The verdict in three sentences
For a London SME processing 1,500 supplier invoices a month, a custom AI extraction pipeline connected to the ERP costs between GBP 10,000 and 26,000 excluding VAT and pays for itself in 12 to 18 months. An off-the-shelf OCR tool is enough as long as your invoices are standard and you accept correcting 15 to 25% of fields by hand. Once you deal with multi-line invoices, cost-centre coding or a less common ERP, custom extraction reaches 92 to 97% of fields captured without correction and frees around 60 hours a month.
Off-the-shelf OCR or custom AI extraction: what you are actually buying
Market tools (Dext, AutoEntry, Xero and Sage capture modules, ERP OCR add-ons) read invoice headers well: supplier, date, number, net amount, VAT, gross. They become fragile on line items, poor-quality scans, credit notes and cost allocations specific to your chart of accounts. Custom AI extraction combines a multimodal language model, business rules (purchase order matching, VAT checks, duplicate detection) and direct posting into the ERP.
| Criterion | Off-the-shelf OCR | Custom AI extraction |
|---|---|---|
| Upfront cost | GBP 0 to 2,600 of setup | GBP 10,000 to 26,000 |
| Subscription or running cost | GBP 260 to 780/month for 1,500 invoices | GBP 65 to 195/month API + GBP 170 to 350 maintenance |
| Header fields correct without edits | 90 to 95% | 96 to 99% |
| Line items correct without edits | 60 to 80% | 90 to 96% |
| Purchase order matching | Basic, ERP-dependent | Custom rules (price and quantity variances) |
| ERP connection (Sage, Xero, NetSuite, SAP B1) | Standard connectors | Direct posting, cost codes included |
| Time to go live | 2 to 4 weeks | 2 to 3 months |
These are 2026 order-of-magnitude figures for an SME with 50 to 250 staff. The gap lies in the line items: if your accounting splits each invoice across several cost centres, that is where the keying time hides.
What each building block costs
A serious AI extraction project is not just 'plugging in a model'. You need a test set of invoices, field-by-field accuracy measurement, a validation screen for doubtful cases and a full audit trail for your auditor.
| Building block | 2026 budget (GBP excl. VAT) | Duration |
|---|---|---|
| Scoping, 300-invoice sample, baseline measurement | 1,300 to 2,600 | 1 to 2 weeks |
| Extraction pipeline (OCR, AI model, VAT rules) | 3,500 to 7,800 | 3 to 5 weeks |
| Purchase order and goods receipt matching | 1,700 to 5,200 | 2 to 3 weeks |
| Validation screen for low-confidence invoices | 1,700 to 4,300 | 2 weeks |
| ERP connector and audit log | 1,700 to 4,300 | 2 to 3 weeks |
| Acceptance testing, training, 1-month hypercare | 400 to 1,800 | 2 weeks |
| Total | 10,000 to 26,000 | 2 to 3 months |
API cost ranges from GBP 0.04 to 0.13 per invoice depending on page count and model. For 1,500 invoices, budget GBP 65 to 195 a month. As more suppliers send structured e-invoices (Peppol, UBL) ahead of the UK e-invoicing roadmap, the pipeline should read those directly without the AI step, which brings this cost down year after year.
The controls that drive accuracy
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Claimed accuracy only matters if it is measured on your invoices. Require a confidence score per field: above the threshold (say 0.9), the invoice is posted; below it, it goes to a validation queue. Also require an arithmetic check (line totals equal net amount, VAT consistent with the rate), duplicate detection by supplier and amount, and retention of the original PDF linked to the posting. These safeguards make the automation acceptable to your external auditor.
Mini case study
Claire, finance director of a 140-person manufacturing SME in London, receives 1,500 supplier invoices a month. Keying takes 4 minutes per invoice on average, so 100 hours a month. After rolling out AI extraction for GBP 19,000, keying and validation drop to 40 hours: 60 hours saved. At a loaded cost of GBP 30 an hour, that is GBP 1,800 a month. Subtract GBP 130 of API and GBP 260 of maintenance: a net gain of GBP 1,410 a month. The project pays back in 13.5 months, then returns around GBP 16,900 a year, not counting a month-end close brought forward by 2 days.
FAQ
What accuracy should we target before going live? Aim for at least 95% of header fields and 90% of line items correct on a sample of 300 real invoices. Below that, the validation queue absorbs too much time and the saving falls under 30 hours a month.
Will e-invoicing make AI extraction pointless? Not in the short term: in the UK, structured e-invoicing is still largely voluntary in 2026 and 50 to 80% of SME supplier invoices arrive as PDFs or scans. The pipeline handles both formats.
Does our invoice data go to an AI vendor? You can choose a UK or EU-hosted model with a no-training commitment, or a self-hosted model for GBP 260 to 700 a month of infrastructure. The choice is made during scoping according to your UK GDPR policy.
Which ERPs are compatible? Sage 200, Xero, NetSuite, SAP Business One, Odoo and most ERPs with an API or structured import. An ERP without an API adds GBP 1,700 to 3,500 for a controlled file import.
What happens if a supplier changes its invoice layout? The AI model works without templates, unlike zone-based OCR. Accuracy usually drops by less than 3 points and the validation queue catches the doubtful cases.
Let's scope your project. Send us a sample of 50 anonymised invoices: we measure achievable accuracy and price a scope with a 2 to 3 month timeline. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
