This is the first of a monthly series. It starts here because almost every conversation we have about AI stalls on the same thing: nobody has told the person what the tool actually is.
Claude is a large language model made by Anthropic. Strip away the branding and it does one mechanical thing: given some text, it produces the text that should plausibly come next. That is genuinely all. Every impressive thing it does — drafting, summarising, extracting, translating, writing code, answering questions about a document you hand it — is that single operation applied to a carefully arranged input.
Understanding that one sentence explains most of what follows, including the failures. It is not looking anything up. It is not consulting a database of facts. It has no memory of your last conversation unless something deliberately gives it one. And it does not know when it's wrong, because "wrong" isn't a category it can perceive — only "less likely to come next."
Cut through the demos and the useful applications for a small organization fall into three groups.
You have a purchase order as a PDF, an intake form as a scan, an email describing a change request. You need it as rows in a system. Historically that was a person retyping, or an integration project with a rules engine that broke whenever a supplier changed their layout. This is the strongest and most boring use of the technology, and it's where most of the money is. Every organization we work with has at least one of these, usually several, usually invisible because the person doing it has always done it.
Not "write me a blog post" — that produces the flavourless text everyone can now spot. Rather: here are my meeting notes, produce the client summary in our format; here is the incident log, produce the report the board needs; here are eleven support tickets, tell me what they have in common. The value comes from you supplying the substance and the model handling the assembly.
Point it at your policies, procedures, or historical records and let staff ask questions in plain language. Done properly this is genuinely transformative for organizations whose knowledge lives in four hundred documents nobody can search. Done carelessly it invents answers with total confidence. The difference is entirely in the engineering, which is why we'll spend a whole post on it in January.
Arithmetic and anything requiring exactness by nature. It can produce a spreadsheet formula, and it can write code that calculates correctly. But asking it to add up a column in its head is asking a text-prediction engine to do accounting. The fix is architectural: have it write the calculation, then let the computer run it.
Knowing current facts about your business. The model knows nothing about your inventory, your staff, or last Tuesday. Everything specific has to be handed to it in the request. When a tool appears to know something current, it's because an engineer built a pipeline that fetches the current thing and includes it. That pipeline is the actual work.
There are three tiers, and the useful distinction is capability against cost per unit of text — Anthropic prices per million tokens, roughly 750,000 words:
All the current models handle up to a million tokens of context in a single request, which in practice means you can hand one an entire contract, a year of email, or a substantial codebase and ask questions about the whole thing.
For an organization just starting out, the more relevant choice is whether staff are using the consumer subscription or whether you're building on the API. Those two paths have very different privacy and control properties, and confusing them is the most common mistake we see. That's November's post.
Not headcount, in our experience — at least not in organizations of the size we work with. What it changes is which work is worth doing at all. Tasks that were previously too expensive to bother with — reconciling two systems monthly instead of annually, reading every piece of customer feedback instead of sampling, keeping documentation actually current — become cheap enough to just do.
The second thing it changes is the cost of a first draft, which is more significant than it sounds. A great deal of organizational work stalls not because the thinking is hard but because starting is. When a serviceable draft of the policy, the report, or the proposal costs ninety seconds, more things get finished.
What it doesn't change: whether your systems talk to each other, whether your numbers reconcile, whether anyone owns the process. AI sits on top of your operations and inherits their quality. If four systems disagree about a customer's address, a language model will confidently pick one. Which is why, when someone asks us to build them something with AI, we usually start by looking at the plumbing.
Ten questions, about three minutes, no email required. It scores your data, process, people, and guardrails, and tells you which one would sink a build. Some results say "this isn't an AI problem yet" — that's a real answer.
Take the AI readiness checkNext month: Claude inside Excel and Word — what the Microsoft 365 add-ins actually do, and the governance question to settle before you turn them on.