Most people use AI in a chat window in the browser tab next door. You type a question, copy the answer, paste it into a document. That works and is manual labour in the long run.
The next step is not a better model but a different place. The assistant should not sit next to the work, it should sit inside it.
For a copywriting studio in Germany we built exactly that: a complete workplace that opens in a browser, with the assistant right beside it.
What It Feels Like#
You open an address, type a password, you are in. On the left a familiar editor window with the files, on the right a chat where the assistant thinks along. No installation, no program that has to live on one particular machine.
In practice that means the workplace is the same on the Mac at home as on the iPad on a train and on a borrowed machine in a hotel. There is no version that is newer on one device than another, and nothing you can forget to bring.
For a one-person business that is worth more than any extra feature. The alternative, a workplace that exists on exactly one laptop, is an outage waiting to happen.
The Assistant Sees The Files, Not Just The Text#
The real difference to a chat window is access to documents.
A client sends a briefing as a Word file, a spreadsheet with target group data, a PDF with requirements. The assistant reads that directly, without anybody copying text out. And it writes back the same way: a Word document with headings, lists and formatting, not a block of text you tidy up by hand afterwards.
That sounds like a small thing and is the point where the time saving actually appears. Copy and paste costs only a minute in any single case. Across a month of jobs it is the difference between a tool and a toy.
The inbox is connected as well, though in one direction. The assistant can read the mail and prepare replies. It sends nothing. That is not a technical limit but a decision, and the most important one in the whole setup.
Memory Is The Part That Stays#
A chat window starts from scratch with every conversation. In copywriting that is the most expensive drawback there is, because knowledge about a client is precisely what you get paid for.
So this workplace has a memory attached. It keeps clients, style rules, earlier decisions and agreements across weeks and months. On the next job the assistant knows which client needs which tone, what worked last time and what is off-limits.
That changes the nature of the collaboration. Without memory you explain the context again before every task, and eventually you stop and do it yourself. With memory, explaining pays off, because it happens once and then holds.
Two Guard Rails That Were Not Negotiable#
The first concerns files. The assistant cannot delete anything unintentionally and cannot blindly overwrite a file. Anyone letting a language model work inside a directory of real client work has to build that barrier before doing anything useful with it. Otherwise the question is not whether, but when.
The second concerns style. Before every reply the assistant is handed the house style rules. Sounds trivial, but it is the difference between usable drafts and the smooth default tone everybody recognises instantly. A copywriter whose drafts sound like a text kit gets nothing out of the tool.
Where The Data Sits, And What That Actually Means#
The server stands in Germany and belongs to the studio. The credentials for mailbox and documents live exclusively there and do not run through our systems.
What belongs in the same breath, honestly: whatever is sent to the language model leaves the EU, because that is where the model runs. This distinction gets blurred a lot. So on every project like this we write down what leaves the server and what does not, and the client reads it before the start, not afterwards.
Whatever access is needed for maintenance is written down in the project, traceable, and under the owner's control. That belongs settled before the start, not afterwards.
Who This Fits#
Not everybody. If you use AI occasionally for a phrasing, you do not need a server of your own.
The setup pays off when three things come together: you work with it daily, you work with documents rather than only text, and it involves clients whose preferences you do not want to explain from scratch every time.
Then the maths tips over. And the difference is not that the assistant gets smarter. It is finally standing in the right place.
