How to Automate Invoice Processing Without Power Automate

Invoice processing is where most UK SMBs first try automation. It is also where many of them hit a wall. The usual path leads to Power Automate and AI Builder because those tools ship with Microsoft 365. The assumption behind that path is that you are already paying for the Microsoft stack, your team already knows it, and Dataverse is the place your business data lives.

For a lot of businesses, none of that is true. You run Xero on Google Workspace. Your finance team uses Gmail and Drive. You do not want another Microsoft licence, you do not want to build on Dataverse, and you do not want to pay per-page AI Builder credits that scale linearly with invoice volume. This post walks through the alternative: a non-Microsoft stack using n8n, a modern OCR or vision model, and your existing accounting system. It also covers the cases where Power Automate is still the right call, because those cases exist.

When Power Automate Is the Wrong Choice for Invoice Automation

Power Automate is the wrong choice when your business does not live inside Microsoft 365, or when your invoice volume makes AI Builder credits cost more than a custom build. Both situations are common in UK SMBs.

We already covered the Power Automate and AI Builder version of this build in an earlier post. It works well for businesses on Microsoft 365 Business Standard or above, with invoices arriving in Outlook and data landing in SharePoint or Dataverse. If that describes you, start there.

If it does not, three problems show up early.

First, the Microsoft licensing. AI Builder credits sit on top of your Power Platform licence. A Per App plan runs around £4 per user per month before you add the AI Builder add-on. For a five-person finance team processing 300 invoices a month, the licence cost alone can exceed the per-invoice API cost of an n8n build.

Second, the extraction model. AI Builder’s invoice processing prebuilt model works well on standard formats, but handles poorly on hand-written notes, unusual supplier layouts, and invoices with multiple line-item tables. In our experience, extraction accuracy drops sharply once you move past typical B2B supplier invoices.

Third, the lock-in. If your data needs to move to Dataverse before it gets to Xero or Sage, you have added a hop. Every Microsoft-specific component makes it harder to migrate later.

The Non-Microsoft Stack You Need

Three components do the heavy lifting in a non-Microsoft invoice automation build.

• An orchestration platform. n8n is our default because it self-hosts, handles complex logic without breaking, and its credit cost is effectively zero on self-hosted. Make is the alternative for teams that want a managed cloud service.
• An OCR or vision model. Mistral OCR, Google Document AI’s Invoice Parser, GPT-4o Vision, and Claude Sonnet with vision all work. Which one depends on invoice variety and volume, which we cover in the comparison section below.
• Your accounting system’s API. Xero, QuickBooks Online, and Sage Accounting all have usable APIs. This is where the extracted invoice data ends up.

The n8n workflow development work we handle for finance teams almost always uses some version of this three-layer pattern. The value is that you can swap the OCR layer without rebuilding the orchestration, and you can swap the accounting integration without touching the extraction logic.

A fourth component is optional but recommended. A validation layer that catches extraction errors before they write to your accounting system. This can be a spreadsheet-based review queue, an Airtable base, or a Slack approval workflow. The point is that the AI never writes directly to your ledger on day one.

How the Workflow Fits Together End to End

The workflow has five stages, triggered by an invoice arriving as an email attachment or a file drop in a shared folder. Each stage has a specific job and fails in predictable ways when things go wrong.

The five stages are ingestion, extraction, validation, enrichment, and delivery. Ingestion watches your inbox or drive folder and picks up new files. Extraction pulls structured fields from the invoice image or PDF. Validation checks the extracted data against rules you define (supplier exists, VAT rate matches, totals reconcile). Enrichment adds context the OCR cannot see, such as general ledger coding or PO matching. Delivery writes the invoice to your accounting system and notifies whoever needs to approve payment.

n8n handles all five in a single workflow. A webhook or email trigger starts ingestion. An HTTP node calls the OCR or vision API. A function node runs validation logic. A lookup node matches suppliers or POs. A final HTTP node pushes to Xero or QuickBooks.

If you are choosing between orchestration platforms, the comparison of how Make, Zapier and n8n compare for AI-heavy workflows covers the trade-offs in detail. For invoice processing specifically, n8n’s error handling and conditional logic make it the right fit at scale. Zapier hits cost walls fast once you cross 500 tasks a month. Make sits between the two.

One pattern to apply from day one. Build the workflow with an approval queue before the final write step, even if you plan to remove it later. Most teams leave human approval in place for invoices above a value threshold anyway, so the queue pays for itself.

OCR and Extraction Options Compared

Four tools cover most invoice extraction needs in 2026. Each has trade-offs around accuracy, cost, speed, and handling of non-standard invoice formats.

• Mistral OCR is the cheapest option for pure text extraction at scale. It does not reason about invoice structure, so you need a second pass with a language model to turn text into fields.
• Google Document AI’s Invoice Parser is the most accurate on standard B2B invoices and returns structured fields directly. It costs more per page and takes longer to set up.
• GPT-4o Vision and Claude Sonnet vision both handle extraction and reasoning in one step. They cost more per invoice than Mistral OCR but eliminate the second pass.

The broader UK document processing guide covers extraction accuracy benchmarks across more document types if you process documents other than invoices.

Numbers here are typical ranges from builds we have shipped and public pricing at time of writing. Your real costs depend on image size, token output, and whether you need multiple passes for validation.

The pattern we most often recommend for UK SMBs is Claude Sonnet vision or GPT-4o Vision when invoice volume is below around 2,000 per month, and a Mistral OCR plus LLM split when volume is higher. Google Document AI wins when your invoice formats are highly standardised and you need the prebuilt parser’s structured output.

How to Handle UK VAT and Multi-Currency Invoices

UK invoice automation has two compliance points that generic invoice processors handle badly. VAT rate validation and foreign currency conversion. Both need rules that live in your workflow, not in the OCR model.

For VAT, your workflow must check three things. That the supplier is VAT registered (match the VAT number against your supplier database). That the VAT rate applied is correct for the goods or services (standard, reduced, zero, exempt, or outside scope). That the VAT amount calculates correctly from the net (a 20% VAT on £100 net should be £20, not £19.95 because of rounding upstream). Getting any of these wrong creates a Making Tax Digital reporting problem.

For multi-currency invoices, your workflow needs the exchange rate at invoice date, not at processing date. HMRC’s monthly exchange rates are the accepted benchmark for VAT purposes if you use period-end conversion. Xero and QuickBooks both handle the conversion if you pass the invoice date and currency correctly in the API call. Sage Accounting is more rigid and needs the GBP amount calculated upstream.

Accountancy firms handling invoice volume on behalf of clients typically build these rules once and reuse them across the whole client base. For a single business, a rules table in Airtable or PostgreSQL works. Map supplier to default VAT treatment, map currency to exchange rate source, and let the workflow check each invoice against those defaults before it writes. One failure mode to plan for: EU suppliers sometimes apply their local VAT to UK customers when the reverse charge should apply instead, so flag any non-UK supplier invoice with VAT on it for human review until you trust the rules.

What This Build Costs Per Invoice

Per-invoice economics are where the non-Microsoft stack most often beats Power Automate. Three inputs drive the headline number.

• OCR or vision cost. Between £0.002 and £0.030 per invoice based on the options in the comparison table above.
• Orchestration cost. Effectively zero on self-hosted n8n. Around £0.01 to £0.02 per invoice on Make. Around £0.04 to £0.08 per invoice on Zapier above the free tier.
• Infrastructure and storage. Around £15 to £40 per month for a self-hosted n8n instance plus PostgreSQL. This is a flat cost, so it drops per-invoice as volume rises.

For a business processing 300 invoices a month, the total cost per invoice on a self-hosted n8n plus Claude Sonnet vision stack typically sits between £0.20 and £0.35 including infrastructure amortisation. At 1,000 invoices a month, that drops to between £0.05 and £0.10 per invoice because the infrastructure cost spreads over more volume.

Compare that to a Power Automate plus AI Builder build at the same volume. Between the Per App plans for your finance team and the AI Builder credits for invoice processing, typical total cost runs £0.40 to £0.70 per invoice at 300 per month. The cross-over where Power Automate makes economic sense is usually when your business already has E3 or E5 Microsoft 365 licences that include enough Power Platform allowance to cover the workflow.

Our general cost guide for AI automation builds in the UK covers the build cost side. Per-invoice running costs are only half the economics. A typical invoice automation build costs between £4,000 and £12,000 to deliver, depending on how many accounting systems, approval paths, and validation rules you need. Payback on that build for a business processing 300 invoices a month at 15 to 20 hours of manual finance time saved per week is usually three to five months.

Common Failure Modes and How to Handle Them

Invoice workflows fail in predictable ways. The most common modes are extraction errors on unusual formats, duplicate invoice detection, supplier mismatch, VAT rate errors, and API timeouts writing to the accounting system. Each has a handling pattern.

Extraction errors on unusual formats are the most common failure in the first month of production. A supplier sends a PDF with a scanned image inside, or a receipt instead of an invoice, or a credit note formatted as an invoice. The handling pattern is confidence scoring. Your language model prompt should return a confidence score per extracted field, and any invoice with a score below your threshold routes to the human review queue.

Duplicate invoice detection catches suppliers sending the same invoice twice, usually because a chase email restates the original. Match on invoice number plus supplier plus net amount. If any two of those three match an existing invoice in the last 90 days, flag as potential duplicate and hold for review.

Supplier mismatch happens when the extracted supplier name does not match any supplier in your accounting system. New supplier, typo in extraction, or a subsidiary name. The handling pattern is fuzzy matching plus a “new supplier” queue that your finance team reviews weekly.

API timeouts writing to Xero, QuickBooks, or Sage are rarer but harder to recover from if you do not plan for them. n8n has built-in retry logic with exponential backoff. Set the retry count to at least 5, with delays starting at 30 seconds. Queue the invoice for manual re-send if all retries fail.

Layer observability on top of all of this from day one. A dashboard showing daily invoice count, extraction success rate, validation pass rate, and average time from receipt to ledger write is enough. Without it, the first sign of a failing workflow will be a finance team member complaining about missing invoices a month later.

When You Should Still Use Power Automate

Power Automate is the right choice in three situations.

• Your business runs on Microsoft 365 E3 or E5 licences that include Power Platform allowance. The marginal licence cost of adding Power Automate to a finance workflow is zero.
• Your invoice data must live in Dataverse or SharePoint for other reasons, such as an existing Power Apps finance application. Adding a second platform doubles the operational surface area.
• Your finance team is already trained on Power Automate and the cost of retraining them on n8n outweighs the per-invoice savings.

In those cases, building on Power Automate removes friction and uses licensing you are already paying for. The per-invoice cost difference matters less when the marginal licence cost is zero.

What should give you pause. Businesses that start on Power Automate often outgrow AI Builder’s credit model at around 500 to 800 invoices per month, depending on complexity. At that point the choice is to move up Microsoft’s pricing tier or rebuild on a different stack. Rebuilding on a non-Microsoft stack at that volume is harder and more disruptive than starting there from the beginning.

If you are reading this post, you already have a clear idea of which side you are on. If you want a second opinion on your specific volume, format variety, and existing stack, the invoice processing builds we deliver end to end cover both Microsoft and non-Microsoft approaches depending on what fits the business.