Automation, workflows and agents get used interchangeably. What the difference is, why a set of books does not want maximum autonomy, and which vendors genuinely take over which task.
Every vendor says AI agent now. In two years the word has gone from technical term to sales argument.
To understand what an agent is, and above all what it is not, you first have to separate three words that get used interchangeably: automation, workflow and agent. They overlap, but the distinction matters, and it comes down to one thing: how much freedom does the system have to decide what it does?
Three kinds, in increasing order of freedom
Rule-based automation. If this, then that. You decide in advance exactly what should happen in every situation: if supplier X, then ledger Y. If the invoice says "restaurant", then cost centre entertainment.
That works precisely and entirely predictably, until something arrives that you did not think of. A new supplier. An order in a shape that did not exist before. Then the automation does nothing, and someone on your team is looking at it again.
Workflows, intelligent or not. Here it is still fixed which steps exist and in what order they come: read the invoice, lay it next to the order, if it deviates, queue it. What AI adds is that individual steps get solved better, even when the input differs slightly from what the system saw before.
The path through the system stays the same. Only the steps on it get smarter.
AI agents, autonomous or not. The name says it: there is agency. Connected to context and equipped with a set of instructions, often called skills, an agent decides within boundaries how to tackle a problem. First fetch the order, then the receipt, then last week's email announcing a surcharge. On the next invoice that route may look different, because the situation is different.
Why this matters commercially
The ratio that settles this conversation is not how many invoices pass automatically, but how the time is distributed.
In practice we see the same thing at Claridy every time: the twenty percent of invoices that does not fit the rules costs more than eighty percent of the time. Precisely the cases you did not foresee are the cases someone spends half an hour on.
That is why rule-based automation is not enough. Not because it works badly on the eighty percent, where it works excellently, but because that eighty percent was never the problem.
More agency is not better
It is tempting to think you want as much agent as possible. More things can be solved, after all.
But the more freedom, the less predictability, and in finance that is not always what you want. An administration needs one property most other systems do not: the same input has to produce the same output. Every time, a year from now too, and when someone checks it.
A system that posts this invoice to ledger 4300 today and 4310 next month because it assessed the situation differently is unusable. However reasonable that assessment is, and however often it turns out right. Your auditor asks why this posting was made this way, and "the model thought so" is not an answer.
So it comes down to knowing what to deploy where. And if you get that right, it does not have to be a choice between a system that always does the same thing and a system that thinks along.
The rule of thumb: standardise recurring processes into workflows as far as you can, but put enough intelligence on top that not every deviation lands on an employee. A new supplier, a new item, a missing PO number: those are exactly the cases where you switch agency on.
In practice you make that call step by step:
Deterministic, always. The posting itself. Which ledger, which cost centre, which project, which VAT code. Which threshold applies and what happens above it. That is code you defined and that runs identically every time.
Agency, where it is needed. Working out why an invoice does not match. Fetching the context a person would fetch. Writing an answer to a supplier's email that matches what that supplier actually asked. Those are tasks where a template always falls short, because the situations do not resemble each other.
The short version, and it is the sentence to put to a vendor: the AI reads and investigates, the rule decides.
AI-first, AI-native and agent are not the same thing
Three terms used interchangeably while describing different things.
AI-first and AI-native are about how something was built. An AI-native product was designed around AI from the first line of code, without a template library or a rule library as its foundation. A product with AI on top is an existing system that had an AI layer added later.
Agent is about what something can do. Namely: does it have agency or not.
Those two axes are independent. You can be AI-native and still run a workflow with zero agency. And you can build an agent on top of a twenty-year-old system. So "AI-first" says something about build history and nothing about what happens to your invoice that does not match.
That is also why the label helps you little in choosing. The question that does help is at the bottom of this article.
What "AI in the reading step" actually means
Almost all existing invoice processing has AI, and it sits in the same place nearly every time. It pays to be precise about which place.
An invoice arrives as a PDF or an image. Before anything can happen with it, that document has to be turned into fields: which supplier, which invoice number, which amount, which VAT rate, and with line recognition also which items, in what quantities, at what price.
That conversion is the reading step. AI makes it demonstrably better. You no longer need a template per supplier, a new layout works immediately, and the lines on the invoice are recognised rather than only the header fields. Basecone states on its own page that it applies AI to the total amount, the subtotal and the VAT percentage. Zenvoices does recognition at invoice-line level. TriFact365 claims the system learns from every invoice processed.
What the reading step does not do is decide.
That the invoice says € 12.96 per unit is a reading result. Whether that agrees with the order, whether a difference of € 22.47 is a rounding or a pricing error, which ledger it belongs to, what to do with the difference and who to email about it: those are all decisions. On rungs 1 and 2 they follow from rules you configured in advance, and where no rule fits, it goes to a person.
Where the work sits
The reason agents exist at all is where the automation stopped.
In the analysis Claridy made of 21,642 Dutch finance job postings, every task named was checked to see whether it is directly automatable. For reading and entering invoices, that is 72.8 percent. For matching against purchase orders it drops to 47.2 percent, and for approvals to 22.1 percent.
That descending series is exactly why a fixed path is not enough. The further you get from the reading, the less can be resolved with rules thought up in advance.
The four tasks where agents genuinely take over work
Matching and posting purchase invoices. Laying the invoice against the purchase order and the goods receipt, investigating differences, coding on your own axes, posting. The most mature area.
The invoice mailbox. Supplier questions about payment status, promised credit notes, explained price differences. Barely any software has been bought for this and a lot of time goes into it. This is also where agency is needed most, because no two emails are the same.
Applying bank transactions. Linking one receipt to dozens of open invoices based on a remittance advice sitting as a PDF in a separate email.
Receivables. Reminders, disputes, payment arrangements.
The vendors, placed on the rungs
Basecone and ScanSys. Captured positions per supplier, a system approval for the fully matching invoice, and AI in the recognition. The processing follows rules.
Zenvoices, TriFact365, Blue10, Klippa (now Doxis SpendControl). Recognition at invoice-line level, and at Zenvoices an extensive library of automation rules across administrations. Where it stops: the path is fixed, and the deviation is flagged and routed rather than investigated. That is the boundary of the category rather than a defect, and below roughly a hundred invoices a month it is exactly what you need.
HighRadius, Esker, Serrala, Quadient. Mature products with real agents for cash application and collections. Where it stops: price and implementation suit an organisation with its own shared service centre. For a hundred to five hundred employees this is too heavy.
Claridy. Runs on Exact Online, NetSuite and Microsoft Dynamics 365, reading from them and writing back, and keeps no books of its own. The deviation goes to the agent first, which does the arithmetic and fetches the order, the receipt and the email; the posting that follows is deterministic code you described in plain language. For payables, receivables and bank reconciliation. Price: € 499 per month flat, three-way matching € 250 extra, cancellable monthly.[^bron1] Where it stops: below roughly a hundred invoices a month there is too little repetition to learn from.
American AP platforms such as Tipalti, Stampli, Ramp and Bill are strong on approval flows, cards and payments, and quick to ship agent features. They are not listed here as an option because Dutch accounting conventions, VAT codes and integrations on Exact or AFAS are not their starting point. If you run internationally with a US entity in the mix, that is a different conversation.
The trap: 95 percent is not good enough
A system that is right 95 percent of the time sounds excellent and is unusable in a set of books. Five percent of two thousand invoices is a hundred wrong postings a month, and finding them all again takes longer than posting them yourself would have.
Worse: a system that is usually right teaches your team to stop checking. Then those hundred errors stay invisible until the auditor trips over them.
This is also exactly why agency alone does not work. The less freedom sits in the actual execution, the more predictable the result.
How to start
Not with everything at once, and not autonomously on day one. Be suspicious of a vendor who proposes that.
The order that works: the system proposes every action and your team approves. You see where it is right and where it is not. Then you set a threshold above which it executes on its own, and you move that threshold based on what you have watched happen.
The risk of trying is low because you are not entering a migration. You put something beside your systems that takes over one job. If it does not work, you switch it off and your administration is exactly as it was.
Frequently asked questions
Software with agency: it decides for itself which steps are needed to finish a financial task, rather than following a path defined in advance. It understands documents, fetches the context and acts within defined rules inside your existing systems, with every action in an audit trail.
Rule-based automation executes if-this-then-that rules thought up in advance, and stops the moment something arrives you did not foresee. A workflow fixes which steps exist and in what order; AI makes the individual steps better, but the path stays the same. An agent has agency and decides within boundaries how to tackle a problem.
No, and in finance certainly not. More freedom means less predictability, and a set of books requires that the same input produces the same output. For the posting itself you want deterministic code. Agency belongs in investigating differences and answering messages.
AI-first and AI-native are about how a product was built, agent is about what it can do. You can be AI-native and still run a workflow with no agency at all. So the label says nothing about what happens to your invoice that does not match.
The place. Almost all existing invoice processing has AI in the reading step: turning a PDF into fields, and with line recognition into items, quantities and prices. An agent starts where that stops, at the question of what should happen to those fields when they do not agree.
That does not even have to be the question. All the work before the posting can be done by the agent: working out why an invoice deviates, finding the PO number the supplier wrote down wrong, proposing a resolution, and drafting the email to the supplier or the customer based on everything it has pulled together. A human in the loop matters, but in the right place. Reviewing the data just before it is written to the ERP is often a good final check. That is something quite different from looking up the order number yourself, tracing the ledger discrepancy yourself and entering the correction yourself. In the first case you are checking finished work. In the second you are doing the work.
No. A chat window answers questions, an agent performs work in your systems. The test: can the thing change something in your ERP, or can it only talk about it?
Ask one question: what happens to the invoice that does not match? Does it appear in a list, or does the system first try to work out why itself, and only reach your team when it cannot?
It depends on the task and the size. Cash application and collections at a large organisation: HighRadius or Esker. On Exact Online, NetSuite or Microsoft Dynamics 365, with payables, receivables or bank reconciliation costing structural time: Claridy. If it is only about reading invoices, you do not need an agent but capture software, and Zenvoices, TriFact365 and Blue10 are cheaper.
More on the architecture behind this: what is a system of action for finance. On what AI does and does not automate in invoice processing: automated invoice processing. All vendors side by side: the comparison guide.
Source: Claridy analysis of 21,642 Dutch finance job postings, 2026. Last checked: 2026-08.