4QueriesFree audit
Case study · AI automation

AI automation for a transfer company in Italy

The owner tracked the day on a paper sheet, while orders arrived from 2 own websites, 7 booking platforms, direct calls and WhatsApp. We built one order hub with a Telegram bot and a phone planner. AI reads the messages. The owner keeps every decision.

The client is anonymous. The demo and all screenshots use fictional data.

11 days
from the first logged sync run to handover
2,027
sync runs logged, 1 failed, and the next run succeeded
182
driver assignments, 169 of them in the phone planner
~$2.64 to ~$0.05
bot cost per message at API prices
The problem

Orders in many places, and the day on paper

What we built

One sheet, and three ways to work with it

One day sheet

All accepted trips are in one list: time, route, source, driver, price.

Telegram bot

A team member sends a booking screenshot, a booking number or plain text. The bot records the order and replies in one line.

Phone planner

A private link opens the day. The owner assigns a driver with one tap, adds notes, creates, moves and removes trips. No app to install.

Google Sheet

The same data for the owner and bookkeeping.

Morning digest

At 06:30 Italian time the chat receives the day’s trips and highlights trips without a driver.

Dispatch draft

The planner proposes a driver for each unassigned trip and flags trips one driver cannot reach in time. The owner accepts with checkboxes. Nothing is assigned automatically.

The demo

See it on fictional data

These frames come from the demo, not from the live system. The demo is a simplified recreation in the browser, and you can click through it yourself.

Demo team chat on fictional data: a booking screenshot, the bot’s one-line reply and a tag showing that AI read the screenshot
Team chat. A screenshot or a line of text becomes an order, and the reply fits on one line.
Demo day planner on fictional data: trips by time with route, source, price and driver, and trips without a driver marked in red
Day planner. One sheet for the day, with trips without a driver in red.
Demo dispatch draft on fictional data: proposed drivers with checkboxes and reasons, plus flags for an unreachable trip and an overlap
Dispatch draft. Proposals with reasons, checkboxes and flags. The owner decides.
Open the demo
Where AI is used, and where it is not

AI reads. Code runs. The owner decides.

AI

Claude Sonnet

Reads chat messages and booking screenshots and turns them into commands. It can run only a short allowlist of order commands.

Plain code

Python

Website and mailbox import, email parsers, sync, journals, planner, digest and dispatch draft. These parts give the same result every time.

Fallback

Slash commands

If the model hits a limit, slash commands work in the chat without AI.

Human

The owner decides

Only the owner confirms an order or assigns a driver.

Results

What the logs show

We counted these figures from the production logs on the server.

FactValue
Sync runs logged2,02718 Oct to 5 Nov 2026
Failed sync runs1 (Google 503, next run succeeded)18 Oct to 5 Nov 2026
Gaps over 30 minutes between runs018 Oct to 5 Nov 2026
Morning digests sent15 of 15 days logged22 Oct to 5 Nov 2026
Booking actions in the journal265 (244 added, 21 removed), including 7 test records19 Oct to 4 Nov 2026
Actions that came from the chat238 of 26519 Oct to 4 Nov 2026
Driver assignments182 (169 in the phone planner, 13 by link)19 Oct to 4 Nov 2026
Platform emails read, read-only17,690stored window, up to 400 days
Bot cost at API prices, general setup$145.26 for 55 messages18 to 24 Oct 2026
Bot cost at API prices, focused setup$2.52 for 47 messages26 Oct to 4 Nov 2026
Automated tests168 passing5 Nov 2026
Python codeabout 4,150 lines5 Nov 2026

The sync log also has 3 skipped starts, because the previous run was still active, and 1 connection error line on 21 Oct. The next run succeeded each time.

Cost per message fell from about $2.64 to about $0.05. The earlier setup also ran on more expensive models and a much larger prompt, so we do not credit one cause alone.

Costs are API-price equivalents. The bot runs on a subscription, so they are not invoices. 25 Oct is left out of the cost rows, because the setup changed that day.

The bot-cost sample is small: 55 and 47 messages.

What we do not claim

No before-and-after numbers

  • We have no “before” measurement, so this page makes no claim about hours saved, more orders, more revenue or fewer mistakes.
  • AI does not plan the day. The dispatch draft is a rule-based algorithm, its travel times are rough estimates, and the owner’s criteria are still open.
Lessons

What we would tell the next client

01

Copying the paper sheet won

The owner rejected the calendar view and asked for the paper sheet layout.

02

Email is a reference, not a source of orders

Platforms send everything, and only the owner’s confirmation makes a trip.

03

An append-only journal keeps the history

A journal that only appends records shows who changed what and when.

04

Review before handover

The code review found three silent data loss risks, and we fixed all three before handover.

05

Less AI was enough

A smaller prompt and a cheaper model were enough for this job.

Questions

About this case

Does AI decide who drives?+

No. The dispatch draft is a rule-based algorithm without AI, and its travel times are rough estimates. Only the owner confirms an order or assigns a driver.

What can the AI do in this system?+

It reads chat messages and booking screenshots and turns them into commands. It can run only a short allowlist of order commands.

What happens if the AI model hits a limit?+

Slash commands work in the chat without AI, and the sync, journals, planner and digest are plain code.

How much time or money did this save?+

We do not know. We have no measurement from before the project, so we make no claim about hours saved, orders, revenue or mistakes.

Is the demo the live system?+

No. The demo is a simplified recreation on fictional data. The chat is a recorded script and no AI runs on that page. The live system and its data are private.

Why is the client not named?+

We publish this case anonymously as a transfer company in Italy. The screenshots come only from the demo, so they show no real people, hotels, addresses or prices.

Do your orders live in chats, emails and paper?

Tell us how your day works now, and we will tell you which parts need AI and which need plain code.

Talk to us