“Couldn’t you just build us a sort of ChatGPT for our accounting?” We hear this question often. Behind it lies an understandable misconception: much of what is sold as “AI” today is a chatbot. And a chatbot cannot book your receipts. It can explain to you how to do it. That is a difference like the one between a cookbook and a cook. Anyone making an investment decision for their SME should understand the difference, otherwise you buy the wrong thing or pay for features you don’t actually need.
The short definition first
A chatbot answers questions in a dialogue. You type or speak, it replies, and that is the end of its contribution. It delivers information, not action.
An AI agent carries out a task. It takes an input, plans several steps, accesses other systems, makes decisions within fixed limits and, at the end, delivers a result rather than just an answer.
Put simply: the chatbot talks, the agent gets something done. Everything else is a gradation of this one distinction.
Three characteristics that pin down the difference
Rather than going by marketing terms, it is worth looking at three sober criteria. They let you reliably identify what you are dealing with.
| Criterion | Chatbot | AI agent |
|---|---|---|
| Context memory | Usually only remembers the current conversation, and often only to a limited extent. Once the window has passed, the thread is gone. | Retains the status of a process over time. Knows what has already been done and what is still open. |
| Decision autonomy | None. It outputs what it knows. The human has to act. | Makes its own decisions within clearly defined limits, such as assigning a booking or escalating a case. |
| Multi-step action | One question-and-answer step. No chaining of actions. | Plans and carries out several steps in sequence: reading out data, checking, writing into a system, reconciling, reporting. |
This is the quotable core: context memory, decision autonomy and multi-step action. If a system has all three, you are talking about an agent. If at least one of them is missing, it is a chatbot, no matter how modern the language model behind it is.
One important point here: “decision autonomy” does not mean the machine does as it pleases. With a seriously built agent, the limits are defined in advance. What falls within those limits, it handles; everything else goes to a human. Drawing those boundaries is the actual work, not the language model itself.
When a chatbot is perfectly sufficient
A chatbot is not the inferior tool. It is built for a different task. For the following cases it is enough, and an agent would be over-specified:
- Providing information: opening hours, delivery dates, frequent questions, directions.
- Making information findable: “Where do I find the holiday rules in the handbook?” The bot searches documents and replies.
- Pre-qualifying enquiries: clarifying what the matter is about before a human takes over.
- First contact in service: catching standard requests and passing the rest on to the responsible person.
The common denominator: it stays at talking and looking things up. No one has to change anything in another system. If that describes your requirement, you can spare yourself the complexity of an agent.
When you really do need an agent
As soon as there has to be an action in a system at the end, a chatbot can no longer help. Typical signs:
- A task consists of several steps that someone currently strings together by hand.
- Data has to be read out of one document and written into another system.
- There are recurring decisions based on fixed rules that currently eat up time.
- The process stretches over hours or days, and you need to know where it currently stands.
The honest rule of thumb: if you need an answer, a chatbot is enough. If you need a completed result, you need an agent. Most expensive wrong purchases happen because these two things are confused.
A practical example: from receipt to booking
Take the classic of everyday fiduciary and SME work, the path from receipt to booking. In many places it works like this today: a person opens the invoice, reads out the supplier, date, amount and VAT, assigns the correct account, types everything into the accounting software and files the receipt. A few minutes per receipt, which adds up considerably over the year.
Here, a chatbot could at most explain which account a particular expense belongs to. The work itself would remain entirely with the person.
An AI agent goes the whole way:
- It reads out the receipt, that is, supplier, date, amount and VAT rate.
- It suggests the appropriate account, based on earlier bookings from the same supplier.
- It detects duplicate entries and anomalies.
- It books clear cases, while it submits unclear or unusual ones to a human for review.
- It records which receipt was booked in what way, so that the process remains traceable.
This is not a concept from a slide presentation. We run such an accounting agent in productive daily use in our own US-based CPA/accounting practice (in the US context this is referred to as a “CPA firm”, not a fiduciary practice in the Swiss sense). It takes over the routine part of receipt processing; the professional responsibility and the tricky cases remain with the human. To be honest: for the Swiss fiduciary market, this solution has not yet been rolled out. The building blocks are running in production; in this sector and under Swiss conditions we are still in development. So we are not selling you a finished Swiss case, but a proven mechanism that we adapt to your requirements.
What you must not forget legally in accounting
An agent in accounting touches on obligations that apply regardless of the technology. These points are not a technical detail but a matter for management:
- Bookkeeping obligation: The obligation to keep proper accounts and financial reports applies under Art. 957 CO to all legal entities, as well as to sole proprietorships and partnerships with a turnover of at least CHF 500,000 in the last financial year. Those below this threshold keep simplified accounts covering income, expenditure and asset position.
- Retention: Business books and accounting records must generally be retained for ten years under Art. 958f CO. For documents relating to immovable property, VAT law requires a longer period: twenty years under Art. 70 para. 3 VAT Act. This also applies to digitally processed receipts.
- Integrity and availability: Anyone archiving electronically must comply with the requirements of the Ordinance on the Keeping and Retention of Business Books (GeBüV). It requires integrity, that is, the immutability of the data (Art. 3 GeBüV), and availability at all times (Art. 6 GeBüV); the legibility of the information media must be checked regularly (Art. 10 GeBüV). An agent must therefore not file a receipt in such a way that it could later be altered unnoticed.
- Responsibility: The obligation to keep proper accounts remains with you, not with the system. If the bookkeeping or retention obligation is breached, this is a contravention under Art. 325 SCC that can be punished with a fine (the maximum amount of a fine is CHF 10,000 under Art. 106 SCC, unless the law provides otherwise). In connection with bankruptcy, failure to keep accounts can under Art. 166 SCC even be punished with a custodial sentence of up to three years or a monetary penalty.
The lesson from this is not that an agent in accounting would be problematic. On the contrary: a cleanly built agent can even improve traceability, because it logs every step. What matters is that the legal guardrails are considered from the outset and that a human retains control. Anyone who takes this seriously builds not a chatbot that pretends, but an agent that works reliably within clear limits.