Can AI Help Me Check Whether Invoices, Purchase Orders and Delivery Notes Match?

8 October 2026

What does AI actually check?

The basic process is called three-way matching. The system compares the supplier invoice with the approved purchase order and the goods received note or delivery note. It checks whether the business ordered the item, whether the item arrived, and whether the supplier charged the agreed amount. For services, the third record may be a completed job sheet, approved timesheet or manager confirmation rather than a delivery note.

Modern document tools can read PDFs, scans, emailed attachments and photographed paperwork. They extract fields such as supplier name, invoice number, purchase order number, product code, quantity, unit price, VAT, delivery date and total. Rules then compare those fields with records in Xero, Sage, QuickBooks, an ERP or a job-management system. AI is most useful where supplier layouts vary or descriptions are inconsistent. A traditional rule may fail because one document says '20 x filter cartridge' and another says 'cartridges, pack of 20'. An AI model can suggest that these probably refer to the same item, while still marking the match for review if confidence is low.

The word 'check' matters. The system should produce one of three outcomes: matched within agreed tolerances, clearly mismatched, or uncertain. It should not silently force documents to agree. A quantity difference, duplicate invoice number, unexpected freight charge, changed bank details, missing purchase order or VAT discrepancy should enter an exception queue with the source documents attached.

HMRC says a UK invoice must include a unique identification number, supplier and customer details, a clear description, supply date, invoice date, amounts, VAT where applicable and the total owed. Those required fields provide a practical minimum extraction checklist. See GOV.UK guidance on invoice requirements.

What a safe matching workflow looks like

A safe workflow starts when an invoice reaches a controlled inbox or supplier portal. The system saves the original file, records when it arrived and checks whether the invoice number has appeared before. It then extracts the key fields and looks for a purchase order in the accounting, procurement or job-management system. If it finds one, it retrieves the order lines and the receipt record. The comparison happens at line level, not just on the grand total.

You define tolerances before automation begins. A business might allow a rounding difference of 1p, but require review for any quantity difference. It might accept a delivery charge up to £15 only when the purchase order explicitly permits carriage. It might reject every invoice without a purchase order, while sending approved utilities and rent through a separate recurring-cost route. These are business rules, not decisions the model should invent.

A useful exception queue explains the problem in plain English: 'Invoice shows 12 units, purchase order shows 10, delivery note confirms 10' or 'Unit price is £48, approved price is £44'. The reviewer should be able to open all three records, add a note, ask the supplier or buyer a question, and either approve, reject or return the item. Every action needs a timestamp and named user.

The strongest design also separates duties. The person who raises a purchase order should not be able to create a supplier, alter bank details and release the payment alone. AI can prepare the evidence, but it should not collapse these controls. Give integrations the narrowest permissions they need. Read-only access to bank transaction data may support reconciliation, but the matching tool does not need permission to send money.

Start in observation mode for two to four weeks. Let the workflow make recommendations while staff continue the existing checks. Compare the result, tune the tolerances and record false matches before allowing any straight-through processing.

What can it catch, and what will it miss?

AI-assisted matching is good at repetitive discrepancies. It can spot duplicate invoice numbers, duplicate amounts from the same supplier, missing purchase order references, incorrect quantities, price differences, VAT calculation errors, invoices received before delivery and delivery notes that do not cover every invoiced line. It can also flag an unusual document layout or a supplier name that is similar to, but not exactly the same as, an approved supplier.

It is less reliable when the documents do not describe the purchase consistently. A purchase order for 'maintenance visit' may result in an invoice containing labour, mileage and replacement parts. Partial deliveries and consolidated monthly invoices can also require context. The same is true for weight-based goods, foreign currency purchases, retrospective price changes, credit notes and substitutions agreed by telephone. The system can assemble the evidence, but a buyer or accounts person may still need to decide whether the charge is valid.

Document quality matters. Folded delivery notes, handwriting, faint scans and mobile photos taken at an angle reduce extraction accuracy. A confidently extracted wrong digit is more dangerous than a field the system marks as unreadable. Require confidence scores, retain the original image and test the software against your worst real documents, not the supplier's demonstration set.

Fraud checks also need separate rules. Three-way matching does not prove that a bank-account change is genuine or that the person who placed the order was authorised. Treat new supplier details and bank changes as high risk. Confirm them through a known telephone number or another independent channel. Never use contact details taken only from the change request itself.

The goal is not a fictional 100 per cent automation rate. A healthy result may be that 60 to 80 per cent of routine invoices pass the configured checks, while the remainder reach people with the evidence already organised. The precise rate depends on your documents and process, so demand a pilot result rather than accepting a vendor benchmark.

Which tools and approaches should a UK small business consider?

If you already use mainstream accounting software, begin with its document-capture and purchasing features. Xero Hubdoc, Sage Accounting, Dext and AutoEntry can capture invoice data, although the depth of purchase-order and receipt matching varies by product and plan. Staying inside an existing platform can be cheaper and easier to support than building a new workflow.

A low-code route can work where documents arrive consistently and the systems have reliable APIs. Microsoft Power Automate with AI Builder, Make or n8n can watch an inbox, extract fields, look up an order and create an exception task. This is flexible, but somebody must own authentication, error handling, logging and changes when an API or document layout changes. A cheap workflow with no owner becomes expensive the first time it stops quietly.

Specialist accounts-payable platforms offer stronger approval routing, duplicate detection and audit trails. They make more sense when volumes are high, multiple entities share a finance team, or purchasing controls are already mature. Ask every vendor to demonstrate line-level matching, partial deliveries, credit notes, tolerance rules, duplicate handling, UK VAT and export of the full audit history. Also ask what happens when you leave, including how you recover documents, decisions and configuration.

Our bias is towards the smallest system that solves the verified problem. If the existing accounts package can remove most rekeying, use it. If matching requires data from email, purchasing and a job system, a managed integration may be justified. Custom AI is sensible only when the document mix, approval logic or operational systems create a genuine gap that packaged software cannot cover.

How much does it cost, and when does it pay back?

For a small UK business, basic document capture may already be included in an accounting subscription or cost roughly £20 to £100 a month. A low-code workflow connecting an inbox, accounting package and approval tool commonly costs £2,000 to £8,000 to design, test and document. A more involved implementation with several entities, job systems, line-level matching and formal approval controls can cost £8,000 to £25,000. Specialist platforms may charge per document, per user or by annual contract, so ask for the total cost at your actual monthly volume.

Include internal costs. Someone must define tolerances, clean supplier records, test awkward documents, train reviewers and own the process after launch. Budget for monitoring and maintenance rather than treating the setup fee as the entire cost. A practical allowance is 10 to 20 per cent of the original implementation cost each year where the workflow has custom integrations.

Calculate payback from real handling time. Suppose 600 invoices arrive each month and the current process averages six minutes per invoice. That is 60 staff hours. If automation reduces 70 per cent of them to one minute and leaves 30 per cent at six minutes, the monthly workload falls to about 25 hours, saving roughly 35 hours. At a fully loaded staff cost of £25 an hour, that is £875 a month before counting fewer duplicate payments, faster month-end work or better supplier relationships. A £5,000 setup would recover its cost in about six months if those savings appear in practice.

Do not count every saved minute as cash unless the team can use that capacity productively. Measure actual outcomes: invoices processed per hour, percentage matched without rekeying, exception rate, false-match rate, duplicate invoices stopped, time to approve, overdue supplier queries and hours spent at month end.

What records and controls must you keep?

Automation does not remove your record-keeping duties. HMRC says VAT-registered businesses must retain all invoices they issue and receive, and keep VAT records for at least six years. It also says a delivery note, statement or pro-forma invoice is not enough on its own to reclaim VAT. Read the current HMRC guidance on keeping VAT records and confirm the rules that apply to your business with your accountant.

Keep the original documents, extracted fields, comparison result, confidence score, rule version, reviewer decision and any correction. If a supplier sends a revised invoice, preserve the relationship between the old and new versions. If the workflow changes, record when the new rule took effect. This gives your accountant, auditor or manager a defensible trail rather than a mysterious green tick.

Personal data may appear in invoices and delivery records, including sole trader names, home addresses, telephone numbers and employee details. Decide the lawful purpose for processing it, restrict access and set appropriate retention. The ICO's guidance on AI and data protection highlights accountability, transparency, lawfulness, accuracy and fairness. The ICO notes that its guidance is under review following the Data (Use and Access) Act, so check the current position before implementation.

Finally, create a manual fallback. Staff need to know how to receive, check and approve invoices if the AI service, integration or accounting system is unavailable. Test that fallback at least once. A workflow is not controlled merely because it usually works.

When this is NOT right for you

Do not automate three-way matching if your business does not consistently raise purchase orders, record deliveries or assign payment approval. You would be automating around missing evidence. Introduce a simple purchasing process first: approved supplier, numbered purchase order, receipt confirmation and named approver.

It is also a poor fit where most purchases are unique, negotiated after delivery or impossible to express as clear order lines. In those cases, better document capture and a well-designed approval checklist may deliver more value than elaborate AI matching.

Do not proceed if the proposed tool needs broad access to email, files, accounts and banking without a clear reason. Do not let it update supplier bank details or release payments. Avoid a vendor that cannot show confidence scores, error logs, data retention terms, export options and a human review route.

If you have fewer than about 50 invoices a month and the current process is reliable, the financial case may be weak. A built-in capture tool and a monthly duplicate check could be enough. If you are unsure, run a four-week sample before buying anything. If you want an independent view, we can help map the workflow and test the numbers. No pressure, and no recommendation to automate unless the evidence supports it.

Is This Right For You?

This is a good fit if your team receives at least 100 supplier invoices a month, repeatedly types the same details into accounts software, or loses time chasing purchase orders and proof of delivery. It is particularly useful where most purchases follow a repeatable process and every payment already has a named approver.

It is probably not right for you if you receive only a handful of invoices, rarely use purchase orders, or cannot reliably record what was delivered. In that situation, fix the purchasing process first. AI cannot match three documents when one of them does not exist. It is also the wrong starting point if you expect software to approve payments independently. Keep bank access, supplier bank-detail changes and final payment approval outside the AI workflow.

If you want to test whether this would save enough time to justify the cost, start with 50 recent invoice packs. Measure how many match cleanly, how many need judgement and how long the team currently spends on each one. That will give you a more honest answer than a software demonstration.

Frequently Asked Questions

Can AI approve supplier payments automatically?

It can recommend that a well-defined, low-risk invoice has passed matching rules, but a person should retain final payment authority. Supplier creation, bank-detail changes and payment release should have separate controls.

Does this work if suppliers send paper invoices or poor scans?

Sometimes, but accuracy falls with handwriting, faint print, folds and angled photographs. Test the system on your worst real documents, require confidence scores and route uncertain fields for review.

Can AI match partial deliveries and split invoices?

Yes, if the purchase order and receipt records identify lines and quantities clearly. Partial deliveries are more complex, so the workflow must maintain the remaining open quantity and prevent the same receipt being matched twice.

Will I still need purchase orders?

Yes. Three-way matching depends on an approved order record. If you do not use purchase orders, AI can still capture invoice data and check duplicates, but it cannot prove that the purchase was authorised on the agreed terms.

Which accounting packages can this connect to?

Common UK options include Xero, Sage and QuickBooks, alongside specialist accounts-payable and job-management systems. The deciding factor is whether your plan provides secure API access to purchase orders, receipts, suppliers and invoice records.

How long does a small implementation take?

A focused pilot often takes two to six weeks, including process mapping, sample testing, permissions, exception rules and staff training. Multiple systems, poor supplier data or complex approval routes can extend this to two or three months.

What is the biggest risk in automated invoice matching?

The biggest risk is a false match that looks authoritative. Reduce it with line-level checks, conservative tolerances, confidence thresholds, human approval, audit logs and regular sampling of invoices the system marked as clean.