After processing thousands of invoices every week, you start to see patterns. The same document processing challenges come up again and again.
Most tools handle the straightforward cases fine. A clean PDF from an enterprise vendor with consistent formatting? Easy. But finance teams don’t live in that world. They live in a world of marketplace settlement reports, cross-border invoices in three languages, and vendors who quietly change their bank details.
Here are five real challenges we see every week, and how our AI agents handle them.
1. Marketplace settlement reports that look nothing like invoices
If you sell on Amazon, or you’re the accountant for a business that does, you know this one intimately.
Amazon doesn’t send a standard invoice. They send a settlement report. It’s a completely different document structure. Fees netted against payouts. Advertising costs bundled in. Refunds deducted inline. Currency conversions applied across multiple lines. FBA storage fees, referral fees, promotional discounts. All in one document that technically isn’t an invoice at all.
A bookkeeper looking at this for the first time has to reverse-engineer which lines are revenue, which are fees, which are refunds, and how to map all of that to their chart of accounts. For one of our customers, an e-commerce brand selling across three European marketplaces, this was taking half a day every week. Just Amazon.
How our AI agent handles it: The agent understands the settlement report structure. It knows that a line labelled ‘FBA fees’ maps to a specific GL code in your system, because it’s cross-referenced how you’ve categorised it before. It breaks the report apart, maps each component to the right account, handles the currency conversion, and posts the entries. The bookkeeper reviews a clean summary. That half day is gone.
The key insight: a settlement report requires fundamentally different extraction logic than a standard invoice. AI invoice processing tools that are built only for invoices will either reject it or extract garbage. You need an agent that recognises the document type and adapts its approach.
2. Same supplier, different entity, different VAT treatment
A French supplier sends a facture. A German supplier sends a Rechnung. A Dutch supplier sends a factuur. Same expense type across all three. Completely different VAT treatment depending on which entity in your group received it.
This is where most automated invoice processing falls apart. It’s not enough to extract the data correctly. You have to know which compliance rules apply to this specific combination of supplier country, receiving entity, and expense type. And those rules change. A reverse charge that applied last quarter might not apply after a regulatory update.
We work with a company that operates across Germany, the US, France, and Japan. Four entities, four sets of tax rules, invoices arriving in four languages. Every invoice that comes in needs to be routed to the right entity and have the correct VAT treatment applied. Before Quid, someone on their team maintained a spreadsheet of rules. When they went on holiday, invoices backed up for weeks.
How our AI agent handles it: The agent checks which entity the invoice belongs to, looks up the supplier’s country and VAT registration status, and applies the correct treatment based on your historical handling of that combination. If the combination is new - a French supplier invoicing your Japanese entity for the first time - it flags it. If the rules have changed since last quarter, it flags that too. The institutional knowledge isn’t locked in someone’s head anymore. It’s in the system, applied consistently on every invoice, across every entity.
3. Vendors who quietly change their bank details
This one is surprisingly common and surprisingly dangerous.
A vendor you’ve been paying for two years sends their monthly invoice. Everything looks normal. Same amounts, same line items, same format. But somewhere in the payment details section, the bank account number is different.
Maybe they genuinely switched banks. Maybe someone in their accounts team updated the template. Or maybe it’s a business email compromise attack - one of the most common forms of invoice fraud - where a bad actor intercepts a legitimate invoice and swaps the bank details.
Either way, if your AP team is processing 200 invoices a week, the odds of a human catching anomalies like a bank detail change on invoice number 147 are low. Especially on a Monday morning.
How our AI agent handles it: The agent compares the payment details on every invoice against the vendor’s historical record in your master data. If the bank details don’t match what we’ve seen before, the invoice gets flagged immediately. It doesn’t quietly process it. It doesn’t assume the change is legitimate. It stops and asks for confirmation before anything moves to payment.
This is one of those cases where cross-referencing historical master data isn’t just about efficiency. It’s a financial control. The same check that speeds up routine invoices also catches the one that could cost you real money.
4. One PDF with multiple documents buried inside
A logistics company sends you a single PDF every month. You open it expecting one invoice. What you actually get: page 1 is an invoice for this month’s shipments. Page 4 is a credit note for a previous overcharge. Page 7 is a pro-forma estimate for next month that absolutely should not be booked.
Your experienced bookkeeper knows this because they’ve been dealing with this vendor for a year. They know to split the document, book the invoice, book the credit note with opposite sign, and ignore the pro-forma. A new team member doesn’t know any of this. And most OCR-based invoice processing tools certainly don’t. They extract the whole PDF as one document and produce a mess of numbers that don’t reconcile.
How our AI agent handles it: The agent detects document boundaries within a single file. It recognises that page 1 is an invoice (book it), page 4 is a credit note (book it with the correct sign), and page 7 is a pro-forma (flag it, don’t book it). Each section gets processed according to its own rules and document type.
This isn’t magic. It’s pattern recognition combined with your historical handling of that vendor. The first time we encounter a new vendor’s bundled PDF, we learn the structure. Every subsequent month, it processes automatically. The bookkeeper who used to spend 20 minutes splitting and categorising each monthly PDF from this vendor now spends zero.
5. Recurring invoices where amounts fluctuate without explanation
You pay a SaaS vendor monthly. Usually it’s around the same amount. But this month it’s 15% higher. Is that normal? Did they add seats? Did the pricing tier change? Is it a billing error?
For a finance team processing hundreds of invoices a week, there’s no time to investigate every fluctuation. So most of the time, it just gets booked. The error, if it is one, only surfaces weeks later when someone reviews the P&L and spots the variance. By then, the payment has already gone out. Getting it back is a different headache entirely.
How our AI agent handles it: The agent tracks the payment history for every vendor. It builds a picture of what ‘normal’ looks like for each relationship: the typical amount range, the usual frequency, the expected fluctuation. When an invoice deviates beyond an acceptable threshold, it flags it before posting. Not after. The accountant can then check whether the increase is legitimate, request a corrected invoice if it’s not, or approve the new amount and update the expected range going forward.
This is what we mean when we talk about applying institutional knowledge through AI. The agent learns what’s normal for your specific vendor relationships and acts on deviations. It’s the same judgment call an experienced AP clerk makes instinctively, just applied consistently across every invoice, every week, without the experienced AP clerk needing to be there.
The common thread: your master data is the missing layer
Every one of these five challenges has the same root cause. The invoice itself doesn’t contain enough information to process it correctly. You need context. Historical context.
How did we handle this vendor last time? What compliance rules applied to this entity? What did the bank details look like on the previous invoice? What’s the normal amount range for this subscription? Is this page an invoice, a credit note, or a pro-forma?
That context lives in your master data. Your transaction history. Your vendor records. Your chart of accounts and the institutional knowledge of how your team has handled thousands of invoices before this one.
Most AI invoice processing tools don’t use any of it. They extract what’s on the page and hand the rest back to your team. That’s why ‘automation’ often just means faster data entry that still needs a human checking every line.
We built Quid to close that gap. Our AI agents don’t just read invoices. They cross-reference your entire historical master data, apply your institutional knowledge, and post the entries that match known patterns automatically. Your team only sees the exceptions that genuinely need a human decision.
Quid is an AI-native finance workflow automation firm based in Europe. We deploy AI agents that automate manual, rule-based finance workflows - from invoice processing and approval routing to vendor intelligence - helping finance teams achieve a faster month-end close.
If any of these challenges sound familiar, we’d be happy to walk you through how we’d handle them for your specific workflows.