AI Business Analysis

What cross-licensing is: how two builders who each own a patent the other needs get out of a standoff by letting each other use their patents, the honest limit of a licence, and the one question to ask.

What gross margin is and why AI companies often keep less of each dollar than software companies: a 90% versus 50% example, the twist that cost grows with use, and the one question to ask.

What separates an AI agent from a chatbot: a model that works in a loop, using tools, until the task is done, why small mistakes add up, and the one question to ask.

How an AI model can read a photo of a broken part: multimodal models, why they matter for work that is not typed, two cautions, and the one question to ask.

Why the cost of AI does not end when training ends: inference, the cost paid on every answer, how it compares with training, and the one question to ask.

How a hidden line in an email can make an AI assistant forward invoices: prompt injection, why a model cannot tell orders from content, and the one question to ask.

What open weights means in AI: a model whose parameters are published so anyone can run it, how that differs from a closed model, what it costs to run, and the one question to ask.

Why an AI model can get a clause on page two wrong after reading a 300-page contract: the context window, what shares it, and the one question to ask.

A 7-minute worked example of AI sovereignty engineering: an agent classifies 12 workloads, traces a request, reads an offer, prices the premium, drills the exit and reads the meters.

A 14-minute course on AI sovereignty as three verbs, read, stop and change: the seven-rung ladder, the map of sellers, the premium, the break-even sum and the meters.

How an AI model that needs 16 GB can be shrunk to 4 GB by storing its numbers with less precision: what quantization is, what it costs, and the one question to ask.

How AI reads every word at once: attention, the transformer, why the design scaled, why doubling the text roughly quadruples the work, and the one question to ask.

Why AI labs keep building bigger models: scaling laws, the steady line that lets them estimate the next model, what each step costs, and the one question to ask.

Why a high test score does not tell you which AI model does your job: benchmarks, their two limits, a manager’s own 20-case test, and the one question to ask.

Why connecting AI models to company systems produces a pile of custom connectors, and how the Model Context Protocol (MCP) replaces 30 of them with 13.

What parameters are in an AI model: the dials that training adjusts, where a model’s knowledge is stored, why more of them costs more, and the one question to ask.

Why an AI assistant greets a returning customer like a stranger, how a notebook called agent memory fixes it, and the three questions the notebook raises.

What stops an AI model from doing something it should not: guardrails, the two places they sit, the cost of setting them too tight, and the one question to ask.

What happens when a model writes its own training examples: synthetic data, what you gain, the quiet risk, and the one question to ask.

Why the same AI model gives the same answer every time or a different one each time: temperature, low versus high settings, and the one question to ask.