Claude Cowork: an honest account of using AI as your SME admin assistant

I have been quietly using Claude Cowork for a few months now, and I want to give you an honest account of how it works in practice for someone running a small business. Not a product review. Not a sponsored post. Just what actually happened when I started handing it real tasks.

The short version: some of what it does has genuinely changed how I spend my time. Other parts require careful thought about what access you give and why.

What is Claude Cowork?

Before getting into the specifics, it helps to understand what Cowork actually is, because it is meaningfully different from using Claude in the browser.

Cowork is a desktop application from Anthropic that gives Claude direct access to folders on your computer. You designate a folder, describe your task in plain English, and Claude executes it. It reads your files, creates new ones, reorganises and renames content, extracts data, and produces outputs like spreadsheets or documents. You do not need to copy and paste anything. You do not need any technical knowledge. You point it at your files and tell it what to do.

The key distinction is this: Claude Chat gives you answers. Cowork gives you finished work.

It is part of a broader category called agentic AI. If you want a more detailed explanation of what agentic AI is and what it means for how we work, I wrote about it here: Agentic AI: your new office assistant.

The VAT receipts example: three hours down to minutes

The task that sold me on Cowork was my VAT return preparation.

Every two months, I would spend three or four hours going through a folder of receipts, categorising expenses, and matching them to the structure of my existing Excel document. It was not complex work. It was just slow, tedious, and completely mine to do.

I gave Cowork access to the folder where I store my receipts and pointed it at the Excel structure I had already set up. It categorised the receipts according to my existing headings, flagged anything ambiguous for my review, and completed in minutes what used to take most of an afternoon.

That part alone was worth the attention.

What I did not expect was the follow-up. Cowork suggested some improvements to how I was organising the receipts in the first place: a more logical folder structure, consistent naming conventions that would make future categorisation faster. It did not just do the job. It thought about the system underneath it.

The permissions question: what access should you actually give it?

This is where I want to be direct, because it is easy to get this wrong.

When you set up a Cowork session, you choose a folder and assign it permissions. The options run from read-only (Cowork can look but not touch) through to full read, write, and delete access. Full access means it can delete files it judges to be redundant.

I would not give full access to any folder unless you are certain about two things: first, that you have a backup; second, that the content is genuinely expendable. AI makes mistakes. If you ask Cowork to remove duplicates and it misidentifies something, you want to be able to recover it.

Where full access makes sense is in a folder where you know there is redundant content and you are comfortable with it being gone. Removing duplicate files, for example, or clearing out an archive folder you no longer need. For anything where the original matters, stick to read or read/write without delete.

The principle here is the same one that applies to any tool with real consequences: start narrow, expand only once you understand the behaviour.

The disorganised folders problem

Once you see what Cowork can do with a well-defined task, you start looking at the rest of your file system differently.

I have folders I have been meaning to reorganise for years. Multiple folders that do essentially the same job, created at different times with slightly different naming conventions, making the whole structure harder to navigate than it should be. It is the kind of thing you know needs fixing but never quite rises to the top of the list.

Cowork is well suited to this. You can ask it to consolidate similar folders, apply consistent naming, and produce a cleaner structure without touching the content of the files themselves. It is also addictive in a mildly alarming way. Once one folder is sorted, the next one becomes obviously worth addressing.

What else can Cowork do?

Document creation is one of the main capabilities, though the regular Claude chat interface handles this too. Where Cowork adds value is when the document needs to draw from files already on your machine: pulling data from multiple spreadsheets into a single report, for instance, or turning a folder of notes into a first draft.

The other significant capability is skills.

What is a skill?

A skill is a text file, usually saved as SKILL.md, that you write once and load into Cowork. It contains standing instructions: how you want documents formatted, what steps to follow for a recurring task, what tone to use in a particular type of output. Once the skill is active, Claude follows those instructions automatically in every session where it is loaded.

The practical value is this: instead of explaining your preferences from scratch every time, you write them down once. The skill becomes a briefing that is always available. For business owners with consistent but repetitive processes, report formats, client communication templates, filing conventions, this is genuinely useful.

Why I think this is the best practical use of agentic AI right now

I have tried a number of agentic AI tools over the past year. Claude Cowork is the one I have kept using.

The reason is that it operates directly on your actual work environment. It is not a separate platform you have to migrate your files into. It is not a workflow tool that requires you to build automations. You give it a folder, describe a task, and it works with what you already have.

For an SME owner, that matters. The friction of learning a new system often outweighs the productivity gain. Cowork keeps the friction low because the interface is just plain language.

If you want to understand the broader context of what agentic AI means for businesses, including where the risks sit and how roles are likely to change, the post I mentioned earlier goes into more detail: Agentic AI: your new office assistant.

The honest caveats

Cowork is currently available on paid Claude plans: Pro, Team, Enterprise, and Max. It is a research preview, which means the product is still developing.

It will make mistakes. For longer or more complex tasks, check the outputs before treating them as final. This is not unique to Cowork. It applies to any AI-assisted work. But it is worth stating plainly.

The file access model is also worth thinking about carefully, as I mentioned above. Be deliberate about which folders you connect and what permissions you assign. More access is not always better.

Frequently asked questions