The model takes the one step that has to understand language or documents. The rest stays ordinary, checkable logic.
01 In short
In a workflow we build, a language model only does what fixed rules cannot: sort an informal e-mail, read an invoice whose layout differs every time, write a draft reply. Everything else — transferring data, calculating, comparing, setting deadlines, filing — stays ordinary logic that does the same thing every time for the same input. Decisions with consequences are made neither by the model nor by the logic, but by a person. Our sentence for it: we automate the recurring work, not the judgement.
02 Why the question matters
A language model is strong at understanding and weak at guaranteeing. It reads a lease in seconds, but it can give two slightly different answers to the same question. A rule like "invoices above a certain amount go to management" is the opposite: it understands nothing, but it does exactly the same thing every time.
A good workflow uses each where it is strong. Let the model also calculate, book or decide, and you get a workflow that works most of the time. In accounting, most of the time is not enough.
03 The three kinds of step
Every step in a workflow belongs to one of three kinds.
| Kind | Examples | Who does it |
|---|---|---|
| Understanding | Sorting an e-mail, reading an invoice, finding a contract clause, drafting a reply | Language model |
| Rule | Transferring data, calculating, matching, setting a deadline, filing, reminding | Ordinary logic in the workflow |
| Deciding | Approving, paying, terminating, negotiating, committing to a tenant | A person |
This table is the core of every project. For each step it is settled which row it belongs to before anything is built.
04 An example: the damage report
A tenant writes: "Since yesterday it has been dripping under the sink, the cupboard is already wet." This is how the report runs through the workflow:
- Arrival. The e-mail lands in the property manager's inbox. Rule.
- Sorting. The model recognises: damage report, plumbing, kitchen, fairly urgent. It reads sender and text to do so. Understanding.
- Matching. The workflow looks up the sender in the property management software and finds the building and the flat. Rule.
- Routing. Following the stored list, the report goes to the plumber responsible for this building, as a draft. Rule.
- Approval. The property manager sees report, classification and draft order, and sends it or changes it. Deciding.
- Follow-up. If the plumber does not respond, the workflow sends a reminder. Rule.
One step in six is AI. The rest could be built without a model — only then every informal e-mail would end up in a catch-all folder.
05 Where the AI does not sit
- In calculations. Rent adjustments, service charges, cost estimates and totals are calculated by the workflow using fixed formulas. A model supplies at most the basis from a document.
- In bookings. What lands in accounting or the ERP is written by the workflow through the system's interface, following rules that can be checked.
- In decisions. Who wins the contract, which invoice gets paid, which tenant gets the flat: no model decides that. As soon as a system rates people, the legal requirements change too — see The EU AI Act for Swiss companies.
- In routing by fixed lists. Which contractor is responsible for which building is in a table. Nobody has to "understand" it.
06 What makes the AI checkable inside a workflow
A model step inside a workflow is easier to control than a chat window, for three reasons:
- Fixed output. The model fills predefined fields, for example category, urgency, amount, property. Free text only where it is needed.
- A log. Every run records which document was read and what was produced from it.
- A threshold. If the model is unsure or a field is missing, the case goes to a person instead of moving on.
Which model takes which understanding step is a separate question: Claude, ChatGPT, Copilot and Mistral for real estate and construction.
07 Data protection in the model step
Only the understanding step sends data to a model, and only the data it needs. Names, dates of birth and account numbers can be replaced by placeholders beforehand and put back afterwards. Where the model runs and who provides it is covered in AI models and data residency.
08 What this means for your project
When you describe a workflow, our first question is: which step really has to understand something? Often it is one, sometimes none. Then no AI is needed at all, just a clean connection between the systems you already have. How a workflow is built in principle is described in What is a workflow, what is an automation.
09 Frequently asked questions
Can a language model not simply do everything?
It can try a lot. But a workflow in which the model also books and calculates is hard to check and does not give the same result every time. We only give it the steps it is needed for.
What happens if the model sorts an e-mail wrongly?
The person sees it when approving and corrects it. If the model is unsure, the case does not move on at all but goes straight to a person.
Does every workflow need AI?
No. If all inputs arrive in structured form, as a form or a record from another system, rules are enough.