Chapter 8
These Concepts in the Real World
Everything covered in Chapters 2 through 4 (a conversation with a model, retrieval grounding an answer in real documents, an agent deciding which tool to reach for) isn't just a teaching device. Each one is a pattern that large organizations have already put into production, at a scale of millions of users and, in some cases, hundreds of millions of dollars in measured impact. This chapter walks through five real examples, one per concept, with the actual companies and the actual numbers they've published.
8.1 Conversational AI
The plain chat interface from Chapter 2 is the most direct route into a business: replacing or augmenting a support queue that would otherwise require a person for every single conversation. The appeal is straightforward: a model can hold thousands of conversations at once, doesn't get tired and can pull up a customer's history instantly. The harder lesson enterprises have learned is that this doesn't mean the human disappears entirely; it means the human's time gets reserved for the cases that actually need judgment.
The Swedish payments company Klarna launched an AI customer service assistant in February 2024 built on OpenAI's models. In its first month it handled 2.3 million conversations (work the company estimated was equivalent to about 700 full-time agents), resolving issues in under two minutes on average, versus eleven minutes for a human agent, and cutting repeat inquiries by a quarter. By 2025 it was handling roughly two-thirds of all customer chats.
The instructive part came next: in May 2025, Klarna's CEO publicly said the company had cut human support too far, and began rehiring agents after customers complained about generic answers on complicated, nuanced cases. The company's own conclusion was that full replacement was the wrong goal: the AI is best at high-volume, well-defined questions where it already has a customer's purchase history in context, while a human is still better for anything that needs real judgment.
8.2 Retrieval-Augmented Generation
RAG's enterprise appeal is almost entirely about trust and access control. A general-purpose model has no way to know a specific company's internal documents, and even if it could be retrained on them, that answer would become stale the moment those documents changed. RAG sidesteps both problems: the model reads current, real documents at the moment it answers, and, because retrieval is a separate step, an organization can restrict exactly which documents a given search is even allowed to touch.
Morgan Stanley Wealth Management built an internal assistant, "AI @ Morgan Stanley Assistant," on top of GPT-4, configured to generate answers exclusively from the firm's own library of roughly 100,000 research reports and internal documents, the same grounding-in-retrieved-passages pattern from Chapter 3, just at institutional scale. The firm reported that the share of that library advisors could effectively search and use jumped from about 20% to 80%, and that queries that used to take an advisor over thirty minutes to research now take seconds. As of the tool's most recent reporting, over 98% of advisor teams were actively using it.
8.3 Vector Databases at Scale
The demo in Chapter 3 searches a few dozen chunks of text: small enough that almost any approach would be fast. The reason production vector databases exist at all is what happens when that number grows past a company's entire catalog, all its users' behavior, or every document it has ever produced: exact comparison against every stored vector stops being fast enough, and the approximate, index-based search described in Chapter 3 becomes a hard requirement rather than an optimization.
In 2023, Spotify's engineering team open-sourced Voyager, an HNSW-based nearest-neighbor search library (the same indexing approach from Chapter 3) built to replace their previous system for powering recommendation features like Discover Weekly. Spotify reported it runs roughly ten times faster while using about a quarter of the memory of what it replaced, and it's queried hundreds of millions of times a day in production. The exact trade-off Chapter 3 describes an approximate index making (trading a small amount of accuracy for a large speed gain) is the whole reason the system exists.
8.4 AI Agents and Autonomous Tool Use
Agents earn their keep wherever a task is better described as "go accomplish this" than "answer this one question", most visibly, so far, in software engineering itself, where an agent can read a codebase, make a change, run the result and iterate, the same think-act-observe loop from Chapter 4 applied to real production code.
GitHub Copilot, an AI coding assistant, reached roughly 20 million users by mid-2025 and is used at an estimated 90% of Fortune 100 companies; developers using it have reported coding meaningfully faster, with a large majority keeping its suggested code rather than discarding it.
A step further up the autonomy ladder, Cognition's agent Devin doesn't just suggest code: it plans and executes entire coding tasks on its own. In July 2025, Goldman Sachs announced it was piloting Devin alongside its roughly 12,000 human developers, describing the goal as a "hybrid workforce" aiming for the equivalent output of about 14,400 developers from the same headcount.
8.5 The Model Context Protocol
MCP is the youngest of these five patterns (Anthropic introduced it in November 2024) but it solves an old, unglamorous problem: without a shared standard, every company connecting an AI system to its internal tools has to build and maintain its own one-off integration for each pairing. A shared protocol turns that many-to-many problem into a one-time cost per tool, reusable by any compliant AI application.
Block (formerly Square) and Apollo GraphQL were among the first companies to partner with Anthropic in shaping MCP and piloting it internally. Adoption then moved quickly across the industry: one early-2025 industry survey found the share of Fortune 500 companies running MCP servers in production roughly doubled in a single quarter. In December 2025, Anthropic handed governance of the protocol to the Linux Foundation's Agentic AI Foundation, backed by AWS, Google, Microsoft, OpenAI, Bloomberg and Cloudflare, the clearest sign yet that MCP had moved from one company's idea to shared infrastructure, the same trajectory USB took from one company's connector to a universal standard.
Five patterns, five real deployments, one throughline: none of this is speculative technology. Every mechanism in Chapters 2 through 4 (the conversation, the retrieval, the agent loop, the shared tool protocol) is already running in production, at a scale most software never reaches, inside companies you almost certainly already know.
Sources & Further Reading
- Klarna, "Klarna AI assistant handles two-thirds of customer service chats in its first month" (official press release)
- Morgan Stanley, "Key Milestone in Innovation Journey with OpenAI" (official press release)
- Spotify Engineering, "Introducing Voyager: Spotify's New Nearest-Neighbor Search Library" (2023)
- TechCrunch, "GitHub Copilot crosses 20M all-time users" (July 2025)
- CNBC, "Goldman Sachs autonomous coder pilot marks major AI milestone" (July 2025)
- Anthropic, "Donating the Model Context Protocol and establishing the Agentic AI Foundation" (December 2025)