A measured case
How much time does invoicing automation actually save a small business?
In the small business where my agent has been running for six weeks: 170 to 180 delivery notes a month, 13 minutes of manual entry each — about 38 hours a month, close to a full working week. The agent prepares the draft invoices overnight; the reading and approval that remain take a fraction of that time. No incident reported in six weeks of service.
The maths, and where the numbers come from
These aren't market averages. They are the figures of one real business, where the agent has run every day for six weeks.
| What is measured | The figure |
|---|---|
| Delivery notes handled | 170 to 180 per month |
| Manual entry time, per note | 13 minutes |
| Total, per month | about 38 hours |
| Over a year | about 455 hours |
| Already handed back since launch | around fifty hours |
| Incidents reported by the client | none in six weeks |
Thirty-eight hours a month is close to a full working week. Not a week spent deciding, selling or solving a problem: a week spent copying lines from a file into a piece of software.
What the agent does, exactly
Nothing magic happens, and that's deliberate. Each step replaces a gesture someone used to make by hand.
- It polls the mailbox. The delivery notes arrive as Excel attachments, several per file, for different customers.
- It reads the file. It recognises the lines, the quantities, the references, and which customer each belongs to.
- It signs in to the business software — here Erplain — with credentials the business entered itself, on its own machine. Two-factor authentication included. I never have access to them.
- It creates the draft invoices, one per customer, through the software's own interface — exactly as a person would.
- It stops there. Nothing is approved, nothing is sent.
One technical detail that matters more than it sounds
The agent calculates no amount itself. It doesn't add up, it doesn't apply VAT, it doesn't decide a discount. It enters the lines into the business software, and it is the software that computes prices, taxes and addresses — with its own rules, the ones the business already configured.
That is the answer to the most reasonable fear I'm given: "what if it gets an amount wrong?" It can't get an amount wrong, because it works none of them out.
Six weeks without incident, but not without work
The client has reported no incident since the agent went live: no wrong invoice, no breakdown that forced anyone back to manual entry.
That doesn't mean I haven't touched it. There have been improvements along the way — cases nobody had foreseen when we wrote the scope, better ways of presenting information, details that only surface once the tool runs daily on real documents.
What it hasn't changed
It's worth being honest about the half of the work that stays human, otherwise the 38-hour figure becomes a lie.
- The review stays whole. Someone opens the drafts and approves them. It's quick, but it isn't zero.
- Exceptions stay human. An unusual note, a customer with a special case, a discount agreed on the phone: the agent doesn't guess. It flags the line and stops there.
- Sending stays a decision. No invoice leaves without someone clicking.
When it becomes worth it for you
The figure that decides isn't the number of documents, it's the total time. A twenty-line delivery note is not a two-line one.
Below roughly ten hours of data entry a month, the installation takes too long to pay for itself, and I'll advise you against it. I say so at the first conversation, before any quote — it has happened.
The case described here sits at 38 hours a month, well above that threshold. That's what makes it a good case rather than a staged demonstration.
- Run the numbers yourself. The calculator asks for three figures you already know: documents per month, minutes per document, loaded hourly cost. That page is in French.
- Then send me three real documents. I'll tell you whether the format is regular enough to be read automatically — and if the answer is no, I'll tell you that too.