Anyone talking about AI in the office today usually means Microsoft Copilot or ChatGPT. Claude, from Anthropic, comes up less often in these conversations, even though it is already running in many companies — often unofficially at first, because individual employees discovered it for themselves. This article sets out what Claude actually delivers in day-to-day work, how it differs from Copilot, and which questions you should have answered before rolling it out.
The decisive difference is not the model
It is tempting to compare AI assistants by their benchmarks. For business use, that is the wrong level to look at. The models from Anthropic, OpenAI and Google are close in capability, and the lead changes hands several times a year. What does not change every quarter is how each one is built.
Copilot is deeply anchored in Microsoft 365. It sees whatever your account has access to — emails, files, appointments — and works directly inside Word, Excel and Teams. That is both its strength and its precondition: where permissions are not maintained cleanly, that becomes visible immediately.
Claude takes the opposite approach. By default, it has no access to your company systems. You bring along the documents you want to work on. That sounds like a drawback, but in many cases it is exactly why departments get up and running faster with it: no licence changes and no permissions clean-up are needed before the first useful step becomes possible.
What Claude actually delivers in day-to-day work
Three building blocks shape everyday work with it.
Conversations are the starting point — a dialogue in which you think through a problem step by step, rather than sending a question and receiving an answer. The difference from a search engine lies in the follow-up.
Projects are workspaces with their own store of knowledge. You upload the documents that keep coming up — quote templates, product descriptions, internal guidelines — once, and every conversation in that project can draw on them. For recurring tasks, this is where the biggest time saving comes from, and it can be shared across a team.
Artifacts are results that do not disappear into the chat history: a document, a spreadsheet, an analysis that is created alongside the conversation and that you refine together until it is right. Anyone who has ever tried to shape a long piece of text inside a chat window knows the difference.
On top of this come connections to third-party systems. An open standard allows you to connect file storage, calendars, ticketing systems or a CRM, so that the assistant can read there and, under certain conditions, write as well. This is exactly where the part that needs a decision begins — more on that below.
Directly inside Excel, Word, PowerPoint and Outlook
The development changing the most in Swiss offices is a recent one: Claude is now available as an add-in directly inside the Microsoft Office applications. Excel, Word and PowerPoint are generally available, and Outlook is in preview. It is delivered through Microsoft AppSource and installed centrally by administrators in the Microsoft admin centre – your employees do not need to set up anything themselves. It is included in the paid plans.
The difference from a browser window is bigger than it sounds. You stay inside the file you are already working on: the assistant sees the actual spreadsheet, not a copy pasted into a chat. And because a single conversation carries context across all four applications, changes carry through – you adjust an assumption in the Excel calculation, and the figure in the Word memo and the chart in the presentation follow suit.
For a training organisation like us, that is the immediately visible benefit: people who are already good at Excel become faster with it. People who are not do not suddenly become proficient just because they have it – tool knowledge remains the foundation, and the assistant is the lever on top of it.
Where the limits lie
Three points that are regularly underestimated in practice.
First: the assistant knows nothing about your company that you have not given it. It does not know your price list, your project history or your internal abbreviations. That is not a shortcoming but a consequence of how it is built — but it means that the quality of the results depends directly on how well you supply the context. That is a skill you can learn, and it is why training achieves more here than licences alone.
Second: confident wording is not proof of accuracy. Wherever figures, deadlines, legal questions or commitments to customers are involved, every statement needs to be checked. The assistant speeds up the route to a draft; it does not replace professional responsibility.
Third: not every task benefits. For short, clearly defined activities, writing a good prompt takes longer than the task itself. The benefit shows up where scope, repetition or research are involved.
Data protection: what Swiss companies should clarify
Anthropic does not train its models on business customers’ conversation content by default; for personal accounts, contributing to model training is a setting that must be actively switched on. According to the provider, deleted conversations disappear from its systems within 30 days. Business plans and the API are subject to their own, more extensive terms.
That does not let you off your own homework. Under the revised Federal Act on Data Protection (FADP), your company remains responsible for what your employees enter into such a tool. Three questions should be answered in writing before the first licence is purchased: Which categories of data may go in, and which may not? Who is allowed to authorise connections to company systems? And how are results that were produced with AI assistance labelled?
Anyone operating in the EU also needs to keep an eye on the AI Act. For general office use, the obligations are manageable, but the requirement to train staff who use AI systems affects practically every company.
Before you roll out licences
Experience from rollout projects is clear: the bottleneck is rarely the tools, but the use cases. Companies that hand out licences broadly without naming the three to five specific tasks the tool is meant to improve typically see a usage curve heading towards zero after six weeks.
What works well: a small group starts with clearly named tasks, builds reusable templates for them, and those templates become the basis for the wider rollout. At the same time, set the ground rules — better to have three that everyone actually follows than a ten-page document nobody reads.
Together with Power Automate: the draft is ready, you decide
This is where it gets really interesting for companies. An AI assistant writes well, but it does not start on its own. A flow in Microsoft Power Automate starts reliably, but writes clumsily. Together, they cover the whole distance.
An everyday example: an enquiry arrives in the shared mailbox. The flow recognises it, retrieves the customer data from the CRM and the matching terms from the price table, and places the reply as a draft in the mailbox – written in your house tone, because the template for it has been set up. A person reads it, changes a couple of sentences and sends it.
That is the honest benefit, and the emphasis is on the last sentence. Not “communication runs by itself”, but: the rough draft is ready by the time you open the mailbox. For recurring enquiries, that saves most of the time, and responsibility for what goes out stays exactly where it belongs.
Three rules have proven useful here. Only automate once the process works cleanly by hand – an unclear process does not become clearer through automation, only faster at going wrong. Do not let anything go out without human sign-off whenever customers, money or commitments are involved. And start with the enquiry that arrives most often, not the most complicated one.
Training: what makes sense
There is no vendor certification for end users of Claude. What counts is the practical ability to write good prompts, supply context and critically check results. At CT Academy, several courses cover this, depending on your starting point:
- Working with Claude: the AI assistant in everyday business — one day of hands-on practice, from the basics to projects, artifacts and connections
- Prompting that works — the craft techniques, independent of the tool
- AI assistants in the office: ChatGPT, Claude and Copilot compared — if the decision is still open
- AI for executives: applications, risks, regulation — for the level that sets the framework
- Microsoft Power Automate / Flow and 55268 Microsoft Flow / Power Automate — the automation side, from the first flows to approval steps
- PL-900 Microsoft Power Platform Fundamentals — the certified grounding for anyone responsible for this topic in the company
We run all courses as an online live webinar, in the classroom in Wangen-Brüttisellen, or as in-house training at your premises.
In brief
Claude is not a replacement for Copilot, nor a Copilot alternative in the strict sense, but a tool built along different lines. Where Copilot draws its strength from its closeness to your Microsoft data, Claude’s strength lies in working with the documents you deliberately give it — and in building reusable structures for tasks that repeat. Many companies end up running both side by side. What matters is not which logo is on the tool, but whether your people know what to use it for, and where to keep their hands off.
