An email says, “Send me another six boxes like last month’s.” Before turning it into an order, you need to know who is writing, which product they bought and whether those six boxes are available. Reading the sentence is a small part of the job.
I would start by deciding what the system can do when one of those details is missing. That decision matters more than calling it an agent.
From text to a draft order
This is a hypothetical scenario. A company receives orders by email and wants to reduce transcription into its management software. The system suggests the product, quantity and customer; it then checks the catalogue and prepares a draft for a person.
The model can interpret unevenly structured text and suggest fields. The connection to the management software allows it to retrieve data or write a record. Rules check permitted quantities, existing product references and required fields. None of these pieces replaces the others: correctly understood text can end up attached to the wrong customer if identification is handled badly.
The n8n documentation on agents describes a model that decides which actions and tools to use. A workflow with predetermined steps may be enough for this order. If the sequence never changes, adding model decisions may complicate control without adding value. According to n8n’s official pricing (accessed: 11 October 2026), all plans include unlimited users: charges are based on monthly workflow executions, not on the number of users. This does not mean the service is free.
“Chatbot” does not necessarily mean a rigid menu either. A conversation can use models and tools; the practical question is what permissions it has and which checks precede an action.
Where I would put the review
In the first trial, the system would only create drafts. A person would check the customer, product reference, units and address before confirming the order. If the email says “like last time,” the previous order used should be shown so that the interpretation can be corrected.
Checking availability needs another boundary: seeing available stock does not mean reserving it. If two orders check the last unit at the same time, the management software must resolve the reservation. I would not let the model promise delivery based on a single inventory reading.
I would also save the email and draft identifiers. If a connection fails after creating the record, a retry should not duplicate it. If a product reference is missing, the case would go to review with the missing detail identified, rather than invented.
n8n lets you configure error workflows and review executions. That helps detect execution failures. An incorrect order accepted by the software needs its own checks: it may not cause any technical error.
A trial that supports a decision
I would take a sample of ordinary, ambiguous and repeated orders. I would measure the manual time needed to leave each order correct and the time with the new process, including review and corrections. I would also count stopped cases and check for any improper action.
If the draft saves transcription but every suggestion needs investigating, it may not be worthwhile. If it works with clear orders, it can remain limited to those. Broader permissions would come after that measurement and after agreeing who handles exceptions.
To start, choose a reversible action, identify the necessary data and write down what the system will do when it cannot find it.
Sources
- n8n: how agents work — Accessed 10 October 2026. The model chooses actions and tools; the workflow may make several calls. This does not guarantee understanding, speed or correctness.
- n8n: error handling — Accessed 10 October 2026. An error workflow can be linked and failed executions reviewed; this does not, by itself, detect every business error.
If you have a task whose inputs and exceptions are identified, you can tell me about it so I can assess which part is worth delegating.