That is the real win.
The architectural pattern Linear uses is worth name-checking: lightweight, narrow tools rather than a single big chatbot. Each tool has a tightly scoped system prompt, a defined output schema, and runs against a specific UI affordance. There is no general-purpose Linear AI. There are 15 small ones. This is the future of AI in B2B software. Forget the chatbot. Build a hundred small AI-powered actions that each save the user 30 seconds and never break.
What Production Teams Actually Get Right
Across all the production deployments we have studied, four patterns separate teams that ship from teams that struggle. First, they treat the model as unreliable. Every output goes through validation, schema-checking, and fallback handling. They never assume the model will return what they asked for, even when it does 99 percent of the time.
Second, they obsess about latency. Streaming is non-negotiable for user-facing features. Caching common requests is mandatory. Routing simple requests to smaller models saves both money and time. Third, they treat the model as a feature, not a product — AI is a way to make existing flows better, never the headline. And fourth, they build evaluation suites. Every prompt change runs against a benchmark. Every model upgrade runs the same. They know exactly what changed and why before shipping. This is the discipline that separates AI features that actually work from the demos that go viral and then quietly get rolled back.
Ready to put this into action?
We build the digital infrastructure that turns strategy into revenue. Let's talk about what DRTYLABS can do for your business.
Get in Touch