How AI learns from your team’s corrections, and what stops it learning the wrong thing

Every month, a freight forwarder in Surabaya sends you the same two-page invoice. Every month, the AI capture in your AP system reads the number in the top right corner, which is the vendor’s delivery note number, as the invoice number. And every month, an AP clerk deletes it and types in the real one from the bottom of the first page. By December they have made the same correction twelve times, and the system is no wiser than it was in January.

Most AI capture tools promise to fix this: correct a field once and it gets it right next time. The promise is real, but how the learning works matters as much as whether it happens. A system that learns from every edit also learns typos and one-off exceptions, and in finance, one of those can be a bank account number.

This guide explains how AI learns from corrections, what it should and shouldn’t learn from, a worked example of a lesson being learned and then unlearned, and how to tell whether the AI reading your invoices and receipts is learning well.

What does it mean for AI to learn from corrections?

When finance software says its AI “learns”, it usually means one of three things:

1. Retraining a shared model. Corrections from many customers are collected and used to retrain the vendor’s model. Improvements arrive slowly, in releases, and your documents help improve a model other companies use.
2. Templates and rules set up by hand. Someone tells the system where each field sits on each vendor’s invoice, or which vendor goes to which category. It’s precise, but someone has to build and maintain it, and a template breaks when the vendor changes its layout.
3. Lessons learned per company. The system turns your team’s corrections into short rules in plain words, such as “on this vendor’s invoices, the invoice number is the one at the bottom of page one, not the delivery note number”. The AI gets your company’s rules each time it reads one of your documents. Nothing is retrained, and the rules are used for your company only.

The third adapts as fast as your team corrects, without anyone writing templates. It also carries the most risk if it’s built carelessly, because whatever your team does becomes the AI’s behaviour. The rest of this guide is about getting that part right. The same questions apply to AI agents, which learn from the runs your team reviews, but here the focus is AI that reads documents.

What a careful system learns from, and what it ignores

• Decisions, not keystrokes. Learn from the value a person submits or approves, not from every edit in a draft. A typo made and fixed before submitting should teach nothing.
• One field at a time. A correction is a specific field: the invoice number, the total, the category. That keeps each lesson small enough to check.
• The right scope. Most lessons belong to one vendor or merchant. A few apply to every document, such as reading dates as day, month and year.
• Evidence, not one example. A single correction might be a one-off. A lesson should only apply once the same correction has been made more than once.
• More evidence for money. A wrongly learned description is an annoyance; a wrongly learned bank account is a payment to the wrong party. Amounts, tax, currency and bank details should need more evidence, including a correction made by an approver.
• A way to unlearn. When people keep changing what a lesson filled in, the lesson should stop being used.
• Counted, not guessed. Ask a language model how sure it is and you get a number that sounds precise but isn’t well calibrated. Counting the corrections that support a lesson and the ones that contradict it gives a threshold you can rely on.
• Your company only. Your corrections should improve your results, not another company’s.

A worked example: one lesson, learned and unlearned

Here is the freight forwarder’s invoice number over six months. Say, as an example, that a lesson applies once two corrections support it, and retires when the evidence against it catches up with the evidence for it. When a lesson fills in a field and the team keeps the value, that counts in its favour; when they change it, that counts against it.

Invoice	What happened	For	Against
January	AP replaces the delivery note number with the invoice number	1	0
February	The same correction again, so the lesson now applies	2	0
March	The lesson picks the right number, and nobody changes it	3	0
April	The vendor changes its layout, and AP corrects what the lesson picked	3	1
May	The same correction	3	2
June	The same again, so the lesson is retired	3	3

Three things are worth noticing:

• March is the point. Nobody touched the invoice number. Repeat that across every vendor with a quirk of its own and the typing saved adds up.
• One odd invoice doesn’t undo a working rule. April’s correction counted against the lesson but didn’t retire it, because the lesson had more history in its favour. A rule with a track record should take more than one exception to overturn.
• The vendor’s change was absorbed. By June the evidence had caught up, the lesson stopped being used and the AI read the new layout fresh. If the team keeps correcting it the same way, a new lesson starts from zero.

Why AI learning goes wrong in practice

• Learning from drafts. Every edit is treated as a lesson, so a typo corrected two seconds later becomes a rule.
• Learning from one example. One unusual invoice, such as a credit note or a one-off service, becomes the rule for every invoice from that vendor.
• The wrong scope. A rule learned on one vendor’s invoices is applied to everyone’s, or one company’s corrections change another company’s results.
• Money fields treated like descriptions. A learned total or bank account looks exactly as confident as a learned description.
• No way to unlearn. The vendor changes its layout, the old rule keeps firing and the team corrects the AI forever.
• Corrections made outside the system. The clerk fixes the invoice number in the accounting system or a spreadsheet instead of in the AP record, so the AI never sees it.

AI capture in Indonesia

A few local habits make learning per vendor especially useful, and show where it needs care:

• Two languages on one document. Invoices often mix Indonesian and English labels, such as “Jumlah” next to “Total” or “Jatuh Tempo” next to “Due date”, and receipts from small merchants may be handwritten. Learning per vendor copes with each vendor’s habits instead of needing a template per language.
• The same vendor, written three ways. “PT Sinar Logistik”, “PT. Sinar Logistik Indonesia” and “SINAR LOGISTIK” are one company. Keying lessons on the vendor’s tax number (NPWP) where the document shows it keeps each lesson attached to the right vendor, however the name is written.
• Dates. 03/04/2026 means 3 April in Indonesia. If an AI reads it as 4 March, that’s a correction worth learning once for the whole company, not vendor by vendor.
• DPP, PPN and the total. Many invoices show the tax base (DPP), the VAT (PPN) and the total separately, and some also show tax withheld. Which figure is “the total” depends on the vendor’s layout. Lessons per vendor help, but these are money fields: they should need more evidence, and approvers should still check them.
• Personal data. Receipts and invoices can carry personal data, such as an employee’s name or phone number, which UU PDP protects. Ask whether your documents and corrections are used to improve results for other companies.

Signs your AI capture isn’t learning well

You probably need to look more closely if any of these sound familiar:

• Your team corrects the same field for the same vendor every month.
• Nobody can tell you what the system learns from, or when a correction becomes a rule.
• A correction for one vendor has changed results for another.
• The system shows a confidence percentage, and nobody knows what 90% means.
• Amounts and bank details are filled in as readily as descriptions.
• Corrections are made in the accounting system or a spreadsheet, where the AI can’t see them.
• Your contract doesn’t say whether your documents are used to improve results for other customers.

An AI that learns from every edit learns your typos. One that learns from your decisions learns how your business works.

How to get the most from AI that learns

1. Correct in the record, before it moves on. Fix the field in the invoice or claim itself before it’s submitted or approved, not in a comment, an email or the accounting system afterwards.
2. Let approvers correct, not just approve. Corrections made at approval carry weight, especially for amounts and bank details, so an approver who spots a wrong field should fix it rather than wave it through.
3. Expect a learning curve for each vendor. The first invoices from each vendor need the most attention. Watch your largest vendors closely in the first months.
4. Keep people on the money. Learning reduces typing, not responsibility. Approvers still check amounts, tax and bank details, and a change to a vendor’s bank account should still need its own approval.
5. Look into the fields that keep changing. If a field is corrected again and again for the same vendor, find out why: a layout change, a vendor error or a rule your team applies inconsistently.

---

How Procapy helps

ProcAI reads invoices and receipts, in Indonesian and English, for Proca to Pay and Capy Expense, with no templates to set up. When an invoice or expense claim is submitted or approved, each field ProcAI captured is compared with its final value, and a field your team changed becomes evidence for a lesson: a short rule in plain words, either for that vendor or merchant, recognised by its tax number where the document shows one, or for your whole organisation. Edits made and undone in a draft teach nothing.

A lesson is only used once more than one correction supports it, and amounts, tax, currency and bank details need more evidence, including a correction made at approval. When a lesson fills in a field and your team keeps the value, that counts in its favour; when they change it, that counts against it, and once the evidence against catches up, the lesson is retired. Lessons belong to your organisation and are never shared with another. On every invoice and claim, each captured field shows whether it is still what ProcAI read or a person has changed it.

Proca Agents learn from the runs your team reviews too, with each correction applied to the next run of that agent and only that agent. Book a walkthrough and bring a few invoices from your most awkward vendors.