Sorting enquiries in a trades office: what can be automated

Two things can be automated in a trades office: collecting enquiries from a form, a mailbox and phone notes into one shared list, and adding a suggestion to each about what it is and how urgent it seems. Whether the business accepts the enquiry, what it commits to and what it charges is decided by the office clerk. There is a practical reason for this: an email is untrusted text, and a language model cannot reliably tell content from instruction in such a text.

What pre-sorting means here

A workflow is a fixed sequence of steps that an event starts, for example someone submitting a form. What that means on a small scale is covered in What process automation is in a small business. There are three ways in for enquiries.

For the form, the n8n Form Trigger generates the form page itself, according to the manufacturer, and starts the workflow when someone submits the form. For the mailbox, the Email Trigger (IMAP) receives mail from a mail server and can mark it as read. It downloads attachments only on request, which according to the manufacturer increases the processing effort. For the phone note no separate technology is needed: my suggestion is that the office clerk types the note into the same form, so that all routes end in the same list.

The list can be a table with the columns received, channel, sender, subject, urgency, draft query and status. One row per enquiry.

What can be automated and what stays in the office

Step	Automatic	Stays in the office
Capture	Entry in the list	Typing in the phone note
Cleaning	A fixed step removes invisible characters and opens neither attachment nor link	nothing
Completeness	Fixed rule: is a name, call-back number or site address missing?	Approve a query about the gaps
Classifying	AI suggests subject and urgency	Check the suggestion, change it
Follow-up query	AI creates a draft	Read, edit, send it yourself
Accepting, appointment, price, commitment	nothing	everything

1. Capture enquiry Form, email or note end up in one list
2. Clean the text Fixed rules, no attachment or link is opened
3. AI suggests Subject, urgency, draft for the follow-up query AI prepares
4. Office decides Check, edit, approve the draft and send it yourself

Text version of the graphic

• 1. Capture enquiry: Form, email or note end up in one list
• 2. Clean the text: Fixed rules, no attachment or link is opened
• 3. AI suggests: Subject, urgency, draft for the follow-up query (the AI prepares)
• 4. Office decides: Check, edit, approve the draft and send it yourself

The classifying step is the part with AI. The n8n Text Classifier sorts a text into categories that you define with a name and a description. It has a default worth knowing about: if it finds no matching category, it discards the item, as the documentation says. An enquiry that fits no category then disappears without a message. The documentation names an alternative, a separate output called “Other”. The unclear cases end up there, and those are exactly the ones that belong in the list.

What is a small matter for the workflow can be an emergency on the phone. A burst pipe in the cellar is not an enquiry that waits for a table. Why certain processes are not automated is covered in When not to automate.

Why a person approves the draft

In “Generative KI-Modelle” the BSI describes a behaviour of language models that applies to every mailbox. They “interpret all inputs in principle in the same way and do not distinguish between instructions and other texts” (translated from the German here and below). Letting such a model evaluate emails therefore opens a door. According to the BSI, attackers can hide instructions “on websites, in emails or in documents that the LLM evaluates”. As possible consequences the BSI names, among others, sending and deleting emails if the application has corresponding access and acts on the model’s outputs by itself. It also names the leaking of information through externally loaded images in the response.

According to the BSI, no protection exists that removes the problem. In its baseline protection of 2026 the BSI says it had already pointed out in 2023 that “no reliable countermeasure exists that completely eliminates the problem”. That baseline protection deals with documents uploaded to AI chats, but it calls the document “representative of an external data source”. In my view a customer email is such a source. The BSI also stresses that its measures do not replace an individual security assessment.

From this I derive for the office what the BSI lists as measures. The rights of an AI application are to be limited to the necessary minimum, and critical actions can depend on “explicit consent of the users, e.g. via a confirmation button”. Outputs with an external effect are to be checked before they are used further. That gives four rules:

• The AI step receives only the cleaned text and no tool. It cannot send anything, delete anything or open a link.
• Before the AI step, a fixed step cleans the text. The BSI describes this in its baseline protection as the first basic measure: invisible characters and forged role separators are replaced or removed. Attacks that argue with clean, inconspicuous text “pass through unchanged”.
• Its output ends up in the list or in a draft. The Gmail node in n8n offers the operation “Create a draft” for this, separate from sending.
• Sending is done by hand. If you use an AI agent with tools, n8n can require human approval before a tool is called. The documentation explicitly names sending external communications as an example. The pattern behind this is described in AI in a workflow: one AI step, fixed rules, human approval.

This limits the damage of an attack but does not prevent it. In the unfavourable case, a wrong suggestion sits in the list. Whether this pattern is enough in practice, I have not measured.

Personal data and Article 28 GDPR

An enquiry contains a name, a telephone number and usually the address of the property. These are personal data, under Article 4(1) GDPR “any information relating to an identified or identifiable natural person” (the Official Journal wording in German: “alle Informationen, die sich auf eine identifizierte oder identifizierbare natürliche Person beziehen”). If a service provider processes them for you, under Article 4(8) it is a processor if it processes them “on behalf of the controller”.

Article 28 sets requirements for this relationship. Under paragraph 1 the controller uses only processors that provide “sufficient guarantees” of suitable technical and organisational measures. Paragraph 3 requires a contract that sets out the subject matter, duration, nature and purpose of the processing and provides, among other things, that the processor processes the data only on documented instructions. Paragraph 9 requires written form, which can also be electronic. Under paragraph 2 the processor needs authorisation before engaging another processor.

For your workflow this means: the mailbox provider, the operator of the workflow and the AI provider are each to be looked at separately. Whether a processor is involved in your case and what the contract must contain is for your data protection officer to clarify. Such a workflow can be operated in line with the GDPR if these questions are answered beforehand. That is not a guarantee.

What it brings in time: no measured values, only a calculation with assumptions

I found no measured figures for how many enquiries small businesses receive or for how long sorting takes. How to collect your own values is shown in Measure time savings instead of promising them.

The following calculation is an example with invented assumptions, not measured. It assumes 40 enquiries a month. Today each one costs 4 minutes: reading, classifying and entering it in the list. With pre-sorting each one costs 2.5 minutes: checking and correcting the suggestion, reading the draft of the follow-up query and sending it. That is a saving of a good third per enquiry.

Today that is 40 × 4 minutes = 160 minutes a month, with pre-sorting 40 × 2.5 minutes = 100 minutes. The difference is 60 minutes, so 1 hour a month. With 100 enquiries a month it would be 400 against 250 minutes, a difference of 2.5 hours. Not included are set-up, maintenance and the cost of the tools. With small volumes the saving is small, and whether it outweighs the upkeep depends on your figures. Put in your own values; the assumptions above do not apply to your business automatically.

What you can do now

1. For four weeks, keep a tally of how many enquiries arrive by form, email and phone, and time ten enquiries each.
2. Decide which details every enquiry must contain, and build the form first, without AI. A shared list alone already brings order.
3. Determine who decides which kind of enquiry and who sends drafts, including during holidays and sickness.
4. Before using an AI service, ask your data protection officer which providers are processors and whether the contracts are in place.