Last month covered what Claude is. This month: the version that lives inside the software your staff already have open, which is where most organizations will meet it first — whether or not anyone decided to adopt it.
Anthropic has spent the past year putting Claude directly inside Microsoft 365. Excel arrived first, in October 2025. PowerPoint followed in February 2026, Word went to public beta in April, and Outlook is still in beta. They are add-ins on paid Claude plans, and since a March 2026 change called Office agents, a single conversation can span the apps — an analysis you do in Excel can feed a deck in PowerPoint without copying anything across.
That is a materially different proposition from a chat window in a browser tab. It is worth understanding precisely, because it changes both what your staff can do and what leaves your organization.
The useful mental model: it can read the sheet, understand its structure, write formulas, and explain what it finds. Concretely, the things we see land well:
What it does not do is guarantee arithmetic. This bears repeating because a spreadsheet feels like a calculator, so people assume the assistant inside it is calculating. It is writing formulas, and Excel does the calculating — that part is trustworthy. But if you ask it to eyeball a column and tell you the total, you are asking a text-prediction engine to do mental arithmetic. Get it to write the formula. Then the answer is Excel's, not the model's.
Word is the newer integration and the more conventional one: drafting, editing, restructuring, and — the part that earns its keep — redlining against a source. Handing it a contract and a set of standard terms and asking where they diverge is a real professional task done competently. So is turning a long report into a two-page summary that keeps the numbers straight, or rewriting a document to a house style you describe once.
The failure mode is subtler than in Excel and worth watching for: it will happily produce confident, well-structured prose about something it has misread. A document that reads authoritative is not evidence that it is right, and the more polished the output, the less carefully people check it. Every workflow that ends in a document going to a client, a funder, or a regulator needs a human who is accountable for the content, and who knows they are.
The add-ins come with paid Claude plans, which means per-seat subscriptions rather than the per-token pricing of the API. For an organization the practical question is how many people genuinely need it. In our experience it is not everyone: finance, operations, anyone whose week involves reshaping data or producing documents from source material. Give it to those people first, deliberately, and see what happens over a month.
Be aware that the alternative is not "nothing." If you don't provide a sanctioned tool, some of your staff are already pasting your data into a free consumer account on their own login. That is the actual comparison, and it is worse on every axis that matters.
Here is the thing nobody asks before turning the add-ins on, and the reason we get called afterward: what happens when someone asks it to summarise a spreadsheet of client records?
The data goes to Anthropic to be processed. That is not a scandal — it is how the product works, it's the same as any cloud service processing your files, and the business plans carry commitments about not training on your content. But it is a decision, and under PIPEDA it is your decision to make and document, not your staff's to make ad hoc on a Tuesday. Three things worth settling in writing before rollout:
None of that is onerous. It's an afternoon, and it turns an uncontrolled situation into a defensible one. We do this with clients as a small piece of work, usually alongside a look at what data is actually sitting in those spreadsheets, which is often more alarming than the AI question.
Turn it on for a small, deliberate group. Pick the people who spend their days reshaping data or drafting from source material. Write the one-page data policy first. Then, after a month, ask them what they used it for — the answers will tell you far more about where AI belongs in your organization than any vendor demo, and they usually point at one specific repetitive job that is worth automating properly.
That job is the thing to build next. Which is what this practice actually does.
That's the right question to ask before rollout, not after. Tell us what you're working with and we'll tell you plainly where the lines are — the AI readiness check covers the guardrails question in about three minutes.
Take the AI readiness checkNext month: running AI on your own hardware — Ollama, Qwen, and the three situations where local beats the cloud.