Reference setup The business in this case study is a model business, not a client case. Sorting, reading and model cost were measured on 30 test emails. The time saved and the daily workflow are assumed.
Evening desk in a workshop office with a laptop and copper fittings, the lit workshop behind it

Case study · Skilled trades · Plumbing, heating, air conditioning

AI for tradespeople: enquiries are sorted before you reach your desk.

An assistant reads the emails, moves emergencies to the top and puts a draft reply in front of you for approval. You decide what goes out. This page shows what worked in the test and what did not.

Cockpit · 06:48
Emergency: heating down
Cuxhaven-Döse, please call back
6 h

Enquiries and questions per week: 30 messages at 12 minutes each.

assumed · model business, 8 employees

An evening and a morning at work

A maintenance enquiry arrives after hours, a breakdown the next morning.

  1. 19:12
    The enquiry arrives.

    Gas boiler maintenance in Langen, house from 1994, 130 square metres, phone number included. The assistant classifies the email and picks out the details.

  2. 19:13
    The draft is ready.

    Maintenance follows your fixed template. If the place or the trade is missing, it becomes a question. Everything else comes to you as a site visit.

  3. 19:20
    You approve.

    A quick read, approve, done. Without your approval, no reply goes to a customer.

  4. 06:48
    The breakdown is at the top.

    Heating down, baby in the house, phone number included. Fixed rules pass the email on as an emergency instead of queuing it.

Evening desk in a workshop office with a laptop and copper fittings, the lit workshop behind it
06:48
Cockpit06:48
Emergency: heating down
Cuxhaven-Döse · please call back
Cockpit19:20
Reply approved
The reply to the customer is on its way.
Cockpit19:13
Draft ready for approval
Maintenance from template
Email19:12
Gas boiler maintenance
Langen · built 1994 · 130 m²

Example flow. Sorting, reading and the route of the enquiry were measured. Draft and approval are described after the holiday rentals test setup. The times show the order of events, not a measured response time.

Before and after, kept honestly apart

What was measured in the test is marked in green. What is assumed for the model business is dashed. We will recalculate with your own figures in a conversation.

measured assumed
Overfilled paper calendar with sticky notes and a phone on a wooden table in the evening

Before

6 h / week
Reading emails, asking back, finding appointments, replying
assumed
30 messages
per week: enquiries, appointments, questions and breakdowns mixed
assumed
One list
Emergencies and advertising sit in the same inbox
assumed
Tidy boiler room with a wall-hung unit and a packed tool case in the morning light

After

1.5 h / week
Reading and approving drafts, deciding special cases yourself
assumed
0.6 US cents
Language model cost per message, averaged over 30 test emails with Claude Sonnet
measured
28 of 30
Test emails on the route set beforehand, all four emergencies spotted (Claude Sonnet)
measured

What the AI does. What you decide.

The assistant handles

  • Sort messages: enquiry, appointment, question, breakdown, advertising
  • Pick out details: trade, place, year built, area, preferred date
  • Spot missing required details so that a question can be asked
  • Move emergencies to the top by fixed rules

You keep

  • every approval before a reply goes out
  • price, quantities and technical feasibility
  • binding appointments for site visits and measuring
  • warranty, funding and anything legal

The limit, stated openly

With Claude Sonnet, one of 30 test emails came back with a preferred date nobody had mentioned. The IONOS model took wrong details from a manipulated email in one of 30. That is why nothing goes out without your approval.

measured · 30 self-written test emails, one run each

What is behind it

Four parts, nothing more. Your mailbox stays as it is.

1
Your mailbox

Customers write to your address as before.

2
Your assistant

Runs on its own server, just for your business.

3
Language model

Only asked to classify an email and pick out details.

4
Your cockpit

In the browser and on your phone: processes, approvals, contact person.

The server and the language model run on your own accounts with the provider. Expect around €10 a month for the server, depending on the provider, plus a few euros for the language model. You pay both directly to the provider.

Your cockpit. You see everything that happens.

No black box and no third-party platform you are tied to. Every process is visible, every approval traceable.

Status at a glance Contact person in the chat Push to your phone Daily backup
Cockpit overview: all services fine, availability, backup, disk space and certificates
Cockpit chat with the coordinator assistant, labelled as AI-assisted

Real screenshots from my own cockpit instance. For you, the cockpit shows your modules, such as enquiries and appointments.

What it costs

Starting prices for the entry package. The exact price depends on the number of your standard services, your special rules and the software you already use.

Starting point

Enquiry assistant with cockpit

one-off
from €1,490
per month
from €290
  • Read and sort messages
  • Reply and question drafts for your approval
  • Cockpit with chat and push
  • Support: technical updates and adjustments to your processes
  • Trial run with your real enquiries and an introduction
Add-on modules
Follow-up
Reminder when an offer stays unanswered
on request
Maintenance reminder
Letters at fixed intervals, without any language model
on request

All prices net plus VAT. With my guide, you book the server and language model directly with the provider: around €10 a month for the server, depending on the provider, plus a few euros for the language model.

For comparison: mini-job

A part-time helper on a mini-job basis (€603) costs you as an employer, with flat-rate contributions of around 31 %, about €790 a month. Source: Minijob-Zentrale, contribution rates 2026 (German).

Calculate with your figures

Want the numbers for your business?

Tell me how your enquiries are handled today, and I will work out with you what would be worth it. The initial conversation is 30 minutes and without obligation.

Book an initial conversation

The questions everyone asks

Does this replace my trade software?

No. The assistant takes care of what arrives by email. If you already use software for offers and invoices, we look at whether and how it can be connected in the initial conversation. None was connected in the test.

Does the AI write my offers?

It prepares. Standard services such as maintenance follow your template and your price list. Everything else is planned as a site visit. Prices, quantities and commitments come from you, withdrawal notices and contract clauses are fixed blocks.

Can I rely on it in an emergency?

Four out of four emergencies were spotted in the test, which is a good sign and not proof. The system does not replace an emergency service or an emergency number. If there is a smell of gas, your customer has to call the network operator, not write an email.

Where is my customers’ data stored?

On your own server. You set up the accounts with the host and the model provider yourself; they run in your name and under your contracts. The language model only receives the email text it needs to classify and read the message.

Which language model is used?

The default is IONOS, with data centres in Germany. On request, we can also connect OpenAI, Anthropic or Google. In the test, Claude Sonnet and the IONOS model were close: 28 and 27 of 30 emails on the right route.

Is this a real client?

No. The business is a model business with eight employees, and the 30 test emails were written for the test. Sorting, reading, routes and model cost were measured. Setup effort and long-term operation were not. The time saved is an assumption that we recalculate with your figures.

Let us work it out with your figures.

30 minutes, no obligation. You tell me how your enquiries are handled today, I tell you what is worth it and what is not.

Some images AI-generated