TL;DR Automating PDF order entry means having an AI read the orders that arrive by email and fill in the file your management software imports, with a person checking before the import. In a real case, a metalworking manufacturer in central Italy, the time per attachment fell from 3-4 minutes to about one by the administrator's estimate, and data-entry errors went down. Over six months: 219 orders, 300 attachments, about $10 of AI usage. Building it took one month of work, not full-time.
The AI was already there. The process was missing.
Automating PDF order entry means handing an AI the transcription of the orders customers send by email, and leaving a person only the check and the import into the management software. In a real case, a metalworking manufacturer in central Italy, the time per attachment fell from 3-4 minutes to about one, by the administrator's estimate, and data-entry errors went down. From 1 April to 28 September 2026 the system handled 219 orders on about $10 of AI usage.
I designed and built this system for the client, and building systems like it is my trade, so I write about it as an interested party. For the same reason, the figures that follow come from the real project rather than from an example. The guide covers what changes, what it costs and how long it takes to build, and closes with a checklist to see whether your case resembles it.
The technical version, with the workflow and the choices behind it, is on the page about how I built the n8n workflow for PDF orders.
Using ChatGPT by hand is not enough
Before, the company's administrator downloaded each PDF order, uploaded it to ChatGPT, had it generate an Excel file and checked it, because something was often wrong. Then they entered the order into the management software, where they could make mistakes of their own. All of this took 3-4 minutes per attachment, by their own estimate.
If you do the same today, AI is already part of your process. The trouble is that every time it has to guess what to produce: which columns, in what order, in which units. Every customer writes orders their own way, and ChatGPT answers its own way each time.
How it works, in practice
- The customer emails the order, as they always have, with the PDF attached.
- The system discards emails that are not orders on its own, without using AI. In the reference case 504 of 723 emails stopped here, at no cost.
- Before analysing a PDF the system estimates what it will cost. If it exceeds the threshold the company set, it stops and flags it, and the administrator decides whether to authorise it. No analysis over the threshold runs without someone having approved it.
- The AI reads the PDF and fills in the file in the format the management software imports, converting units of measure with the company's table.
- The file lands in Google Drive, sorted by customer, year and month, and the administrator gets a Telegram message. They download the file, check it, import it and confirm with a button.
The rules (senders to ignore, keywords that identify an order, the spending threshold, the unit table) are files in Google Drive that the company edits alone, without calling the developer.
Process first, then the AI
It is easy to assume the hard part is connecting an AI to the inbox. In the reference case the most important part was deciding, with the company, what an order looks like: a single format, the one the management software imports, and a table that says how to handle each unit of measure customers use. The month of building included this digitisation and standardisation work.
Only then did the AI enter the process, with a precise job: fill in that format, following instructions the company wrote.
The result, according to the administrator: in six months no order has needed correcting before import. Errors went down on both sides, the ones ChatGPT used to make and the ones made by whoever copied orders into the management software by hand, because that transcription no longer happens.
What changed, in numbers
In the reference case, from 1 April to 28 September 2026:
| Before | After |
|---|---|
| 3-4 minutes per attachment (estimate) | about 1 minute per attachment (estimate) |
| ChatGPT by hand, a different format each time | AI filling in the management software's format |
| Frequent errors to correct | No corrections in six months (according to the administrator) |
| - | 723 emails tracked, 219 orders, 300 attachments |
| - | about $10 of AI usage, about $0.03 per attachment |
Across 300 attachments, going from 3-4 minutes to about one means 10-15 fewer hours of work in six months. The times, before and after, are the administrator's estimates: the system records emails and costs, not time spent in front of the management software. The other figures in the table come from the system's own log.
You don't have to change what you already have
A solution like this should adapt to the company, not ask it to reinvent itself. In the reference case:
- the management software stayed Danea Easyfatt, even though it cannot be driven through an API: the system produces the file Easyfatt already imports, and a person does the last step;
- email, archive and configuration live in Google Workspace, which the company already licensed;
- notifications arrive on Telegram, which costs nothing.
The operator did not have to learn new software, and their job moved from copying orders to checking them.
What a person still does
The system removes the transcription. In the reference case three tasks remain with the administrator:
- checking and importing each file into the management software, then confirming on Telegram. Until the management software has an API, this step stays manual;
- deciding whether to authorise emails that exceed the spending threshold: 13 out of 219 in six months, all authorised;
- keeping the rules up to date when a new customer, an unfamiliar unit of measure or a sender to ignore turns up.
There is also a limit worth knowing. An email is recognised as an order by keywords in its subject: if a customer writes a different subject, the order is discarded. It doesn't go unnoticed, because every email gets a label visible in the mailbox, but someone has to look.
Checklist: does it make sense for your business?
Answer yes or no to these four conditions:
Orders arrive as PDF email attachments, from several customers, each with their own format.
Your management software imports a file (Excel or CSV), even if it has no API.
You receive at least one order a day, regularly.
Whoever enters orders today has other work to do, and every order copied by hand is time taken from the rest.
No yes: the volume or format of your orders doesn't yet justify a dedicated system.
One or two yes: it can be done, but it has to be designed around how your case differs from this one.
Three or four yes: your process is very close to the reference case.
What it costs and how long it takes
In the reference case, building the system took one month of work, not full-time, including standardising the process with the company.
Running it cost about $10 of AI usage over six months, roughly $0.03 per attachment, with OpenAI's GPT-5.4. On top of that, about $10 was spent before going live, on development and testing. The workflow runs on n8n, self-hosted on an Oracle server on the free tier, and there are no other recurring costs.
The cost of building it, on the other hand, depends on many factors: how many customers and formats there are, what the management software is like, how orderly the starting process already is. That is why it has to be assessed case by case.