AI Anatomy
Morgan Stanley and Klarna use the same AI models as every competitor they have. Anyone can rent the same intelligence tonight for the cost of a coffee. What separates them is what they built around it: how it remembers, how it reasons, how it decides. Most organisations bought the brain. Almost none built the body.
Nadim A. Massih9 June 2026 · 9 min read
The Body Nobody Built
Three companies from completely different industries are outperforming their competitors on AI. Not because they found a smarter brain. Every model they use can be rented tonight for less than the cost of a coffee.
Morgan Stanley (the global investment bank) uses the same frontier models as any competing bank. Klarna (the Swedish payments company) handled 2.3 million customer service conversations through AI in a single month, cut resolution time by 82 per cent, and added $40 million in annual profit. Cursor, the AI coding environment used by thousands of professional software teams, now merges more than 30 per cent of its pull requests (the formal code submissions a developer packages for review) from autonomous agents. The brain in each case is available to anyone with a credit card. What is not: the body they built around it.
This is the story of that body: every part of it, and how to build yours.
The reasoning models available today are extraordinary by any prior standard. Claude Opus 4.8 (Anthropic’s current flagship; Anthropic is the AI safety company behind Claude) solves professional software engineering tasks at the level of a senior developer. GPT-5 (OpenAI’s flagship model) scores 94.6 per cent on AIME 2025 (the American Invitational Mathematics Examination, the world’s hardest pre-university mathematics competition). Gemini 3.1 Pro (Google’s current flagship reasoning model) reads a thousand-page document in the time it takes you to read this paragraph. By April 2026, the cheapest frontier-class API (DeepSeek) processed a million tokens, roughly 750,000 words, for $0.0035. Frontier-level intelligence is no longer an expensive private service. They are APIs (Application Programming Interfaces, the standardised gateways through which software systems communicate), available to any organisation on earth.
The brain has never been more capable. The brain has never been cheaper.
So why did Gartner find that more than half of all generative AI proofs of concept were abandoned by end of 2025? Why does McKinsey describe two-thirds of organisations as stuck in “pilot purgatory”: running experiments that never become products?
The brain is not the problem. Most companies never built anything around it: no memory, no nervous system, no hands, no ability to take action.
They built a brain in a jar.
The six components of a complete AI body. Most organisations have only the Brain. The winners connected all six. (Patient Comet analysis, 2026)
The Brain: What Happens the Moment You Hit Send
Most people experience AI as a conversation. The conversation format makes it feel like a person. It is not a person.
Your words get broken into tokens (the unit of text the model processes, roughly three-quarters of a word each) and converted into numbers. Those numbers go to the LLM (large language model, the AI’s brain). The model has one job: calculate which number comes next. It picks one, converts it back to a word, and repeats thousands of times. That is inference: one processing cycle per token.
There is no understanding. There is sophisticated mathematics, running very fast. Knowing this matters for one reason: it explains what the model can and cannot do, and why each body part that follows solves a specific limitation of the brain.
One inference cycle produces exactly one token. A 500-word response requires approximately 670 cycles. The output from each cycle feeds back into the next input. (Patient Comet analysis, 2026)
The Lungs: How Much the Brain Can Hold at Once
The context window (working memory, the amount the model holds at once during a single conversation) determines how much of the world the AI can see at one time. The previous generation of models held roughly 128,000 tokens, about the length of a standard novel.
Claude Opus 4.8 solves 88.6 per cent of professional software engineering tasks in independent testing and holds one million tokens simultaneously. Gemini 3.1 Pro scores 95 per cent on competition mathematics benchmarks and holds the same. One million tokens is roughly eight novels. A year’s worth of company emails. An entire medium-sized product codebase.
A larger context window means the brain sees more. It holds the full conversation, the full document, the entire codebase. Context window is not a technical footnote. It is the difference between a brain that can only see the room it is standing in and one that can see the building.
The 128K context window of previous-generation models held roughly one novel. Current models hold one million tokens, enough for an entire codebase, a year of emails, or ten books read simultaneously. (Anthropic / Google DeepMind, 2026)
The DNA: How Meaning Becomes Mathematics
Take four words: king, queen, man, woman. A five-year-old understands what they mean. Now ask the model: take the numbers for “king,” subtract the numbers for “man,” add the numbers for “woman.” The result is the numbers for “queen.” Not as wordplay. As arithmetic.
This is vectorisation and embeddings: each word in the model's vocabulary is converted into a number, then placed in a mathematical space where similar meanings sit close together. This is why an AI can retrieve “refund policy” when you ask about “how do I get my money back”. The two phrases sit close together in that space, even though they share no words.
Every word the model knows has a position in that space. Words with related meanings sit near each other. Words with opposing meanings sit far apart.
The word “doctor” sits near “nurse” and “hospital” in that space, so a query about “patient recovery” finds documents about “post-operative care” automatically. The model does not understand either phrase. It is doing geometry on numbers that represent meaning. This is also what makes the Memory work.
Words that share meaning cluster together in mathematical space. “king − man + woman = queen” works as arithmetic, not metaphor. This geometry is why AI can retrieve “refund policy” when you ask about “how do I get my money back.” (Word2Vec / embedding research)
The Memory: Why AI Forgets (and How to Give It a Past)
Every conversation starts from zero.
Ask Claude something at 9am. Open a new chat at 9:15am. It has no memory of 9am. It has no memory of any previous conversation. The context window (the only thing holding the exchange) is wiped clean when the session ends. This is not a flaw. It is a design property. And it has a known fix.
Short-term memory is whatever is still inside the active context window. Within one conversation, the model remembers everything said. That memory ends when the session closes.
Long-term memory comes from connecting the brain to a vector database (a storage system built for meaning, not keywords, so a search for “what is our refund policy” also finds documents titled “customer compensation process” and “returns guidelines”). The mechanism connecting that database to the AI’s responses is called RAG, or Retrieval-Augmented Generation: the process by which an AI searches the knowledge store before answering, rather than relying solely on what it learned during training.
Without RAG, the brain answers from training data that may be months or years old. When it does not know the answer, it invents one that sounds plausible regardless of whether the facts are there. This is what people call hallucination. RAG reduces hallucination rates from roughly 50 per cent to 13.9 per cent, according to a peer-reviewed study published in 2024.
Morgan Stanley equipped their financial advisors with an AI assistant connected to a vector database of 100,000 internal research documents. Before the system, advisors spent more than 30 minutes per search and reached roughly 20 per cent of the material. After it, queries return in seconds, advisors reach 80 per cent of relevant documents, and more than 98 per cent of advisors use the system daily.
No better brain than their competitors. Better memory.
“The brain in a jar can tell you what it knows. The brain with memory can tell you what your business knows.”
Without RAG, the LLM answers from training data, which may be months out of date and prone to hallucination. With RAG, it searches your knowledge store first, then answers from the retrieved documents. Hallucination rates drop from roughly 50 per cent to 13.9 per cent. (arXiv peer-reviewed study, 2024)
The Nervous System: Why Your AI Can’t See Your Business
A brain without connections to the outside world is still just a brain.
It can reason, write, and analyse any document or data you give it. What it cannot do, without a nervous system, is reach live systems: check a calendar in real time, run code, send a message, or update a record. Without that connection, the AI sits behind a chat interface and waits.
In November 2024, Anthropic released MCP (Model Context Protocol, an open standard now adopted across the industry for connecting AI models to external tools and systems). The analogy that stuck: USB-C. Before USB-C, every device used a different connector. Every AI integration required bespoke engineering: a custom bridge between the model and each external system. MCP standardised the connector.
SDK (Software Development Kit, the toolkit developers use to build on a standard) downloads reached 2 million per month at launch. By March 2026, that figure was 97 million. That is how quickly the industry decided the nervous system problem was solved.
Tool calling (the mechanism by which an AI instructs an external piece of software to perform an action and return the result) gives the brain hands. The AI does not browse the web; it calls a browsing tool. It does not execute code; it calls a code execution tool running in a secure, isolated environment. It does not query your database; it calls a database tool through a standardised API.
An AI connected via MCP to a company’s Slack, Notion, GitHub, and CRM can answer “what did the engineering team decide about the login system last week?” by querying all four systems simultaneously. Without MCP, it answers from training data. With MCP, it answers from your actual business.
The Muscles: What Agents Are and Why Most Companies Don’t Have Any
The companies actually winning at AI are not talking about which brain to use. They are talking about the body.
An agent (an AI system that takes a sequence of actions over time to complete a goal, rather than producing a single response to a single prompt) is the difference between a brain that knows things and a body that does things.
A chatbot answers: “Invoice 4721 is 14 days overdue.”
An agent finds the invoice, checks whether payment has cleared, emails the client if it has not, logs the outreach in the CRM, and schedules a follow-up for next Friday.
Same brain. Entirely different body.
The execution loop
Every agent runs on a continuous cycle called the ReAct loop (Reason + Act: the agent observes what is happening, reasons about what to do next, acts, observes the result, then reasons again). It is not one inference. It is a chain of inferences, each feeding the next, until the goal is reached or the agent encounters a decision it cannot make without a human.
This loop (observe, reason, act, observe) is the agent’s heartbeat. The single most important architectural idea in modern AI.
The ReAct loop is the agent’s execution engine. Each cycle produces one action. The agent runs loops until the task is complete or it needs a human decision. Chains of hundreds of loops handle complex multi-step tasks. (ReAct: Yao et al., 2022)
The four levels of a body
Agent maturity runs on a measurable spectrum. Most organisations are sitting at the bottom.
Level 0, the Chatbot: a brain with no body. Answers questions, forgets everything between sessions, takes no actions. This is where most people encounter AI.
Level 1, the Copilot: a brain with short-term memory and the ability to suggest next steps, still unable to act independently. GitHub Copilot (Microsoft’s coding assistant embedded in development environments), Microsoft 365 Copilot: assistants that stay inside the human decision loop.
Level 2, the ReAct Agent: a brain with memory, tool calling, and a running execution loop. It accepts a goal and works toward it autonomously. Claude Code (Anthropic’s agentic coding system), OpenAI Codex (a cloud-based coding agent that accepts a task, works in an isolated environment, and returns a completed pull request for review), Cursor.
Level 3, the Multi-Agent Orchestrator: one brain coordinating a team of Level 2 agents. Each specialist handles one domain. The orchestration layer (the coordination system that routes tasks, synthesises outputs, and manages failures across multiple agents working in parallel) handles the overall goal.
The jump from L1 to L2 requires memory and tool calling working together inside an execution loop. (Menlo Ventures / Databricks, 2025)
Only 16 per cent of enterprise AI deployments qualify as true agents, according to the 2025 State of Enterprise AI report from Menlo Ventures (a Silicon Valley technology investment firm). Only 19 per cent of organisations have deployed agents at all, according to Databricks (the enterprise data platform). The remaining 81 per cent have access to a frontier brain and are running it at Level 0.
The Full Body in Action: Three Workflows That Changed the Maths
The brain gets the credit. The body does the work.
The coding body
When a software team connects all parts of the body, the developer's job changes. The developer describes a feature in plain language. Claude Code or OpenAI Codex (cloud-based, powered by OpenAI’s reasoning models) spins up a pipeline: one sub-agent writes the code, one writes the tests, one checks for security issues, one prepares the pull request for human review. Codex handles longer asynchronous tasks in the background: the developer briefs it on a module refactor, it works in an isolated cloud container with the full codebase, and 20 minutes later returns a completed pull request.
At Cursor, more than 30 per cent of all merged pull requests now originate from autonomous agents. Not suggestions. Full pull requests, written and submitted by agents briefed by a developer and left to complete the work.
The brain writing that code is the same Claude Opus 4.8 available to everyone. What changed: the memory (full codebase context in the vector store), the hands (code execution, git operations, test runners), and the orchestration (the pipeline coordinating the agents through each stage).
The creative body
A video production team running the full creative body is built to compress a 25-day production cycle toward single digits; that is the design goal of the pipeline, not a measured industry average. For a mid-sized agency, time recovered at that scale is the difference between taking on more clients and hiring another person. The tools in that pipeline are specific and measurable.
Claude Sonnet 4.6 (Anthropic’s mid-tier model) or GPT-5 writes the script. Flux 1.1 Pro generates storyboard frames. Veo 3.1 (Google DeepMind’s video model) converts them to video with synchronised dialogue and ambient sound in a single pass. ElevenLabs adds narration. Suno v5 scores the music.
Each tool is a specialist. Each specialist has a brain. The pipeline connecting them is the body.
The business body
In the first month after deployment, Klarna handled 2.3 million customer service conversations through an AI system. Resolution time dropped by 82 per cent. Klarna put the annual profit improvement at $40 million.
The brain Klarna used was available to any of their competitors. The body they built was not: memory connected to customer account history, hands connected to billing and subscription systems, an execution loop capable of processing a refund rather than just telling the customer how to request one themselves.
“The brain answers. The body acts.”
Building Your Body: Four Steps Every Organisation Needs to Take
Most organisations can locate themselves on the body maturity scale in under five minutes.
If your AI use is primarily chat interfaces (staff asking questions, getting summaries, drafting emails) you are at Level 0. You have a brain. No body. If your AI cannot take any action without a human approving each individual step, you are at Level 1. Still no body.
This is where the body starts. None of this requires building from scratch. The infrastructure exists. The missing ingredient is architecture.
Give it memory. Connect a vector database to your most important documents. Pinecone (a managed vector database service) is the easiest starting point. pgvector works if you already run PostgreSQL. This turns your AI from a general reasoner into something that knows your company.
Wire the nervous system. Use MCP to connect your AI to one live system: your knowledge base, your CRM, or your project management tool. One connection first. The one where your team wastes the most time searching for information.
Give it hands. Enable tool calling so the AI can perform at least one action rather than just describing it. The first action is almost always the highest-friction manual task your team does every day on a predictable pattern.
Know your level honestly. The tools are available to anyone: Pinecone, MCP, Claude Code. What is not available to everyone is the architecture: knowing which parts go where, for which workflows, in which order. That is the work. The jump from Level 1 to Level 2 requires memory and tool calling working together inside an execution loop. The jump to Level 3 requires an orchestration architecture managing multiple agents. Both are buildable. Neither happens by accident.
None of this is simple. Memory requires integration work. MCP requires engineering. Agents require orchestration design. And a fair concession: the body costs real engineering money and months of work, and some of the 81 per cent are not asleep; they are waiting for the build to get cheaper. I think the wait costs more than the build, but it is not an irrational bet. The organisations pulling ahead hired for architecture. The 81 per cent of organisations stuck at Level 0 are not stuck because the brain is weak. They are failing because they never asked what a body would look like.
The same brain is available to both columns. The only difference is what was built around it. The brain is free. The body is the work. (Patient Comet analysis, 2026)
The Brain Is Free. The Body Is the Work.
Every organisation on earth has access to a better brain than any technology company had five years ago.
That is the headline. Here is the part that is less reported: the brain is not where the gap is. The gap is memory. The gap is the nervous system. The gap is the ability to act rather than answer.
The companies pulling ahead are not using a smarter model. They are using the same model as everyone else, and they built everything else around it. Memory that knows their company. A nervous system that sees their live data. Muscles that can act on a decision rather than describe it.
The 81 per cent running brain-only is not a model problem. It is an architecture problem. The brain is ready. The body is what is missing. And unlike the brain, the body does not arrive pre-built. That is the work.
The Level Audit
Spend 20 minutes with your team answering: does your AI take any autonomous action at all? Can it query a live system? Does it retain anything between sessions? The answers locate you on the body maturity scale. Most teams discover they are at Level 0 even if they believe they are further along.
AI leads, technology directors, product managersThe First Memory
Identify your organisation’s single most-searched internal document category: policies, procedures, client records, technical specifications. Connect a vector database to that one category. Do not try to connect everything at once. A brain that knows 20 per cent of your business well outperforms one that knows 100 per cent of it badly.
Technical leads, knowledge management, operationsThe First Action
Name the one task your team performs repeatedly on a predictable pattern that requires looking something up and updating a record. That is your first tool to wire. Not the most complex task. The most repeated one. Repetition is where the ROI accumulates.
Operations, customer success, finance teamsThe Honest Level
Tell your organisation what level you are actually at, not the level you claim in presentations. An honest Level 0 with a clear path to Level 2 is a stronger position than a claimed Level 2 that cannot demonstrate autonomous action. The companies that are building the body started by being honest about how little of it they had.
Leadership, strategy, anyone setting AI expectationsQuestions, answered first
Why do most AI projects fail despite access to powerful models?
Because most organisations built only the Brain (the large language model) and nothing else around it. The models are available to everyone. What is not available to everyone is the architecture: memory, nervous system, hands, and agents. Without those four components, even the most capable model cannot take action, access live data, or retain anything between sessions. That is the architecture gap behind the abandoned pilots and the stalled 81 per cent.
What is the difference between an AI chatbot and an AI agent?
A chatbot answers questions in a single response and forgets everything between sessions. An agent takes a sequence of actions over time to complete a goal, using memory, tool calling, and a continuous reasoning loop. A chatbot tells you an invoice is overdue. An agent finds the invoice, checks payment status, emails the client, logs the outreach, and schedules a follow-up, without a human directing each step.
What is RAG and why does it reduce hallucination?
RAG (Retrieval-Augmented Generation) connects the AI to a vector database before it generates a response. Instead of answering from training data that may be months old, the model first searches your knowledge store and grounds its answer in what it finds. Without RAG, a model generates confident-sounding answers even when it does not know the facts. With RAG, hallucination rates drop from roughly 50 per cent to around 14 per cent, according to peer-reviewed testing (arXiv, 2024).
What is MCP (Model Context Protocol)?
MCP is an open standard released by Anthropic in November 2024 for connecting AI models to external tools and systems. It works like USB-C for AI: one standardised connector that lets a model reach any compatible system (your CRM, your calendar, your project management tool, your code repository) without bespoke engineering for each integration. SDK downloads went from 2 million per month at launch to 97 million by March 2026.
What are the four levels of AI body maturity?
Level 0 is a chatbot: brain only, no memory, no actions. 81 per cent of organisations sit here. Level 1 is a copilot: brain plus short-term memory and suggestions, but human approval required for every step. Level 2 is a ReAct Agent: brain with memory, tool calling, and a running execution loop that works toward a goal autonomously. Level 3 is a multi-agent orchestrator: one brain coordinating a team of Level 2 agents, each handling a specialist domain. Only 16 per cent of enterprise deployments have reached Level 2 or above.
How do I start building an AI body for my organisation?
Four steps. First, give it memory: connect a vector database (Pinecone or pgvector) to your most-searched document category. Second, wire the nervous system: use MCP to connect your AI to one live system: your knowledge base, CRM, or project management tool. Third, give it hands: enable tool calling so it can perform at least one action autonomously. Fourth, know your level honestly and name the single highest-friction repeatable task to automate first. The infrastructure already exists. The missing ingredient is architecture.
Sources & references
DeepSeek API documentation, April 2026. api-docs.deepseek.com/news/news260424. Context: pricing benchmark used to illustrate how cheap frontier inference has become.
Gartner research, 2025. More than half of all generative AI proofs of concept were abandoned before reaching production by end of 2025. Consistent with the “pilot purgatory” finding in McKinsey’s parallel survey.
McKinsey & Company annual global AI survey, 2025. Two-thirds of organisations described as running experiments that never become products. McKinsey terminology for the failure mode of perpetual pilots with no production deployment.
arXiv peer-reviewed study, 2024. Empirical measurement of hallucination rates in LLM responses with and without Retrieval-Augmented Generation. The 50% baseline and 13.9% RAG rate are from the same controlled study population.
Anthropic, 2026 (released 28 May 2026). Independent verification via Vellum.ai: vellum.ai/blog/claude-opus-4-8-benchmarks-explained. SWE-bench Verified tests AI agents against real software engineering tasks from professional GitHub repositories.
Google DeepMind / BenchLM.ai / DataCamp, 2026. Competition mathematics benchmarks (MATH dataset and related competition tests). Current Google flagship reasoning model. benchlm.ai/models/gemini-3-1-pro
OpenAI, August 2025. openai.com/index/introducing-gpt-5/. AIME (American Invitational Mathematics Examination) 2025 result from the official OpenAI GPT-5 announcement. Independent testing of AIME for later versions (GPT-5.4/5.5) shows inconsistent results; original 94.6% figure retained from verified announcement.
Anthropic / ZBuild, 2026. Mid-tier model benchmark result. Used in the creative pipeline section as context for Sonnet’s capability tier.
Anthropic / Digital Applied, 2026. Model Context Protocol monthly SDK download figures. Launch: November 2024 (approximately 2 million monthly downloads). March 2026: 97 million monthly downloads. Growth demonstrates industry adoption speed.
OpenAI case study / Financial Times reporting, 2024–2025. AI assistant connected to 100,000 internal research documents. Metrics: 30+ minutes → seconds for document retrieval; 20% → 80% relevant document access; 98%+ daily advisor usage. Primary case study for RAG’s business application.
Klarna press release, February 2024. One-month deployment data. 2.3 million customer service conversations handled by AI. 82% reduction in average resolution time. $40 million annual profit improvement attributed to the deployment. Klarna is the Swedish buy-now-pay-later payments company.
Cursor / Anthropic, 2025–2026. More than 30% of all merged pull requests at Cursor now originate from autonomous agents. Full PRs, not suggestions. Primary evidence for Level 2+ agent deployment in production software teams.
Menlo Ventures State of Enterprise AI, 2025. Of all enterprise AI deployments surveyed, only 16% qualify as true agents (defined as AI systems that take autonomous action sequences toward a goal). Silicon Valley technology investment firm; annual enterprise AI deployment survey.
Databricks State of Data + AI, 2025. Only 19% of organisations surveyed have deployed AI agents of any kind. The remaining 81% are running AI at Level 0 (chat interface only). Source for the standfirst statistic.
TechCrunch / TechBuzz, February 2026. AI music generation platform. $300M ARR milestone published February 2026. Suno v5 is the current version, featuring improved audio quality and full commercial licensing. Used in the creative body section as an example of a specialist creative tool.
Google DeepMind, October 2025. deepmind.google/models/veo/. Veo 3.1 is the first video generation model to produce synchronised dialogue, sound effects, and ambient noise in the same pass as the video output. Previously required separate audio synthesis steps.
More articles
Seven Trends Reshaping Design: AI and the Next Three Years of Creative Work
The AI on your team will execute everything and decide nothing. Here are the seven shifts that follow.
The Taste Problem: AI Can Match Anyone’s Output. It Cannot Match Their Judgement.
When everyone rents the same intelligence, judgment becomes the moat.
The Second Customer: Your Product Has Two Users Now. One Cannot Read Your Homepage.
AI-sourced traffic now converts 42% better than human traffic.
Anyone Can Make It Now: When the Mona Lisa Took Eleven Seconds
Google made its film studio free. WPP cut a third of its creative headcount. The tools gap closed.
The Last Human Reader: How AI Became Your First Audience
The pages you publish are no longer primarily read by people.
LLMflation: The AI Gets Cheaper. The Bill Keeps Growing. Neither Is Your Fault.
Microsoft cancelled its Claude Code licences after engineers burned through its entire annual AI budget in weeks.
The Cheap Code Problem: What Snap’s Memo Got Right About AI and Engineering
Snap fired a thousand people because AI writes 65% of its code.
Own Your AI: Why Renting Intelligence Is About to Look Foolish
AI is shifting from a service you subscribe to, to a feature you ship.
Everyone Can Build It Now. Building It Is the Easy Part.
An AI-built social network was fully breached three days after launch. The gap between AI-generated code and production-safe code is not closing.