Supplier invoices land in a shared inbox as PDFs, delivery notes arrive as phone photos in WhatsApp or Telegram, and completion certificates from contractors come as crooked scans with a stamp across the totals. Someone in finance opens each one, retypes the supplier, the amounts and the line items into the accounting system, checks them against the purchase order or contract, and only then posts. Those hours could go to reconciliations, VAT and cash-flow conversations with the owner. This article shows how AI-assisted document capture works in practice: which documents it reads reliably, which checks must run after extraction, where the line of responsibility sits, and how to roll it out without putting your books at risk.
What leaves manual work and what does not
Start with what AI does not replace. Structured e-invoices are already data. In Saudi Arabia, invoices issued through ZATCA's FATOORA system are machine-readable by design; the UAE is introducing its own e-invoicing framework; across Europe, Peppol networks carry invoices as structured files. If those documents already flow into your ledger automatically, there is nothing to "read" — they need to be accepted, matched and approved.
Manual keying survives wherever a document arrives as an image or an unstructured PDF:
- supplier invoices and pro-formas sent by email;
- completion certificates and service sign-offs, often signed by hand and scanned;
- delivery notes and packing lists that travel with the goods;
- invoices from overseas suppliers in Arabic, Turkish, Chinese or Russian;
- supplier statements used for month-end reconciliation;
- receipts and expense claims from staff.
These are the documents AI turns into data. The decision to post, the choice of nominal code in unusual cases, and the approval to pay stay with your finance team.
The journey of one document, from attachment to draft
A working setup has five stages. None of them writes anything final to the ledger.
- Collect. An agent picks up attachments from a dedicated mailbox (for example, invoices@yourcompany) and from a work chat or bot where managers forward photos. Everything lands in a single queue tagged with source and sender.
- Classify. The model decides what each file is: invoice, credit note, delivery note, statement, contract, or "not a document" (marketing emails, replies). A 12-page PDF that bundles three invoices is split into three.
- Extract. Header fields and line items are pulled out — see the table below.
- Check. Arithmetic, contract or PO match, duplicate search, supplier master-data match.
- Draft. An unposted bill or a row in a review register is created in your accounting system, with the original attached and a list of issues.
If your books live in 1C, the draft is created through the same integration route as other connections; the options are covered in our piece on connecting an AI agent to 1C. Xero, QuickBooks, Zoho Books and most ERPs accept draft bills through their APIs in a similar way.
Which fields to extract and how to check them
Agree the field list with your finance lead up front. Extra fields slow review down; missing ones bring manual keying back.
| Field | Where it comes from | How it is checked |
|---|---|---|
| Supplier name, VAT/TRN or company number | Header, letterhead, stamp | Matched against the supplier record; an unknown number always goes to manual review |
| Invoice number and date | Header | Duplicate search on "supplier + number"; date not in the future or in a closed period |
| PO or contract reference | "PO no." or "as per agreement" line | Exists in the system, still valid, scope matches |
| Lines: description, quantity, unit price | Line table | Quantity × price = line total; mapped to your item list |
| Net, VAT, gross | Totals block | Lines add up to the total; VAT rate fits the supplier's status |
| Bank details | Footer | Same as the stored IBAN; any change is a red flag |
| Signature and stamp | Image | Presence only — AI does not verify authenticity |
A word on bank details. Changing the account number on a "supplier invoice" is a well-known fraud pattern — in the UK it is usually called mandate fraud. The rule is simple: if the bank details differ from the stored ones, the bill does not enter the payment run until a person confirms the change by calling the supplier on a number you already hold, not one printed on the new invoice.
The exceptions queue: how not to drown in review
If the accountant has to reopen every document and compare it with the draft, there is little saving. The point is to split the flow in two.
- Fast lane. Known supplier, valid contract or PO, all checks passed, amount within the usual range. The accountant scans the list and approves in a batch, spot-checking originals.
- Exceptions queue. Everything else, each with a reason: "line 4 total unreadable", "contract expired 30 Sept", "number matches the invoice of 12 Oct", "bank details changed".
The model must be able to say "not sure". For each field the system records whether it was found explicitly or inferred from context, and anything inferred goes to exceptions. That matters more than any accuracy figure in a vendor's brochure: an error the system flags costs a minute, while one it passes confidently turns up at the quarter-end reconciliation.
Here is what an exception entry can look like:
Invoice 87, 21 Oct, supplier "…", 3 lines.
Checks: TRN matches ✓, contract 14 valid ✓, arithmetic ✓.
Issue: the contract rate for "Equipment servicing" is lower than on the invoice. The difference is shown on the card. Original: page 1.
Action: approve / return to supplier / correct manually.
Who does what once it is running
| Stage | What AI does | What stays with finance |
|---|---|---|
| Collection | Pulls attachments from email and chat, removes duplicate files | Tell suppliers where to send documents |
| Reading | Reads text, tables, handwritten notes, several languages | Handle the documents that could not be read |
| Contract/PO match | Finds the agreement, compares prices, dates and scope | Decide what to do about a mismatch |
| Bookkeeping | Creates a draft, suggests a nominal code from similar past bills | Review and post; choose codes for unusual items |
| Payment | Builds the list of bills due, with dates | Approve payment, confirm changed bank details |
| Storage | Links the original to the record, finds documents on request | Keep records for the period your tax authority requires |
On that last row: retention rules vary. In the UK, for example, HMRC generally expects VAT records to be kept for six years (HMRC guidance on VAT record keeping); check the rule that applies in your jurisdiction.
A useful side effect is search. Once every document is data with the original attached, "show me every invoice from this contractor for the last six months" becomes a query, not a dig through a "Scans 2026" folder. How AI answers questions over your own documents is explained in our article on RAG systems for business.
A four-step rollout
- Sample. Gather 50–100 real documents from recent months — different suppliers, good and bad scans. Test reading on these, not on demo files.
- Shadow mode. For 2–4 weeks the AI processes documents alongside your team but creates nothing in the ledger. Differences are reviewed: where the AI was wrong, where the person was, which checks are missing.
- Drafts. The AI creates unposted bills; the accountant approves. The fast lane starts narrow: regular suppliers only.
- Widen. As evidence builds up, more document types join the fast lane. Posting and payment remain with people.
Before you start, check:
- is there an up-to-date supplier master file with tax numbers and bank details — without it, the checks have nothing to compare against;
- where contracts and POs are stored, and can they be found by number;
- who owns the exceptions queue and how quickly it is cleared;
- where documents will be processed and who has access — invoices carry commercial data about your suppliers, so data-protection rules apply;
- which documents are kept out of automation entirely (payroll and HR files, for example).
How we build an agent for this kind of process and connect it to your systems is described on the AI agents page. More on internal processes is in the Operations section.
FAQ
How well does AI read poor scans and phone photos?
It depends on quality: a clean PDF is read almost every time; an angled photo with glare, less so. That is why accuracy is measured on your own documents in shadow mode, not on someone else's numbers. Anything read with low confidence should go to the exceptions queue, not the ledger.
Can documents be posted automatically from day one?
We do not recommend it. Posting affects VAT and reporting, and your finance team is accountable for it. A sensible boundary is a draft that a person approves — in batches for routine bills.
Do structured e-invoices need AI reading?
No. E-invoices from systems like ZATCA's FATOORA or Peppol already arrive as data and can be imported directly. AI is useful for everything that arrives as an image or PDF: supplier invoices, delivery notes, completion certificates, overseas invoices, receipts.
Can it handle documents in Arabic and other languages?
Current models read invoices in English, Arabic, Russian, Uzbek and many other languages. Item descriptions are still mapped to your own item list, and new descriptions go to review rather than being guessed.
What does it cost?
Cost depends on document volume, the number of sources and which accounting system the drafts go into. The first conversation is for looking at your documents and working it out. You can get in touch via the contact page.