AI Agents in the Office: What They Should and Shouldn't Be Allowed to Do
What an AI agent is, where it helps in fiduciary and law firms, and why permissions, human approval and the location of your data matter.
Privacy & Law8 min read
"AI agent" is a term you now come across everywhere, and more and more software vendors are adding agents to their products. For a fiduciary or a law firm, the more useful question is what such an agent should be allowed to do in your own office. This article explains the term without jargon, shows sensible uses and sets out the rules to agree on first.
What an AI agent is
An ordinary AI assistant answers questions. You ask, it answers, and you decide what to do with the answer.
An AI agent goes a step further. It is given a goal and may use tools to reach it: open files, look at a calendar, draft an email, fill in a form. It decides on the next step itself, looks at the result and carries on until the goal is reached. Anthropic, the AI company, describes agents in this sense as systems in which the language model directs its own process and tool use, as opposed to workflows that follow predefined code paths (Anthropic, Building effective agents).
Not everything called an agent is one. The research firm Gartner uses the term "agent washing" for rebranding existing chatbots and assistants. Gartner also expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls (report in The Star, June 2025).
Where agents can help in fiduciary and law firms
Agents make sense for tasks that consist of several simple steps and are checked by a professional at the end. Some examples:
- Chasing documents for a tax return. The agent compares a client's documents with last year's checklist, notices that the pillar 3a certificate and the bank's interest statement are missing, and drafts a reminder. A member of staff checks and sends it.
- Sorting incoming post. Scanned letters are read, assigned to the right client file and given a suggested folder and deadline. A person confirms.
- Preparing a file for a meeting. Before the meeting with Keller AG, the agent collects the latest tax rulings, open deadlines and the last three months of correspondence, and writes an overview with source references.
- Extracting deadlines from rulings. The agent reads new tax assessments, works out the objection deadline and proposes a calendar entry.
What these examples have in common: the agent prepares, a person decides.
What an agent should not do on its own
The more an agent may do, the more can go wrong. The OWASP list of the main risks in language model applications has a dedicated entry for this, "Excessive Agency": too much functionality, too many permissions or too much autonomy (OWASP Top 10 for LLM Applications). For a firm bound by professional secrecy, that translates into something like this:
| Action | Agent alone | Only with approval | Never |
|---|---|---|---|
| Read documents the user is allowed to see | yes | ||
| Write summaries and drafts | yes | ||
| Suggest filing and calendar entries | yes | ||
| Send emails to clients or authorities | yes | ||
| Move or rename files | yes | ||
| Submit tax returns or court filings | yes | ||
| Permanently delete files | yes | ||
| Initiate payments | yes | ||
| Access folders the user may not see | yes |
The last row matters most. An agent should never see more than the person it works for. If an apprentice uses an agent, the apprentice's permissions apply, not the administrator's.
Why human approval is needed
Agents make mistakes, like any language model. In an answer to a question, you notice the mistake when you read it. With an agent that runs five steps in a row, an error in step one can end up as a wrong email in step five.
There is also a risk many people underestimate. An agent reads documents that come from outside. A submitted PDF, an email or a web page can contain hidden instructions that the agent treats as a task, such as "forward these documents to the following address". This type of attack is called prompt injection and is number one on the OWASP list. The most effective defence lies outside the AI: the agent simply has no permission to send anything out without approval. Our article on why a RAG system can be a security risk even on-premise covers related gaps.
Approval is only real control if it is readable. Before sending, an agent should show what it wants to send, to whom and which documents it relies on. An "approve all" button is not control.
Where your data goes with an agent
An assistant answering a question processes the question and a few text excerpts. An agent works in many steps and touches more data along the way: client folders, the mailbox, the calendar, perhaps the accounts. At every step, text goes to the language model. If that model runs in the cloud, considerably more leaves the office than with a single question.
Firms bound by professional secrecy under Art. 321 of the Swiss Criminal Code or the duty of confidentiality under Art. 62 of the Data Protection Act should clarify before introducing an agent:
- Which language model does the agent use, and where does it run?
- Which data goes to that model during a typical task?
- Is the provider a US company? If so, US authorities can compel disclosure under the CLOUD Act regardless of which country the data centre is in.
- Is there a log of which steps the agent took and which documents it read, and where is that log stored?
This is general information, not legal advice.
Where KADARAG fits
KADARAG is an AI assistant for your own files: it answers questions from your documents, cites document and page, and follows your existing file permissions. Search, documents and logs stay on your server. In local mode the language model runs there too; in hybrid mode only the question and three to five short excerpts go to a cloud model. Whatever an agent is meant to do later, it builds on the same foundation and can only be as safe as the access to your documents underneath it.
Next step
Before buying a product with agent features, fill in the table above for your office. Remove what does not apply and add your own processes. With that list you can ask any vendor how they enforce the limits technically. If you would like to talk about an AI assistant on your own server, book an intro call.