Chapter 0
Separate the model from the harness
When tools look like magic, judgment fails
The talk around language models is loud, and almost none of it explains how the software actually works.
The market pulls in two opposing directions at once. Salespeople promise that software will read every company report, calculate every ratio, and draft your research memos before breakfast. Skeptics warn that it will invent figures out of thin air, leak private client records, and embarrass you in front of an audit committee. Both sides make the tools sound like magic.
Magic is a terrible basis for financial work. When you cannot see the moving parts of a tool, you are forced to choose between blind trust and stubborn refusal. You either accept whatever the screen prints, or you refuse to touch the keyboard at all. Neither posture is defensible.
This book has one purpose: you will learn to reason about the LLM. When it drafts a company report, you will know what just moved across the wire. When it changes a revenue figure between two identical runs, or drops a quarter from a table, you will know exactly why it happened. Better yet, you will know how to catch it.
You need very little technical background to understand this. You do not need a computer science degree. You do not need matrix calculus or neural network theory. You only need plain English and curiosity about how your daily tools operate. Soon enough, the mystique is gone, and an ordinary instrument takes its place.
The window is a wrapper around a distant brain
You probably already know how to drive a chat window. You type a question into a white box, press return, and watch complete sentences appear word by word. It feels like sending an instant message to a very fast, polite colleague who never sleeps and never complains.
That familiar window is actually two completely separate things bolted together: the LLM and the harness.
The first piece is the LLM. This is the brain, the genuine novelty, and what most people mean when they say "AI." An LLM is tens or hundreds of gigabytes of mathematical parameters, probably running across specialized processors in an industrial data center. It consumes immense electricity, reasons across vast corpora of text, and solves difficult analytical problems. Yet for all its problem-solving power, it is completely blind and immobile on its own. It takes text in, calculates probabilities, and emits text back. It does not know who you are, what firm you work for, what subscription tier you pay for, or what time it is in your city. Left to itself, it sits entirely inert.
The second piece is the harness. The harness is an ordinary, traditional piece of software. It does not think, reason, or predict words. Instead, it acts as the intermediary between you and the LLM, giving the model hands and eyes to interact with the world. Without a harness, the LLM has no way to receive your questions, browse the web, read your spreadsheets, or store the work it finishes. The harness runs in your web browser and on the vendor's servers. It remembers your login, stores your previous conversations, formats text into tidy paragraphs, accepts file uploads, and bills your corporate credit card every month.
When you drag a financial statement into the chat window, the LLM never sees a file. It cannot read documents or open attachments. Instead, the harness cracks open the file, extracts the raw letters from each page, pastes those letters directly alongside your question, and sends the whole bundle across the wire. The harness acts as a courier carrying text. The original file never leaves your desk.
The harness enables the LLM to search websites, read filings, and talk to you like a colleague. Take the harness away, and the LLM sits in total silence.
APIs remove the courier
The companies that sell chat subscriptions also sell direct access to the LLM alone.
Engineers call this direct access an API, which stands for Application Programming Interface. You can picture an API as a doorway built for software instead of people. A human needs buttons, text boxes, and scrollbars. Software only needs an open network connection and a clean line of text.
Direct access changes how an organization operates far more than the chat window ever did. A chat window sits beside one analyst, waiting for someone to copy and paste paragraphs into a browser. Direct access connects the LLM straight to your firm's data pipelines. A script can wake up at dawn, pull the latest earnings releases for two hundred public companies, extract reported revenue numbers, and write the verified results into an internal database before anyone arrives at the office. An earnings release is a company's public statement of its quarterly performance.
Before your firm wires an LLM into pipelines that handle real capital, you need to understand it on its own. With direct access, there is no polite chat software to smooth over mistakes, check file sizes, or apologize when numbers drift.
Never call the wrapper an AI
Much of the confusion in this field comes from sloppy vocabulary.
Marketers use the label "AI" to describe three completely different things: an entire academic discipline, an underlying mathematical model, and a commercial product you download from an app store. When one word means three things at once, nobody can talk clearly about what they are buying.
In this book, we keep the nouns strictly separated:
An LLM is the raw model. It predicts the most likely next word from the words that came before it.
An AI product is the commercial harness. It pairs an LLM with a user interface, file processors, cloud storage, security filters, and a billing meter.
The chat window is not the LLM. It is just the harness the vendor built to deliver it to your desk.
The Leader's View: Rent the harness or own the pipeline
Treat the LLM and the harness as two completely different purchasing decisions.
When you buy a commercial product, you pay for the convenience of a ready-made harness. You accept the vendor's interface, their file size limits, their document retention terms, and their price increases. If the startup behind the product runs out of funding or gets acquired, your internal workflow breaks with it. A prebuilt harness works well for an individual analyst drafting a memo, but it creates silent dependencies across a department.
When you build with direct access, your firm keeps custody of the workflow. You decide where customer data travels. You control how source documents are parsed, which external calculators verify arithmetic, and whose signature must approve a figure before it reaches a client or an audit committee. You own the workflow instead of renting it.
| Dimension | Rent the Harness (Commercial Product) | Own the Pipeline (Direct API) |
|---|---|---|
| Data custody | Uploaded documents traverse vendor servers and retention policies | Data routes directly from your internal perimeter to the model |
| Workflow lock-in | Tied to vendor UI features, pricing changes, and startup solvency | Workflows live in your internal codebase; models can be swapped |
| Verification & audit | Output appears in chat; human must manually paste and inspect | Scripts run automated calculators, schema checks, and audit logs |
| Cost structure | Fixed monthly seat license per user | Variable cost billed strictly per token processed |
| Best fit | Ad-hoc drafting, individual memos, qualitative exploration | Repeatable data extraction, batch filings, regulated reporting |
Commercial products come and go with venture capital cycles. The startups dominating headlines this quarter will look different in three years. But the mechanical properties of an LLM will remain exactly the same. Once you understand the core mechanics, you can evaluate any vendor pitch with clarity.
This book builds that mental model across four operational perimeters:
- The Single Call: exploring prompt mechanics, next-token prediction probabilities, sampling temperature, and why identical prompts produce drifting numbers across distributed hardware clusters.
- The Conversation Notebook: opening multi-turn dialogues, tracing the replayed transcript notebook, compounding token costs, statelessness, and attention decay across long financial filings.
- The Project Boundary: establishing repeatable enterprise workspaces, grounding models in audited document coordinates, and connecting external deterministic calculator tools.
- The Governed Pipeline: establishing institutional controls, deploying defensive checks against predictable failure modes, decomposing workflows across specialized agent pipelines, and building audit-defensible compliance dossiers.
Throughout this journey, you will follow one public company, Acme Industrial Technologies, and trace one specimen revenue figure: $4,280 million. Follow the number.
The next chapter opens the engine to see how next-token prediction actually works, and why generating text is fundamentally different from retrieving a financial record.
Zoom-out: Distinguish between brands, harnesses, and LLM models
Enterprise software vendors sell unified brand names. When an account executive pitches your investment committee, they present a single package with a monthly subscription and a proprietary logo.
As an analyst or a decider, your first task is to break that brand apart. Ask two questions: which software harness manages the workflow, and which mathematical model performs the inference? The table below separates the major commercial offerings of late 2026 into their moving parts.
| Brand | Harness | LLM Models |
|---|---|---|
 OpenAI |
ChatGPT on web and mobile, plus Codex for coding | GPT-6 Astra is the top model, with the cheaper GPT-5.6 family below it |
 Anthropic |
Claude on web and mobile, plus Claude Code and Cowork | Opus, Sonnet and Haiku, with Fable as a new top tier above Opus |
 Microsoft |
Copilot inside Word, Excel, Windows and GitHub | A mix of three: OpenAI's models, Anthropic's Claude, and Microsoft's own MAI models |
 Google |
The Gemini app and Google Workspace | Gemini Pro for hard questions, Gemini Flash for everyday speed |
 Cursor |
A code editor for developers, now owned by SpaceX | Its own Composer model, plus Grok, and it can still call Claude and GPT |
 Grok |
The Grok app and website, built into X, plus Grok Build for coding and Cursor since the SpaceX deal | The Grok 4 series, currently Grok 4.6, marketed as fast and cheap for its quality |
 Kimi |
Kimi chat on web and mobile, plus a coding tool | The Kimi K3 model, which anyone can download and run themselves |
 DeepSeek |
DeepSeek web and phone app | The V4 family, in a large version and a fast cheap version |
 Z.ai |
Z.ai chat and a coding agent called ZCode | The GLM 5 series, free to download and self-host |
 Mistral |
Vibe, the French chat app formerly called Le Chat | The Mistral model family, sized from Small to Large |
 Cohere |
North, a private workspace built for large companies | The Command family, aimed at searching company documents |
Notice how quickly the layers decouple. One brand sells three different models behind a single search bar. Another harness swaps between competing frontier providers depending on the task. Startups get acquired, brand names merge, and subscription tiers shift every quarter.
The marketing changes constantly. The mathematical machine underneath does not. Stripped of its harness and its logo, every model on this table performs the exact same mechanical operation: it takes an incoming stream of tokens and estimates which token should follow.