It is the third working day of the month and the group controller is still waiting for two reconciliations: intercompany balances with three subsidiaries, and the bank clearing account, which has carried 47 unapplied receipts for a week. The CFO has promised the board a flash result by Friday. At the same moment procurement forwards a complaint from a supplier who invoiced two weeks ago and has not been paid. The reason is sitting in a warehouse manager’s inbox: the goods receipt was never confirmed, so the invoice never cleared matching.

A year earlier the same company bought an invoice capture tool. It works well. The problem is that a growing share of supplier invoices no longer needs capturing, because they arrive as structured files, and the manual work never disappeared. It moved to places nobody automated: matching, cash application, reconciliation.

This article is for CFOs and the finance and IT leaders who prepare automation decisions for them. It defends one position that not everyone will like: finance automation should start in the middle of the process, with matching and reconciliation, not at its edges. Document capture at the front is increasingly being solved by e-invoicing mandates, and the decisions at the back (releasing payments, changing supplier bank details, writing off receivables, booking estimates at close) should stay with people. What automation costs and how to calculate its payback is covered separately in our article on what business process automation really costs. Here the question is sequence: what to do first, what next, and what never to hand to a machine.

Why is the starting point for finance automation moving?

For the last decade, a typical finance automation project looked the same. Supplier invoices arrived as PDFs by email or as paper, so step one was to scan and read them. The market answered with OCR, then machine learning models, and most recently large language models that pull invoice numbers, dates, amounts and lines out of a document. Capture was the bottleneck, so that is where everyone started.

E-invoicing mandates change the starting point. Poland is the clearest current example. According to the Polish Ministry of Finance, structured invoicing through the national KSeF system became mandatory on 1 February 2026 for taxpayers whose 2024 sales exceeded PLN 200 million including VAT, and on 1 April 2026 for everyone else, with a transition until the end of 2026 for the smallest issuers. More important for accounts payable: receiving invoices through KSeF has been mandatory since 1 February 2026. A domestic supplier invoice in Poland is no longer an image to read. It is an XML file with a known schema and a system-assigned identifier.

At EU level the direction is set. The VAT in the Digital Age package was adopted on 11 March 2025 and published as Directive (EU) 2025/516. Since its entry into force on 14 April 2025, member states may introduce mandatory domestic e-invoicing under specific conditions, digital reporting requirements will apply to cross-border B2B transactions from 1 July 2030, and member states with domestic real-time reporting must align their systems with the EU model by 1 January 2035. For a company operating in several EU countries, structured inbound invoices are a question of when, not whether.

That does not make capture obsolete everywhere. US companies face no comparable federal mandate, and suppliers outside a mandate’s scope still send whatever they like. But even where capture is still needed, it is the wrong first project, for a reason that holds regardless of regulation: capture errors flow straight into matching. If matching is manual, a better capture tool only gives people cleaner data to reconcile by hand. If matching is automated, capture quality becomes measurable through the exception rate, and you know exactly how much a better capture tool is worth.

The second consequence is less obvious. Many finance teams already own software that could automate matching and reconciliation. It was never configured properly because OCR data was too unreliable to drive automatic decisions. Structured invoice data changes the error tolerance calculation and makes rules viable that did not make sense before.

What actually happens in accounts payable, receivable and the close?

Before automating anything, break the three processes into steps, because “finance automation” covers dozens of activities with very different risk profiles. Our glossary entry on process automation covers the general techniques; here the question is which finance steps suit them.

Accounts payable (AP). Receiving the invoice, identifying the supplier, order and legal entity, matching it to the purchase order and goods or service receipt, coding it to accounts and cost centres, business approval, posting, scheduling payment, building the payment run, verifying the supplier’s bank account, releasing the payment and reconciling it against the bank statement.

Accounts receivable (AR). Issuing and transmitting the sales invoice, monitoring due dates, reminders and soft collections, applying incoming cash to open items, resolving short payments, overpayments and lump-sum remittances, managing credit limits, bad-debt provisions and write-offs.

Month-end close. Bank reconciliations, customer and supplier ledger reconciliations, intercompany reconciliation, accruals and prepayments, provisions, depreciation, foreign currency revaluation, completeness checks (are all period invoices posted?), tax reporting and the management pack.

Laid out like this, the steps fall into three kinds of work. The first is moving and joining data: from the e-invoicing channel into the ledger, from the bank statement into the sub-ledgers, from the ledgers into tax reports. The second is comparing against a rule: is the invoice price within tolerance of the order, does the receipt match an open item, does the account balance agree with the statement. The third is judgement: should we accept an invoice with a discrepancy, should we pay to a new account, how large should the provision be, is this receivable still collectable.

In those terms, the thesis is: automate the first two kinds aggressively, support the third with data, but leave it with people. Most disappointments in finance automation come from confusing the second kind with the third, trying to replace judgement with a rule or a model where an error is expensive and hard to detect.

What does e-invoicing change in accounts payable, and what does it leave untouched?

First, it changes input data quality. A structured invoice has defined fields: seller and buyer tax IDs, dates, lines, rates and amounts. Nobody needs to read them or check whether a model misread a digit. In Poland’s case, the system assigns each accepted invoice a unique KSeF number, which gives a reliable identifier for everything downstream.

Second, it can change the legal moment of receipt. Under the Polish rules, for a domestic VAT taxpayer the invoice is received on the date the system assigns its KSeF number, not on the day someone in the company opens it. That matters for VAT deduction and for payment terms counted from receipt. An invoice “sitting in an inbox” stops being a valid excuse, because the system knows when you received it regardless of when you downloaded it. Other national systems define receipt in their own ways, so check the rules in each country where you operate.

Third, it ties tax reporting to the identifier. Poland’s Ministry of Finance states that since 1 February 2026 the VAT return file (JPK_VAT) must report the KSeF number of each invoice in its records section. The VAT ledger therefore has to be linked to system identifiers, which is another argument for automatic retrieval over manual keying.

What does e-invoicing not change? It does not tell you which purchase order an invoice belongs to. There is a field for the order reference, but suppliers fill it inconsistently or not at all. It does not tell you whether the goods arrived, or in what quantity. It does not tell you which cost centre a consulting invoice belongs to or who in your company ordered the work. And it does not tell you whether the bank account on the invoice is one you should pay. All of these are questions about your internal data, not about the document.

Mandates also come with fallback modes. In Poland, an invoice issued in offline mode must reach KSeF no later than the next working day, and deadlines extend during system outages. AP automation has to handle a supplier presenting an invoice with a QR code before it appears in the system. That is not a reason to hold back; it is a reason to design an exception path from day one.

In short, e-invoicing solves “what does the invoice say”. It does not solve “what does this invoice mean for us”. The second problem is now the main source of manual work in accounts payable.

Why is matching invoices to orders and receipts the first candidate?

Because it meets the three conditions of a good automation candidate: it is frequent, it runs on rules, and its errors are detectable. Three-way matching (invoice, purchase order, receipt) checks that the supplier bills what was ordered, in the quantity that arrived, at the ordered price. Two-way matching (invoice and order) applies to services, where there is no goods receipt.

The rule is simple to state: supplier and order agree, quantity within tolerance, unit price within tolerance, total within tolerance. When every condition passes, the invoice is posted and scheduled for payment with no human touch. When one fails, it goes to an exception queue with the reason attached.

The most important decision in this project is not the tool. It is setting the tolerances. Too tight, and rounding differences of a few cents flood the queue, so automation achieves nothing. Too loose, and a supplier’s systematic overpricing passes unnoticed. Tolerances should differ by spend category and be agreed by the CFO with procurement, not left at a system integrator’s defaults. They are a risk decision that automation merely executes.

The second condition is discipline upstream. If half your spend has no purchase order in the system because managers order by phone, three-way matching has nothing to match against. If receipts are confirmed a week late, invoices wait in the exception queue not because of a discrepancy but because of missing data. Before you automate matching, measure on last month’s invoices what share had a purchase order and a confirmed receipt on the day the invoice arrived. That single number tells you what automation can deliver in its first quarter.

For spend without purchase orders (utilities, rent, subscriptions, professional services) history-based rules work better: same supplier, amount within tolerance, same day of the month, same coding as previous periods. An office rent invoice that has looked identical for two years does not need monthly business approval. It needs an alert when the amount or the bank account changes.

The effect of good matching shows up immediately in two numbers: the touchless rate (share of invoices posted without human intervention) and average time from receipt to posting. Capture both before go-live as a baseline.

What should you automate in accounts receivable: cash application, reminders or credit limits?

In receivables, the largest source of manual work is usually cash application. A customer pays three invoices with one transfer, quotes an order number instead of an invoice number in the remittance, deducts a disputed amount, or pays from a subsidiary’s account. Someone has the bank statement on one screen and the open items on another and has to work out what is what.

The best rules are boringly simple: exact amount plus invoice number in the remittance; exact amount and the customer’s only open item; several open items summing to the receipt. Each rule should have a measured hit rate, and any receipt that matches no rule should go to a queue with a suggested match, never be applied to the “closest” open item. A wrongly applied receipt is worse than an unapplied one, because it hides two errors: an overstated balance on one invoice and an understated one on another.

Structured identifiers will make this much easier where regulators push them into the payment message. Poland is again ahead: from 1 January 2027, according to the Ministry of Finance, an active VAT taxpayer paying another for a structured invoice by bank transfer must quote the KSeF number or a batch identifier covering up to 500 invoices, including in the split payment mechanism. The ministry notes that some implementing rules are still in the legislative process. For groups with Polish subsidiaries, that means the cash application engine should be able to expand a batch identifier into its invoices and handle a receipt that does not equal their sum. Now is a better time to test this than December.

Dunning reminders are the second natural candidate. “Send a reminder three days before the due date and seven days after” is trivial, yet in many companies a person runs it from a spreadsheet. Keep one decision with people: excluding a customer from the automatic sequence. A key account in the middle of contract renegotiation should not receive an automated demand because someone forgot to update a flag.

Credit limits deserve more caution. Automatically blocking orders once a customer exceeds the limit is a rule and should be automated. Deciding to raise a limit, to ship despite a breach or to change payment terms is judgement that depends on information outside the system: the commercial relationship, market conditions, what the account manager knows. The system should prepare that decision (payment history, ageing, exposure) and a person should make it.

How do you automate the month-end close without making mistakes faster?

The close is where speed is most easily confused with quality. “We close in three days instead of eight” looks good on a slide, but if it is achieved by deferring reconciliations or booking estimates without review, the company gets a faster result that needs correcting more often.

A better goal is continuous reconciliation: moving as much work as possible from the last days of the month into the month itself. A bank reconciliation does not have to wait for the close. If statements are pulled daily and cash is applied automatically, a handful of items remain at month end instead of hundreds. The same applies to intercompany: where group companies invoice each other through a structured e-invoicing channel, both sides hold the same identifier and the same amount, so reconciliation can run daily and flag only differences.

What in the close suits automation? Balance reconciliations where both sides come from systems (bank, sub-ledgers, clearing accounts). Completeness checks: are all inbound e-invoices for the period posted, have all sales documents reached the ledger. Scheduled accruals and prepayments where amount and period are known (insurance paid in advance, an annual licence). Currency revaluation at the reference rate. Tax report generation from ledgers.

What stays with people? Anything that requires an estimate or judgement: bonus provisions, accruals for uninvoiced costs, provisions for disputes, impairment of receivables and inventory, revenue recognition on long-term contracts, decisions on whether a difference is material. Automation can prepare the data and propose an amount from history, but posting should require approval from the person accountable for it.

The practical rule is: automate close steps that can be reconciled against an external source of truth (the bank, the e-invoicing system, a counterparty, a schedule), not steps that create a number themselves. A step that can be reconciled carries its own control. A step that creates a number needs a control from outside, and that control is a person.

One more thing gets forgotten: the management pack. If the close shrinks but the board report is still assembled by hand from several spreadsheets, the time saved disappears at the last stage. How to build reports the board actually reads is covered in our article on business intelligence implementation.

Where must a person stay: supplier bank details, payment release, write-offs and estimates?

Some finance activities combine an expensive, hard-to-reverse error with strong appeal to fraudsters. Full automation there is a bad idea however good the tool.

Changes to supplier bank details. This is one of the best-known fraud vectors: a criminal impersonates a supplier and asks for payment to a new account. Automatically checking supplier accounts before every payment run is good automation. Automatically accepting a bank-detail change requested by email is bad automation. Any change should require confirmation through a second channel and approval by someone other than the person who entered it.

Banks are adding a check of their own. The EU Instant Payments Regulation requires payment service providers to offer Verification of Payee: before a credit transfer is authorised, the payer is told whether the payee name matches the account, with results such as match, close match or no match. According to the European Central Bank, this has applied in the euro area since 9 October 2025 and applies in non-euro member states from 9 July 2027. When a payment run returns a batch of “close match” and “no match” warnings, someone has to decide what to do. If supplier master data is messy, there will be many warnings, and people will quickly learn to click through them. Why supplier data is usually the first domain worth cleaning is covered in our article on data management services.

Releasing the payment run. Preparing the run (selecting invoices by due date, checking accounts, calculating amounts, applying early-payment discounts) is worth automating completely. Releasing it at the bank should belong to someone with authority who sees a summary: number of payments, total, new payees, deviations from the usual pattern. That takes a few minutes and protects against the most expensive kind of error.

Write-offs and adjustments. Deciding that a receivable is uncollectable or that inventory has lost value hits profit directly and often has tax consequences. A model can flag items for review, but the decision should carry a name.

Estimates at close. Judgement-based provisions and accruals are, by definition, judgement. The auditor will ask who set the amount and on what basis. “The system” is not an answer.

What these activities share is not technical difficulty but the shape of their risk: errors are rare but costly, and not reliably detectable by another rule. How to design human involvement so it is more than a rubber stamp is the subject of our article on human in the loop.

How do you keep segregation of duties when a bot does part of the work?

Financial internal control rests on a simple principle: the same person should not be able to create a supplier, post its invoice and release the payment. Segregation of duties assumes each step has an owner with defined rights. Automation introduces a new actor that often receives broader rights than any employee.

A software bot or integration account is a system user. If it can create supplier records, post invoices and build payment runs, it combines three roles that the organisation deliberately split between three people. Who controls the bot? Who can change its rules? Who notices when someone does?

A few principles help. Each automation gets its own technical account with rights limited to its task, not a shared “robot” account for every process. Changing an automation’s rules (tolerances, thresholds, exception lists) is a change to a control and requires approval from finance, not just from IT. Every action the automation takes is logged: what it did, under which rule, in which version. And at least once a quarter someone in finance reviews a sample of cases the automation handled without human involvement.

For US-listed companies this is not optional. Section 404 of the Sarbanes-Oxley Act requires management to assess internal control over financial reporting in the annual report, and the SEC’s Section 404 guide for smaller public companies frames the question directly: is the control automated, and if so, how sturdy are the relevant IT controls; or is it manual, and if so, what is the risk of human error? It is a useful frame for any company, listed or not. Automation does not remove risk. It changes its kind, from a person’s occasional mistake to a wrong rule, changed without finance’s knowledge and applied consistently to thousands of transactions.

The practical advice for a CFO: before approving an automation, ask for the permission matrix of its technical accounts and a description of who approves rule changes. If the implementer cannot answer, the problem is not the technology.

Where does AI make sense in finance, and where is it an expensive substitute for a rule?

With structured invoices spreading, AI has a narrower role in finance than vendor material suggests, but a real one. Three situations are worth separating.

AI where data is unstructured. Invoices from suppliers outside any mandate, contracts, balance confirmations, customer correspondence about payments, free-text remittance information. Here a language model or document-processing tool does something a rule cannot: it extracts meaning from text. For example, a customer replies to a reminder saying they dispute part of an invoice because of a quality claim. A model can recognise that this is a dispute, not a promise to pay, and route it to the right person.

AI as a suggestion where rules run out. Coding non-PO invoices, suggesting the most likely open item for a receipt without an invoice reference, flagging unusual transactions for review. The model learns from accountants’ past decisions and proposes an answer that a person accepts or corrects. A well-designed process measures how often suggestions are accepted unchanged, and only on that evidence decides whether a category can move to sample-based review.

AI where a rule would do. Matching a structured invoice to an order that carries its reference, applying a receipt that quotes the invoice number, checking a bank account against a register, sending a due-date reminder. Using a language model for these is more expensive, slower and less predictable than a rule, and harder to explain to an auditor. If a vendor proposes AI for a task you can state as a one-sentence condition, ask why.

Sequence matters. A company that automates the rules first is left with a smaller, harder set of exceptions where AI genuinely helps. A company that starts with AI uses an expensive tool on easy cases and still has no process for difficult ones. What each option costs, and how payback changes with the touchless rate, is set out in the automation cost article linked at the start.

There is also the data question. Invoices, bank statements and customer correspondence contain personal data and commercially sensitive information. If the model is a cloud service, you need to know where data is processed, whether it is used for training and how long it is retained. Under GDPR those are contract questions, not demo questions.

What data must be in order before you automate anything?

Finance automation stands on three sets of master data: the supplier master, the customer master and the chart of accounts with its dimensions (cost centres, projects, legal entities). If they are messy, every rule produces exceptions and people start working around them.

Supplier master. Duplicates (the same supplier under two IDs), outdated bank accounts, missing tax IDs, suppliers unused for years but still active. For matching structured invoices, the tax ID on the invoice must point to exactly one supplier record. For payments, the account on file must be one you have verified. With Verification of Payee, the supplier name on file must also match the account holder’s name, or every payment run will generate warnings.

Customer master. Groups in which a different entity pays from the one invoiced. Customers with several bank accounts. No link between payer and bill-to party. Each of these breaks automatic cash application until it is described in the data.

Chart of accounts and dimensions. If coding depends on one person’s knowledge (“this supplier always goes to project X unless the invoice is about Y”), that knowledge must be written down as a rule or as data before automation takes over. Otherwise the bot will code correctly only for as long as that person stays and quietly fixes its mistakes.

Cleaning this data is usually the least glamorous and most profitable stage of the programme. Run it as a separate step with its own measures (number of duplicates, share of suppliers with verified accounts, share of customers with a documented payer) rather than “along the way”. If your finance system is old and you are considering replacing it, make that decision before automation, not after; the criteria are in our article on custom ERP: build, extend or buy.

A final note on data: integrations. Finance automation nearly always connects several systems: the e-invoicing channel, the ledger, procurement or warehouse, the bank, sometimes CRM. Our glossary entry on system integration covers the basic patterns, and our article on pricing an API integration project shows how to estimate the work before you talk to vendors.

In what order should you roll out finance automation over three quarters?

The order follows from the thesis: clean the data and automate the middle of the process first, then extend to exceptions, and support the decisions at the edges with data. The table below is a starting layout to adapt. Timings are indicative and depend on volume, data quality and systems.

StageProcessWhat to automateWhat stays with peoplePreconditionMeasure
Quarter 1Master dataDuplicate detection, bank-account verification, mismatch reportsMerge decisions, bank-detail changes confirmed via second channelNamed owner of supplier and customer mastersShare of verified supplier accounts, number of duplicates
Quarter 1Accounts payableAutomatic retrieval of structured invoices, two- and three-way matching with tolerancesSetting tolerances, resolving discrepanciesPOs and receipts in the system on invoice dayTouchless rate, time to post
Quarter 2PaymentsPayment run preparation, account checks, discount captureRun release, new payees, Verification of Payee warningsClean supplier masterPayments stopped by checks, time to prepare run
Quarter 2Accounts receivableRule-based cash application, reminders, order block on limit breachSequence exclusions, limit changes, disputesDocumented payer–bill-to relationshipsAuto-applied cash rate, age of unapplied cash
Quarter 3CloseContinuous bank and ledger reconciliation, completeness checks, scheduled accrualsEstimates, provisions, write-offs, materiality callsDaily statements, intercompany on structured invoicesDays to close, post-close adjustments
Quarter 3ExceptionsModel suggestions for coding and cash matching, correspondence classificationApproving suggestions, sample reviewTwo quarters of decision historyShare of suggestions accepted unchanged
OngoingControlAudit trail of automated actions, alerts on rule changesQuarterly sample review, approval of rule changesSeparate technical accounts, permission matrixRule changes without approval (target: zero)

Two things in this table deserve emphasis. First, AI appears only in the third quarter, and as a suggestion, because only then does the company have the set of exceptions that rules left behind and a history of decisions for a model to learn from. Second, every row has a “what stays with people” column. An automation plan without that column is a plan to transfer accountability to a system, and a system does not answer to the board or the auditor.

For groups with Polish entities, add a 2027 line: preparing cash application and payment runs for KSeF numbers and batch identifiers in remittances, and for Verification of Payee at Polish banks. Both changes are announced with known dates, so neither should be a December surprise.

How should you measure the result so the CFO can hold it to account?

The most common mistake in evaluating finance automation is measuring it in saved hours that nobody realises. If AP gains ten hours a week and nobody plans to redeploy them, the P&L does not change. That does not make the automation worthless; it means its value must be described differently.

A better set of measures combines throughput, quality and risk. Throughput: touchless invoice rate, auto-applied cash rate, time from invoice receipt to posting, working days to close. Quality: post-close adjustments, invoices paid late because of the process, duplicate payments. Risk: payments stopped by bank-account checks, bank-detail changes without second-channel confirmation, rule changes without approval.

Each measure needs a baseline from before go-live. Without it, six months later nobody knows whether a 70% touchless rate is an achievement or roughly where things already were. Measuring a one-month sample takes a few days and decides whether the programme can be held to account at all.

Track the exception queue separately: how many items are in it and for how long. Automation that handles 80% of cases but leaves the other 20% in a queue nobody works daily moves the problem instead of solving it. The age of the oldest exception is a good health indicator to show the board alongside the automation rate.

If you want to turn these measures into money, do it conservatively. Count saved time only where you know what work it will be redeployed to, and count avoided costs (late-payment interest, missed early-payment discounts, fraud losses) from last year’s actual data rather than an optimistic scenario. A worked structure for that calculation is in the cost article referenced above.

What are the key takeaways before deciding on finance automation?

E-invoicing mandates are moving the centre of gravity in finance automation. Reading a supplier invoice is turning from a technical problem into a regulatory given, starting with countries like Poland and spreading across the EU under ViDA. The manual work that remains sits in the middle of the process: matching, cash application and reconciliation. That is where automation delivers most today at the lowest risk, because it runs on rules and its errors are detectable.

Remember six things. First, start with invoice-to-order matching and cash application, not with capture or approvals. Second, before automating anything, clean the supplier and customer masters and measure what share of invoices has a purchase order and a receipt on arrival. Third, treat tolerances and thresholds as risk decisions owned by finance, not as tool settings. Fourth, keep bank-detail changes, payment release, write-offs and estimates with people, and give automation the job of preparing data and proposals. Fifth, give every bot its own account, minimal rights and an audit trail, and treat changes to its rules as changes to internal controls. Sixth, plan for what is already announced: structured identifiers in payment messages and Verification of Payee will reshape cash application and payments, and their dates are known.

If you are planning to connect an e-invoicing channel, your bank and your ledger, or to build matching and reconciliation rules around your existing finance system, see how we work in custom software development and system integration, or book a call. We start with a one- to two-week discovery phase that defines scope and architecture before you commit to a full build.