How to automate accounts payable without replacing your ERP: 9 steps

Automating accounts payable doesn’t have to mean replacing your existing ERP or overhauling your entire finance operation. With the right approach, you can reduce manual invoice processing, streamline approvals and improve financial control while keeping your existing accounting systems in place.

The challenge is knowing where to start. From invoice capture and AI-powered data extraction to approval workflows, ERP integrations and payment reconciliation, each stage needs to work together without disrupting your month-end close.

In this practical guide, we’ll walk through nine steps to automate accounts payable, explain how to integrate AP automation with your existing ERP, compare implementation options and costs, and show you how to measure success.

Key takeaways

• Map the existing workflow before changing anything, because the current process encodes decisions worth keeping.

• Automate intake and extraction first, then routing; they pay back fastest and carry the least risk.

• Keep the ERP as the system of record and add an AP layer that exports journal entries natively.

• Configure matching tolerances and exception routes so people handle judgement calls, not data entry.

• Run a parallel close before cutover and measure touchless rate and cycle time, alongside exceptions, from day one.

This guide is for finance controllers, AP specialists and finance operations managers at UK and European businesses. It covers implementation decisions, fraud controls and measurable KPIs, so you can build the business case and roll out automation without compromising your close.

Why automate accounts payable now

Top-performing AP teams process an invoice for about 79% less than everyone else, and that gap is the business case. Ardent Partners’ State of ePayables 2025 puts the all-inclusive cost of an invoice (staff, benefits, technology, and overhead) at USD 9.84 on average. It is USD 12.42 for the “All Others” group and USD 2.65 for the “Best-in-Class” group, defined as the lowest 20% by cost and cycle time. The sample is 204 AP professionals, about a third of them in EMEA, and the figures are global in USD. For a team handling 400 invoices a month, the distance between “All Others” and “Best-in-Class” is close to USD 3,900 a month.

Most of that cost sits around the invoice rather than inside it. Re-keyed amounts and duplicate payments spotted at reconciliation rather than at intake both land in the last week of the close. Approvals that stall in someone’s inbox do the same. The AP specialist absorbs the first two, the controller absorbs the third, and neither has a number to show for it until something goes wrong.

Late payment carries a UK-wide cost as well. The Department for Business and Trade’s July 2025 research estimates that late payments cost UK businesses almost GBP 11bn a year, with GBP 26bn owed at any one time. Late payments also create pressure for UK businesses and suppliers. Treat these figures as context for the business case, not as a legal or payment-term recommendation. If approval delays are pushing suppliers past their agreed due dates, reducing those delays is a process improvement.

None of this means the current process was wrong. A shared inbox and a spreadsheet are the right tools at low volume, because they cost nothing and everyone understands them. They stop being the right tools when volume and entity count grow, or when approvers multiply faster than the two people maintaining them.

How to automate accounts payable: the nine steps

The order matters because each step produces the data the next one depends on. Each entry names the software category involved and the point where it touches your ERP.

1. Map your current AP workflow. Start without software. Map the workflow against the ERP’s chart of accounts and dimensions, followed by the supplier master.

2. Centralise invoice intake. Use invoice capture in the AP platform. The ERP touchpoint is the supplier master used to identify who sent what.

3. Extract invoice data with AI-powered OCR. Document extraction sits in the AP platform, while the ERP supplies the VAT codes and field formats the ledger expects.

4. Auto-code to the general ledger. Bookkeeping automation uses GL accounts and cost centres, with projects where applicable.

5. Configure approval routing rules. Approval workflows control the point at which an approved bill can post to the ERP.

6. Introduce purchase order and invoice matching. Procure-to-pay uses the purchase order and goods receipt records held for the ERP workflow.

7. Define exception handling. Workflow configuration determines which exceptions block posting to the ERP.

8. Connect to your accounting or ERP system. Use a native integration with NetSuite, Xero, DATEV, QuickBooks, Sage, or Microsoft Business Central. The journal entry is the final ERP touchpoint.

9. Execute and reconcile payments. Payment execution connects payment status with bank reconciliation.

Step 1: map your current AP workflow

Write down who touches an invoice today, in what order, before you configure anything. The existing process holds decisions nobody has documented, and the map is what stops those decisions disappearing during the rollout. For a typical mid-market team, the map covers:

• Every intake channel, including supplier email, post, portals, and structured e-invoices

• Each approver and the threshold they apply in reality, not only on paper

• Who decides the GL code and cost centre, plus who determines VAT treatment and from what reference

• Payment methods in use, from bank portals and SEPA batches to cards

• How payments are reconciled back to invoices and by whom

The output should fit on one page and should expose the supplier master’s condition. A master with duplicates and three spellings of the same supplier takes weeks to validate rather than days, and it will slow every later step. Cleaning it now is the single most useful piece of preparation you can do.

Step 2: centralise invoice intake

One inbox for every invoice gives the rest of the workflow a single starting point. Suppliers forward PDFs to a dedicated address, paper and scans are uploaded, and the platform records each arrival with its source and timestamp. Where structured e-invoicing rules apply to your entities, the inbox also needs to accept structured formats rather than only PDFs; check the requirement for each entity with your adviser.

Duplicate detection belongs at intake, not at reconciliation. At intake, you can compare supplier, invoice number, amount, and date against everything already received and flag a likely duplicate before anyone approves it. That distinction matters most when a supplier resends a reminder copy that looks identical to the original.

In Spendesk’s accounts payable automation, you see repeats flagged at capture, before they enter the approval workflow. Its invoice management features let you handle capture, approvals, matching, and payment tracking in one place.

Step 3: extract invoice data with AI-powered OCR

Optical Character Recognition (OCR) software turns data entry into a review task, because each invoice is read and the fields the ledger needs are already filled. AI-powered OCR goes further than a template-based scanner. You get supplier name, invoice number, dates, net amount, VAT by rate, total, and line items extracted even from layouts it hasn’t seen before. VAT by rate is the field to test hardest, because a single total VAT figure on a mixed-rate invoice is not enough to reclaim correctly.

A good extraction layer assigns a confidence score to each field. Blurry scans, handwritten amounts, multi-invoice PDFs, and credit notes should arrive in the review queue with the uncertain field highlighted, rather than being pushed through as clean data you discover at audit. The next step is to ask any vendor to show you that failure path on your own worst invoices during evaluation.

Step 4: auto-code to the general ledger

Coding runs on two mechanisms, and the distinction decides how much the team still has to touch. Rules-based coding is deterministic: if the supplier is your cleaning contractor and the cost centre is Facilities, apply the general ledger code (GL code) for premises costs every time. Intelligent categorisation covers everything the rules don’t, by suggesting a GL code based on how similar invoices were coded before. In both cases, the accountant reviews and confirms; the platform never posts an unconfirmed field.

Published suggestion-accuracy figures come from vendors and their customers, so test suggestions against a month of your own coded invoices rather than relying on a headline percentage. Spendesk’s bookkeeping automation combines those deterministic rules with machine-learning suggestions, and finance confirms the prepared fields before export.

Step 5: configure approval routing rules

Routing rules replace the email thread with a record of who approved what, when, and against which threshold. Multi-condition rules combine department and amount band, with supplier-specific conditions where needed. For example, a team could route a GBP 3,000 marketing invoice to the marketing lead, while sending a GBP 30,000 capex invoice to the marketing lead and then the finance director. Approval workflows can be sequential, where each approver waits for the previous one, or parallel, where two approvers act at once and the invoice moves when both are done.

Escalation handles the approver on leave. A rule that reassigns after a set number of days keeps invoices from ageing in someone’s queue, and mobile approval means the budget owner can act between meetings. The ERP touchpoint is simple: only an approved invoice is allowed to post, so the ledger never carries an unapproved liability.

Step 6: introduce PO and invoice matching

Finance teams use matching to turn approval from a judgement into a check. A purchase order (PO) records what was ordered, from whom, at what price. Two-way matching compares the invoice to the PO on price and quantity. Three-way matching adds the goods received note, so the software also checks the invoice against what arrived.

You decide how much variance can pass without a human by setting tolerance thresholds. These can combine a percentage with an absolute amount: “5% up to a maximum of £100” behaves differently from “2.5% or £25, whichever is lower”. Within tolerance, you can allow optional automated approval for an invoice that matches an approved PO, which is where touchless processing comes from. Outside it, the workflow routes the invoice to a person. Matching only works if purchase orders exist. The prerequisite is a purchase request process people will follow, one that creates the PO from an approved request so the supplier invoice has something to match against.

Step 7: define exception handling

Every invoice the automation cannot settle should land with a named owner and a reason, and that list is the real design decision. The common routes to a human are:

• Price or quantity variance beyond tolerance

• Missing PO or missing goods received note

• Supplier not found in the master, or bank details changed since the last payment

• VAT breakdown that doesn’t reconcile to the total

• Suspected duplicate

• Extraction confidence below your threshold

The next step is to assign each type to a role, not a person, and set a response time. The point of automation is to concentrate the team’s expertise on these calls, so supplier disputes and credit notes against partially delivered orders stay with humans. A person should also review any change to who gets paid.

Step 8: connect to your accounting or ERP system

Prefer a native integration or a mapped export rather than a CSV that your team has to reformat manually. The export should carry the GL account, cost centre or dimension, VAT code, supplier ID, invoice number, and attachment, mapped to the field names your ledger uses. If the integration does not sync the fields your bookkeeping depends on, the gap shows up at month-end rather than during the demo.

Depending on the connector, you can synchronise accounting master data and export approved payables with mapped fields and supporting documents from Spendesk. Some payment or settlement flows remain file-based, so confirm the exact behaviour for your ERP, entity, and plan.

Test the supplier master in both directions. The useful behaviour is a warning when a supplier on an invoice does not exist in the ERP, so the entry cannot post against nothing. Ask each vendor to demonstrate that warning on a test invoice from a supplier you have deliberately left out of the ERP.

Spendesk offers accounting integrations and export routes for a range of accounting and ERP systems. Connector capabilities vary by configuration, entity and plan, so confirm the available synchronisation, export and payment flows for your ERP.

Step 9: execute and reconcile payments

Decide early whether the AP platform pays suppliers or only prepares the payment, because the two are different controls. Managing AP means approving the liability; executing payment means moving money. Some teams keep execution in the bank portal or the ERP and use the platform to produce an approved payment batch. Others pay from the platform and let the status flow back to the invoice.

Payment timing depends on the payment method, bank, currency, destination and provider. If the platform executes payments, confirm cut-off times, settlement windows and eligibility for your entity and plan. Spendesk’s AP automation supports international invoice payments in 70+ currencies, subject to applicable requirements. When payment status is tracked against each invoice, reconciliation becomes a confirmation step rather than a search.

A worked 2-way vs 3-way matching example

A GBP 40 variance passes or fails depending on the tolerance you set, not on the GBP 40 itself. Take a PO for 10 monitors at GBP 120 each, GBP 1,200 in total, and an invoice that arrives at GBP 1,240 because the supplier added a GBP 40 delivery charge. The variance is GBP 40, or 3.3%.

The £40 figure here is a configurable example, not a published standard. Published UK institutional examples include “10% and/or £10, whichever is lower” in 2016 NHS procurement standards and 5% up to a maximum of £100 at the University of Cambridge. Pick a tolerance by asking what variance you would accept without a phone call, then make the platform enforce the answer. The fourth row shows why three-way matching earns its extra document: the price was fine, and only the goods received note revealed that two monitors never arrived.

Which AP tasks to automate first

Centralised intake and OCR extraction come first because they pay back immediately and break nothing if they fail. An unread invoice still sits in a queue a person can see, which is already better than a PDF in someone’s personal inbox. Approval routing follows, because it removes the chasing and gives you cycle-time data for the first time. Coding rules come third, once a month of extracted invoices shows which suppliers and categories recur.

Matching comes after that, and only when purchase orders are in use for the categories you want to automate. Payment execution is last; move it only when the approval and audit trail has run cleanly through at least one close.

Supplier onboarding and bank-detail changes should stay manual, so that a person reviews every change to who gets paid. The same goes for disputes, intercompany invoices, and one-off capital purchases. Even top-performing teams in Ardent Partners’ 2025 global benchmark reach a touchless rate of 51.0%, so roughly half their invoices still need a person. The next step is to set expectations with your team accordingly: automation changes what the work is, from typing to deciding, more than it changes how many invoices need a person.

Three routes to automation compared

The three routes differ most in what they control once the invoice is read. The table compares a dedicated AP platform, an ERP-native AP module, and a no-code integration flow such as Make.com with an OCR service.

For a controller, the main difference is where the controls live. With the ERP-native route, the controller keeps the workflow in one vendor’s hands. This route tends to stop at capture, so the team may still manage approval routing and tolerance logic elsewhere.

The iPaaS route costs the least to start, but the finance team faces the hardest audit task. The builder must construct matching and approvals, along with duplicate checks and other controls. Control coverage depends on how the workflow is configured. Before selecting an iPaaS route, confirm audit-log retention, approval controls, duplicate checks and rollback behaviour.

The dedicated platform costs more per month and requires an integration project. In exchange, dedicated platforms typically include approval and matching, plus duplicate detection and a chain of custody, which you should confirm per vendor. Choose based on whether budget and IT time create the main constraint, or whether the audit trail does.

What AP automation costs and the invoice volume that justifies it

Published AP automation pricing varies considerably between vendors, with costs depending on user numbers, invoice volumes, payment methods, integrations and additional modules. Some providers publish starting prices, while others require a tailored quote.

• Tipalti: Accounts payable plans start at £99 per month in the UK, with unlimited users. Additional transaction fees apply to invoice processing and payments, and pricing varies according to modules, entities and usage.

• BILL, AvidXchange and Ramp: BILL publishes US plans starting at $49 per user per month for Essentials, $65 for Team and $89 for Corporate, with additional transaction fees, including $0.59 for eligible ACH payments. AvidXchange uses a quote-led model without readily available public list prices. Ramp offers a UK Free plan at £0 per user per month, while Ramp Plus is custom-priced for UK businesses.

• ERP-native solutions: Microsoft Dynamics 365 Business Central Essentials starts at £61.50 per user per month in the UK, paid annually, with Payables Agent usage billed separately through Copilot Credits. Sage 50 Accounts Standard starts at £115 plus VAT per month, including 50 AI document captures, with additional approved captures charged at £0.20 each.

• DATEV and NetSuite: DATEV’s Rechnungsdatenservice 1.0 lists a charge of €2.10 per month for transmitting up to 200 invoices, excluding VAT, under its 2026 pricing schedule. This is a specific service fee rather than the total cost of DATEV. NetSuite offers invoice capture capabilities, but pricing for the relevant modules and implementation should be confirmed directly with the provider.

Spendesk prices on a fixed monthly platform fee plus variable transaction-usage fees, with users included and no per-user, per-card, or per-login charge. Modules such as AP are added to the base plan, and specific prices come from a specialist rather than a price list.

To size the volume that justifies the spend, divide the monthly cost of the platform by what you save per invoice. Using the Ardent gap earlier in this article as the ceiling, moving from the “All Others” figure to “Best-in-Class” saves about USD 9.77 an invoice. If you model half of that gap as a conservative first-year planning assumption, the saving is roughly USD 4.90. On that basis a platform costing USD 1,000 a month breaks even at around 205 invoices a month, and one costing USD 500 at around 100. Your own quote and your own fully loaded cost per invoice (total AP cost divided by invoices processed) should replace those placeholders. The staff time you free only counts as a saving if it goes somewhere the business values.

Can AI actually process invoices?

Yes, for extraction and suggestion; no, for the decision to pay. Current products combine AI-powered OCR with intelligent categorisation. Rules-based coding provides a separate deterministic mechanism. Each has a different reliability profile. AI-powered OCR reads the document and extracts fields with a confidence score. Intelligent categorisation suggests a GL code from patterns in your past coding. Rules-based coding applies the deterministic mappings you wrote yourself. OCR and categorisation do not, by themselves, make the payment decision, so the next step is to keep approval, exception, and payment controls explicit.

The failures cluster in predictable places. Extraction drops on poor scans, handwritten annotations, multi-invoice PDFs, foreign-language layouts, and VAT split across rates. Suggestions drift when a supplier starts billing for a new kind of service, because the pattern they learned no longer applies. Published accuracy figures range from 81% to 99% across vendors, but they measure different things. Some measure field extraction or coding coverage; others measure approval recommendations. They aren’t comparable with each other or with your own data.

The practical design is review for anything below a confidence threshold, confirmation for everything else, and an audit record of what the machine proposed versus what the person decided. With Spendesk’s AI capabilities, finance keeps review and control while OCR, intelligent categorisation, GL-code suggestions, duplicate-invoice detection, and anomaly detection prepare the data. You can see how that machine-learning automation is scoped before deciding how much of your queue to let it pre-fill.

Integrating with your ERP without replacing it

The AP layer sits on top of the ERP, and the ERP stays the system of record. Invoices arrive and get read in the AP layer, then get routed for matching and exception work. The ERP receives a journal entry with approved coding and its attachment. It keeps the ledger, together with its reporting and audit history. The rollout doesn’t change the chart of accounts or dimensions. It also leaves the close calendar in place.

That is the answer to the rip-and-replace objection. The objection is usually about risk rather than cost: a controller who has spent two years getting NetSuite dimensions right doesn’t want a new tool rewriting them. A native integration respects that by syncing the ERP’s supplier and account lists, followed by its cost centres, into the AP platform. Coding then happens against the real values, and the export lands in the right fields. The integration test to run before go-live is a round trip: create a supplier in the ERP and confirm it appears in the platform. Then code an invoice, export it, and check the journal entry in the ledger field by field.

Spendesk is an all-in-one spend management platform consolidating smart company cards, expense management, accounts payable, procurement, and budgeting. For an ERP-first finance team, the gap it closes is the handoff between an approved invoice and the ledger: you export the entry natively with mapped fields rather than rebuilding it from a CSV. Because integration support varies by configuration, the next step is to review Spendesk’s current integration catalogue during evaluation to confirm which connector or export route fits your stack.

Fraud controls and audit trails under automated payment release

Automating payment release raises the stakes on who can change what, so the controls have to be designed before the first automated payment, not after. Invoice fraud is the most common type UK businesses face. The Home Office’s Economic Crime Survey 2024, covering 3,477 businesses with employees, found 11% had experienced fake invoice fraud and 7% mandate fraud, where a criminal persuades the buyer to change a supplier’s bank details. A Finextra analysis makes the uncomfortable point that duplicate checks and approval workflows alone may not stop bank-account redirection, because the invoice itself is often genuine.

Segregation of duties is the first control, and it is one a platform can enforce through roles. Use this checklist to reduce the chance that one person can create and complete an inappropriate payment without another check:

• Separate invoice capture from payment execution

• Assign invoice approval to a different role from supplier setup

• Keep payment confirmation independent of approval

• Require a documented independent review when the team is too small to separate every duty

Payment approval limits add a second layer. A rule that routes any payment above a threshold to two approvers catches one part of the mandate-fraud pattern. The next step is to route any payment to a new or changed bank account to finance regardless of amount, which covers the part of the pattern that a threshold alone misses.

Because the invoice itself is often genuine, the next step is to treat every supplier bank-detail change as an exception. Confirm it by phone on a number you already hold, not one in the email, and log who verified it. Some banks and payment providers offer payee or beneficiary checks, but these are an additional control rather than a substitute for independent verification.

The audit trail is the digital chain of custody for each invoice. It should retain the source email and OCR output, then record every edited field and who changed it. Each approval needs its timestamp and threshold, while the payment reference and ERP export ID complete the trail. Auditors will also ask who changed the rules, so the platform should log edits to approval workflows and coding rules as carefully as it logs invoices.

Spendesk applies role-based access controls and multi-factor authentication. It also states that it holds ISO 27001:2022 certification. Its security materials cover the controls and processing details a security assessment needs.

Running a parallel process during implementation

A parallel run through one full close is what protects the month-end while the new workflow proves itself. The next step is to pick a pilot scope you can isolate, such as one entity or the suppliers of one department, and process its invoices in both the old and new workflow for a month. Then compare the results invoice by invoice: amount, VAT, GL code, approver, and posting date. Discrepancies in that comparison are configuration problems you fix before cutover, not errors a supplier finds after it.

Timing is the part most often got wrong. The next step is to start the parallel run in the first week of a month, so it finishes before close rather than during it, and to avoid scheduling go-live for the week the auditors arrive. The old process then keeps running for the pilot scope until the first fully automated close reconciles cleanly, after which you retire it deliberately.

PayStream guidance gives 4 to 12 weeks for straightforward cloud deployments and 3 to 6 months for complex ones. Corpay, an AP vendor, gives a range of 6 to 16 weeks for most mid-market organisations. Spendesk states that implementation takes roughly two to six weeks.

Treat each range as guidance rather than a guaranteed schedule. The project team should set the timeline according to ERP integration depth, entity count, and internal readiness. A clean supplier master and a documented workflow reduce the validation work, while poor data quality adds to it.

A 90-day rollout plan and the KPIs that prove it worked

Ninety days is enough to move a pilot scope from manual to measured, provided you capture the baseline before anything changes. The plan runs in three blocks:

• Days 1 to 30, baseline and build. Complete the workflow map, clean the supplier master, measure current cost per invoice, cycle time, and exception rate for the pilot scope, configure intake, OCR, and routing, and connect the ERP integration in a test environment.

• Days 31 to 60, parallel run. Process the pilot scope in both workflows through one close, resolve every discrepancy, add coding rules from the first month’s data, and switch on matching for PO-backed categories only.

• Days 61 to 90, cutover and measure. Retire the old process for the pilot scope, run the first fully automated close, compare the four KPIs with the baseline, and decide the next scope to add.

The KPIs below are the ones with an independent benchmark. The Ardent 2025 figures are global, about a third EMEA, and best read as direction rather than targets for a European mid-market team:

Your before-and-after comparison is the number the business case rests on, and the baseline from days 1 to 30 is what makes it credible. A team that moves from a 20% exception rate to 12% on the pilot scope, with cycle time down from ten days to four, has evidence to extend the rollout to the next entity. A team without a baseline has a feeling.

The ERP you started with is still the ERP at day 90, and the spreadsheet that ran AP for years can be retired with the respect it earned. The right setup depends on entity structure, location, integrations, transaction volume, and current plan terms, so if those variables are central to your decision, talk to a Spendesk specialist about the configuration that fits your ERP and your close calendar.

Frequently asked questions

The right answers depend on how the platform connects to your ledger and where your team needs human review. These questions cover software selection, the volume case, and controls against duplicate payments and fraud.

What is the best software to automate accounts payable?

The best option exports a mapped journal entry to your specific ERP and enforces the approval and matching rules you already apply. It should also price in your currency. The next step is to test, during evaluation, for a native connector for your ledger rather than a generic CSV, duplicate detection at intake, configurable tolerance thresholds for matching, and a visible exception queue with named owners. For UK and European teams, local accounting connections such as DATEV or Exact Online are worth adding to that list. Structured e-invoice support matters for the countries where it is mandated, and so does SEPA payment execution if you intend the platform to pay suppliers.

What invoice volume justifies automating accounts payable?

The volume that justifies automation depends on your own cost per invoice and the vendor’s monthly fee, not on a rule of thumb. The next step is to apply the break-even formula from the cost section to your own quote and fully loaded cost per invoice. A controller may still approve automation for a low-volume team with several entities and currencies, especially when many approvers are involved. Control and audit requirements can support that case before the cost calculation closes. The deciding question is whether the team will redeploy the hours currently spent re-keying and chasing to work the business values.

How does AP automation stop duplicate payments and invoice fraud?

Duplicate payments stop when every incoming invoice is compared against supplier, invoice number, amount, and date before it enters approval, and a likely repeat is blocked until a person clears it. Fraud takes more than that, because fraudulent invoices are often genuine documents with changed bank details. Role separation is the control that works here: no one person can set up a supplier and approve its invoices. A separate person releases the payment. The next step is to add an approval rule that routes every bank-detail change to finance for independent verification. An audit trail that records each field edit, approval, and payment reference lets an auditor trace the whole path.