AI invoice and receipt capture
Also called AP automation, bill capture, receipt scanning, invoice OCR, accounts payable automation
Does AI invoice and receipt capture work for a small business? Independent accuracy evidence, real costs, the data it needs and where it fails.
Capture tools read supplier bills and receipts and draft them in your accounting system: expect to stop typing, not to stop checking. The reading is well tested: in an independent 2025 benchmark, the best AI model got 96% of fields right on clean invoices but 87% on scanned receipts, and a small model managed only 54%. Even in one vendor's own high-volume deployment, roughly one invoice in four still needed more than light manual work. Entry plans start around R177 or £15 a month.
At a glance
| Status | In production |
|---|---|
| What it does | Reads supplier invoices, bills and receipts from email, phone photos or uploads, pulls out supplier, dates, amounts and VAT, and drafts the bill in your accounting system for approval. |
| Typical cost | From $13, £15.45 or R177 per month for 50 credits, excl. VAT (vendor list prices, checked Oct 2026); also bundled into some accounting plans |
| Data it needs | Every supplier document routed to one inbox, a clean supplier list with default account codes and VAT treatment, and correctly coded past bills. |
| Setup effort | Low: set up a forwarding address and connect your accounting system; the real work is tidying supplier records first. |
| Evidence | Moderate. One independent accuracy benchmark, one peer-reviewed enterprise deployment written by the vendor's own researchers, and analyst surveys; no independent small-firm outcome study. |
What is AI invoice and receipt capture?
It is software that does the data entry on the bills you receive. Suppliers email invoices, staff photograph till slips, and the tool reads each document. It pulls out the supplier, invoice number, dates, line items, totals and VAT, then creates a draft bill or expense claim in your accounting system for someone to approve.
Older tools relied on templates and optical character recognition (OCR). Newer ones use large language models that can read an unfamiliar layout without a template. Either way, the job changes from typing to checking.
It is one of the most common uses of AI in finance teams. Among finance functions that had implemented AI, a 2025 Gartner survey of 183 finance leaders found 37% using it to automate accounts payable, second only to knowledge management at 49%. That survey is not focused on small firms, so read it as a sign the technology is mature, not as small-business adoption.
How does it work?
Most setups follow the same sequence.
- Intake: a dedicated email address that suppliers or staff forward documents to, plus a phone app for paper receipts.
- Extraction: the model reads the document and fills in supplier, date, invoice number, net, VAT and total. Line-item detail often costs more; one vendor charges two credits per invoice instead of one.
- Coding: the tool suggests the supplier record, expense account and tax rate, usually based on how you coded that supplier before.
- Review and posting: a person checks the draft, fixes anything flagged and approves it, with the image attached.
Keep that image. HMRC accepts a scanned image in place of the paper original if it shows every detail needed for VAT, but data typed into software does not count as a copy of the invoice. VAT records generally have to be kept for at least six years.
What it needs to work
Capture works well only when supplier records and past bills are already clean, because the tool copies your existing coding habits and reads whatever lands in its inbox.
- One inbox or forwarding rule that receives every supplier invoice, so nothing bypasses capture
- A cleaned supplier list in the accounting system: one record per supplier, with VAT number, default expense account and tax rate
- At least the last six months of bills from regular suppliers coded correctly, so coding suggestions have a sound pattern to learn
- Purchase orders or delivery notes held in a system, if you want each invoice checked against what was ordered and received
- A named reviewer who approves drafts, and a written rule that any change of supplier bank details is confirmed by phone
- The document image stored with each bill; in the UK, VAT records are generally kept for at least six years
Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.
What does the evidence show?
The reading step is well tested. A 2025 benchmark by Fraunhofer researchers ran eight current AI models over 1,850 invoices and receipts. The best scored 96% of fields correct on clean, computer-generated invoices, 93% on scanned invoices and 87% on scanned receipts; the weakest small model managed 54% on receipts.
At scale, IBM researchers reported that their own system handled 76% of about 80,000 invoices with low or no manual intervention at two clients. That is a vendor describing its own product at high volumes, not small businesses.
The problem it targets is real in both markets. In a UK government survey, 36% of businesses blamed administrative errors for paying suppliers late and 23% blamed technical issues such as lost invoices. By the end of the third quarter of 2024/25, South African national departments had paid 38,153 invoices after the 30-day limit set by the Public Finance Management Act, citing misfiled, misplaced or unrecorded invoices among the reasons.
We found no independent study measuring time saved or error rates in small businesses using these tools.
AI capture vs manual capture
| Compared on | AI capture | Manual capture |
|---|---|---|
| Speed per invoice | Seconds to read; a person still reviews each draft | Minutes of typing per invoice, longer with line items |
| Error profile | Misreads on poor scans and fields such as bank numbers; errors repeat if past coding was wrong | Typos and transposed digits; errors are random rather than systematic |
| Cost model | Monthly subscription by document volume or bundled into accounting plans | Staff or bookkeeper time |
| Best for | Steady volumes of emailed PDF invoices from repeat suppliers | Very low volumes, or mostly handwritten and one-off documents |
Status and evidence
Rated production because capture is sold off the shelf, including inside standard accounting plans, and independent testing shows current models read most invoice fields correctly. Evidence is moderate, not strong: the benchmark measures reading accuracy rather than business outcomes. The only large deployment figures come from a vendor's own researchers at two high-volume clients, and we found no independent study of time saved in small businesses.
Where does it go wrong?
- Poor photos. Accuracy drops on scanned receipts compared with clean PDFs, and bank account numbers (IBANs) were read noticeably worse than other fields. Till slips need more checking than emailed invoices.
- Copying bad habits. Coding suggestions learn from your history. If past bills were coded inconsistently, the tool repeats the mess faster.
- Duplicates. The same invoice arrives by email and again on a statement, or is forwarded twice. Without a check on supplier plus invoice number, both get paid.
- Changed bank details. The tool reads whatever is on the document, including a fraudster's account number. Confirm any change of supplier bank details by phone on a number you already hold.
- Rubber-stamping. When most drafts are right, reviewers stop looking. Spot-check totals and VAT every week.
- Nobody owns it. In a study of 12 small firms given shared support to automate their accounting, the support helped them see the value but did not make the tools feel easier to use.
Numbers it should move
Ask the demo business
Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.
We take about 49 days to pay suppliers, in line with 45-day terms. But customers take about 42 days to pay us and stock sits for about 60 days, so we fund roughly 53 days of trade from our own cash: the gap between paying for goods and being paid for them.
How this was worked out
Money owed to suppliers at month end, divided by the last 91 days of supplier bills, multiplied by 91. The first two months are left out while the bill history builds up. Days of trade funded = debtor days + stock days − creditor days.
Where the data usually lives
Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.
Frequently asked questions
How much does receipt and invoice scanning software cost?
List prices are modest. One capture vendor charges $13, £15.45 or R177 a month excluding VAT for 50 credits, where a simple invoice uses one credit and an invoice with line items uses two. Xero's South African plans include a document capture feature on every tier, from R450 a month including VAT, though the cheapest plan caps bill entry. The bigger cost is the time spent cleaning supplier records and reviewing drafts.
Do I need a developer or IT team to set up invoice capture?
No. Documents reach the tool by email forwarding, upload or a phone app, and cloud accounting systems connect through a sign-in screen. What you need is a bookkeeper who knows your chart of accounts and can tidy supplier records before you switch on automatic coding. If you run a desktop accounting package, check the capture tool supports it before you buy.
How accurate is AI at reading invoices?
On clean, computer-generated invoices the best current model got 96% of fields right in an independent 2025 benchmark; on scanned receipts the best was 87%, and a small open-source model got only 54%. A field only counts if it is exactly right, so even a 96% score leaves mistakes on many documents. Plan for a person to review every draft until you trust each supplier's pattern.
Can I throw away paper receipts after scanning them?
In the UK, HMRC says you can discard the paper original if the scanned image contains all the details needed for VAT, although some records, such as import VAT certificates, must be kept in their original form. Records generally need keeping for at least six years. Typed data is not a copy of the invoice, so always keep the image. South African rules differ, so check with your accountant.
Does invoice capture software need to comply with POPIA or UK GDPR?
Yes, where documents contain personal information. Under UK GDPR, information about a limited company is not personal data, but details of sole traders and named contacts are. POPIA goes further: personal information can include an identifiable, existing juristic person, so company details can count too. POPIA also requires a written contract obliging the capture provider to keep the data secure. This is general information, not legal advice.
More in cash & finance
Sources
- 1Multi-Modal Vision vs. Text-Based Parsing: Benchmarking LLM Strategies for Invoice ProcessingarXiv (Fraunhofer IAIS and Lamarr Institute), 2025Study
- 2AI Driven Accounts Payable TransformationProceedings of the AAAI Conference on Artificial Intelligence (IBM Research authors), 2022Vendor
- 3Gartner Survey Shows Finance AI Adoption Remains Steady in 2025Gartner, 2025Survey
- 4Late payments research: understanding variations in payment performance and practices across business sectors and sizes (HTML executive summary)Department for Business and Trade, 2024Government
- 5Government departments commended for timeous payment of supplier invoiceSAnews (South African Government News Agency), 2025Government
- 6Collaborative Ecosystems for Increasing Automation in Accounting Processes in Small FirmsIFIP Advances in Information and Communication Technology (PRO-VE 2024), 2024Study
- 7Pricing - Data Entry Automation Software | AutoEntryAutoEntry, 2026Vendor
- 8Pricing Plans | Xero ZAXero, 2026Vendor
- 9Record keeping (VAT Notice 700/21)HM Revenue & Customs, 2024Government
- 10What is personal data?Information Commissioner's Office, 2026Government
- 11Protection of Personal Information Act 4 of 2013Government of South Africa, 2013Government
This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.
- 1Multi-Modal Vision vs. Text-Based Parsing: Benchmarking LLM Strategies for Invoice ProcessingarXiv (Fraunhofer IAIS and Lamarr Institute), 2025Study
- 2AI Driven Accounts Payable TransformationProceedings of the AAAI Conference on Artificial Intelligence (IBM Research authors), 2022Vendor
- 3Gartner Survey Shows Finance AI Adoption Remains Steady in 2025Gartner, 2025Survey
- 4Late payments research: understanding variations in payment performance and practices across business sectors and sizes (HTML executive summary)Department for Business and Trade, 2024Government
- 5Government departments commended for timeous payment of supplier invoiceSAnews (South African Government News Agency), 2025Government
- 6Collaborative Ecosystems for Increasing Automation in Accounting Processes in Small FirmsIFIP Advances in Information and Communication Technology (PRO-VE 2024), 2024Study
- 7Pricing - Data Entry Automation Software | AutoEntryAutoEntry, 2026Vendor
- 8Pricing Plans | Xero ZAXero, 2026Vendor
- 9Record keeping (VAT Notice 700/21)HM Revenue & Customs, 2024Government
- 10What is personal data?Information Commissioner's Office, 2026Government
- 11Protection of Personal Information Act 4 of 2013Government of South Africa, 2013Government