How to Automate Accounts Payable Step by Step (Without Changing ERP)

Automating accounts payable is, in most companies, the AI project with the fastest and easiest to demonstrate return: high volume, clear rules, and a cost per invoice that can be measured before and after. And contrary to popular belief, it doesn't require changing your ERP or touching your accounting: the agent connects to what you already have—SAP, Business Central, Odoo, Holded, A3—and works on top of it.

This is the step-by-step sequence we follow when implementing it, with indicative numbers for each phase. The full process, from zero to production, takes between 4 and 8 weeks in a company processing between 200 and 2,000 vendor invoices per month.

Step 0: Measure what it costs today

Before automating anything, two weeks of measurement. You need four figures: invoices received per month, hours dedicated to the full cycle (receipt, registration, matching, approval, payment, filing), incident rate (duplicates, allocation errors, late payments), and cost per invoice. The benchmark for manual processing in Spain is between 8 and 15 euros per invoice when all the time of everyone who touches it is counted.

Without this baseline, the project won't be able to prove its return later. With it, the decision usually makes itself: 500 invoices a month at 10 euros is 60,000 euros a year of pure processing.

Step 1: Centralize reception

The first bottleneck is not technological, it's fragmentation: invoices arriving at five different mailboxes, through a large client portal, scanned on paper, or in someone's personal inbox from purchasing. The solution is a single point of entry—typically a billing address that the agent monitors—and communication to vendors to redirect the flow.

This step is organizational and costs one or two weeks of persistence, but it determines the ceiling of everything else: what does not enter through the channel does not get automated.

Step 2: Data extraction and validation

This is where AI makes the difference compared to classic template-based OCR. The agent reads any format—native PDF, crooked scan, mobile photo, electronic invoice—and extracts vendor, Tax ID, dates, lines, bases, VAT types, and due dates without anyone having configured a template per vendor. It then applies automatic validations to the extracted data: Tax ID against your master vendor list, arithmetic squaring of bases and taxes, and duplicate detection against historical data.

The practical difference compared to the previous generation of tools: a template-based OCR cleanly processes 60-70% of the volume; a well-implemented agent exceeds 90-95%. If you're doubting between agent, RPA, or classic automation for this step, the comparison is in AI agent vs RPA vs automation.

Step 3: Matching with purchase order and delivery note

The heart of the circuit is matching: cross-referencing each invoice with its purchase order and its delivery note. When the three documents match, the invoice can proceed on its own; when they don't, the real work begins. An agent resolves most common discrepancies without intervention: cent differences due to rounding, partial deliveries invoiced separately, added freight, different units of measure between order and invoice.

Define explicit thresholds: for example, discrepancies up to 2% or 20 euros are approved automatically and logged; above that, the agent prepares the case with all the information and escalates it to the responsible person. This way, the team stops reviewing 500 invoices to review 30.

Step 4: Approvals that don't get stuck

The second time-eater is approvals: invoices waiting for days in the inbox of a manager who doesn't know they are there. The agent routes each invoice to the right person based on amount, cost center, and project, presents it with the context already prepared (order, delivery note, vendor history, detected deviation), and politely chases whoever doesn't respond, with reminders and escalation if the due date approaches.

The valuable side effect: no more late payment surcharges and vendor calls asking about their invoice—and early payment discounts that were previously unreachable because you never made it in time are now open.

Step 5: Accounting and payment proposal

Once the invoice is approved, the agent posts it to your ERP with the correct allocation—account, cost center, project—learned from your history and rules, and incorporates it into the payment proposal for the corresponding due date. Payment execution follows the principle of limits: below the threshold you define, it can go in the batch automatically; above it, it requires explicit human approval. An agent without amount limits is not automation, it's an operational risk.

Step 6: Measure against the baseline and expand

After 4-6 weeks in production, compare against step 0: rate of invoices processed without intervention (goal: exceed 90%), team hours (an 80-90% reduction in cycle time is a normal result), cost per invoice (from 8-15 euros to less than 2), incidents, and late payments. With those numbers on the table, the natural expansion is bank reconciliation and collections, reusing the connections already built. We develop the full reasoning of the circuit and the agent's role in accounts payable automation with AI agents, and the map of the entire finance department is in the complete guide to financial automation with AI.

The three mistakes we see repeated

• Automating chaos. If today every person approves with a different criterion, unify the criteria before automating; the agent will execute what you define, whether it's consistent or not.
• Chasing 100%. The last few odd invoices—that vendor who invoices on carbon paper—are not worth the effort. With 90-95% automated, the rest is exception management, and it's fine that it is.
• Skipping step 0. Without a baseline, in six months the savings will be a discussion of opinions instead of a figure in a report.

If you want to know what this project would return in your specific case, start with the free diagnosis of how much manual work is costing you: with your invoice volume and current hours, you'll have the estimate in euros before talking to anyone.

Frequently Asked Questions

Can I automate accounts payable without changing my ERP?

Yes, and it is recommended. The agent connects through native connectors with your systems to the ERP you already use—SAP, Business Central, Odoo, Holded, A3—and reads and writes in it just as a team member would. Changing ERP to automate is reversing the sequence.

How long does it take to automate accounts payable?

Between 4 and 8 weeks from start to production, including the two initial weeks of measurement. Factors that lengthen the timeframe are the fragmentation of reception channels and the number of systems to connect, not the volume of invoices.

What percentage of invoices are processed without human intervention?

A well-implemented agent exceeds 90-95% automatic processing, compared to the typical 60-70% for template-based OCR. The rest are real exceptions—discrepancies above the threshold, new vendors—that the agent prepares and escalates with full context.

What real savings can I expect?

The cost per invoice drops from 8-15 euros to less than 2, and the team's time in the cycle is reduced by 80-90%. In a company with 500 invoices per month, that's between 40,000 and 55,000 euros a year, plus avoided surcharges and early payment discounts that become achievable.