An agent decides its own next steps. Most office processes do not need that — they need fixed sequences.
01 In short: AI agents explained simply
Put simply, an AI agent is a language model that is given a task and decides for itself which tools to use in which order until it considers the task done. A workflow, by contrast, has fixed steps that someone set in advance. For office processes in real estate and construction — invoices, enquiries, damage reports, reports — the fixed workflow is almost always the better choice: it is checkable, does the same thing every time and costs less. An agent pays off where the route to the result really is different every time.
02 What makes an agent
Three things set an agent apart from a model step in a workflow:
- A goal instead of an instruction. Not "read field X from this invoice" but "find out why this invoice does not match the order".
- Tools. The agent may, for example, search the CRM, read e-mails, open documents or look up an invoice.
- A loop. It picks a tool, reads the result, decides on the next step and repeats until it is done or reaches a limit.
The loop is the decisive part. In a workflow it is settled in advance that step 3 follows step 2. With an agent it is only settled when it happens.
03 Workflow and agent compared
| Workflow with an AI step | AI agent | |
|---|---|---|
| Who decides the steps | Set in advance | The model, while it works |
| Same input, same route | Yes | Not necessarily |
| Traceable | Every step logged, always the same sequence | Can be logged, but every run looks different |
| Testable | With sample cases, result predictable | Only by sampling |
| Cost per case | One or a few model calls | Many model calls, hard to predict |
| Suited to | Recurring workflows with exceptions | Open tasks with a changing route |
04 Why most office processes do not need an agent
An incoming invoice is recognised, assigned, checked, approved, booked and filed. A tenant enquiry is sorted, matched to the property, answered or forwarded. These workflows are not open. They have exceptions, but the exceptions can be listed and each handled with a rule or a hand-over to a person.
Handing such a workflow to an agent means having it search for a known route every time. That is slower, more expensive and harder to check, without a better result. Where the AI belongs in the fixed workflow is described in Where the AI sits in a workflow — and where it does not.
05 When an agent makes sense
Some tasks really do take a different route every time:
- Research across several sources. A question about a project whose answer is sometimes in the minutes, sometimes in an e-mail, sometimes in a drawing register.
- Clarifications with follow-up questions. A discrepancy between invoice, order and delivery note, where you only know where to look once you have started.
- Assistance in conversation. An employee asks in their own words for information from several systems.
There too: the agent searches, reads and proposes. What follows is decided by a person.
06 AI agents: examples from real estate and construction
Three tasks where an agent has its place, and how you can tell it is one:
- Clarifying an invoice discrepancy (construction). A subcontractor's invoice is above the order. The agent searches on its own: the ERP for the order and variations, the e-mails for a commitment by site management, the site diary for the work done. It does not know in advance which source holds the answer. It puts a summary with references in front of site management.
- Answering a project question (property development). "What was agreed with the structural engineer about the slab thickness on the third floor?" The agent searches minutes, drawing revisions and correspondence and names the place where the agreement is recorded.
- Assembling a dossier (property management). For a meeting with an owner, the agent gathers what is open for a building: damage reports, current orders, decisions from the last owners' meeting.
Not an agent but a fixed workflow: capturing and assigning invoices, sorting tenant enquiries, producing weekly reports. There the route is the same every time.
07 The limits an agent needs
When we build an agent, we fence it in:
- Read-only tools wherever possible. Writing, booking and sending run through a fixed workflow step with approval.
- A ceiling on steps and model calls per case.
- A log of every tool call, so it is clear what the agent saw.
- A hand-over to a person as soon as the agent gets stuck.
As soon as a system rates people, for example applications or tenant dossiers, stricter rules apply. More in The EU AI Act for Swiss companies.
08 Agent and workflow together
In practice it is rarely either/or. A fixed workflow runs the process and calls an agent at one point, for example to clarify a discrepancy. The agent returns a result in fixed fields, and the workflow carries on as usual. The tools we build with can do both: n8n has its own building block for AI agents next to the ordinary steps, and can require approval by a person before specific tools of the agent run. More about the tool in What is n8n.
09 How to tell what you need
One question is enough: can you write down today's workflow step by step, with its exceptions? Then you need a workflow. Can you only describe the goal but not the route, because it is different every time? Then an agent can help. What a workflow is, is described in What is a workflow, what is an automation.
10 Frequently asked questions
Are AI agents not the future?
For open tasks, yes. For recurring workflows a fixed workflow remains the cheaper and safer solution, even as models get better. A route you already know does not need to be searched for every time.
May an agent send e-mails for us?
That is your decision. We build it so that an agent writes drafts and a person approves them before anything leaves the building.
Does an agent cost more than a workflow?
Per case usually yes, because an agent needs several model calls to find its route. How many depends on the case and can only be capped in advance, not stated exactly.
Sources
- n8n documentation: AI Agent node
- n8n documentation: Human-in-the-loop for tools