Invoice processing that captures, matches and codes, and only interrupts AP for exceptions
Read every invoice from any supplier, match it to purchase orders and goods receipts, code it to the ledger and post the clean ones automatically, with a traceable record for the auditor.
Accounts payable spends its time on the invoices that are fine
Invoices arrive by email, supplier portal and post, in hundreds of layouts. Someone keys the header, then the line items, then searches for the purchase order, then chases the person who ordered the goods when the quantities do not match. Month-end becomes a backlog, early-payment discounts are missed and the occasional duplicate gets paid twice.
Most of that work is deterministic once the data is extracted. The hard part was always reading the invoice reliably. With vision-capable models that part is now solvable, which means the matching and coding logic can run before a person sees anything.
How it works
From supplier email to posted ledger entry
Extraction is the only probabilistic step; matching, duplicate checks and coding are rules you control. An invoice posts without a person only when every rule passes, and every decision is archived with the source image.
1. Capture and extract
Invoices are collected from mailboxes, portals and scans. A vision model extracts supplier, number, dates, totals, tax and every line item, each with a confidence score and its position on the page.
2. Match and check
Line items are matched to the purchase order and goods receipt within your price and quantity tolerances. Duplicates are detected on content, not just invoice number. Supplier details are checked against the master record.
3. Code, route, post
General ledger account, tax code and cost centre are proposed from history and rules. Clean invoices post to the ERP; exceptions reach the AP queue with the specific failing check highlighted.
What you get
What the finance team receives
Multi-channel capture
Shared AP mailboxes, supplier portals, scanned post and e-invoicing feeds all land in one queue, with attachments separated from covering emails and multi-invoice PDFs split.
Line-item extraction
Full line-level data from any layout, including multi-page tables, using the same engine as our computer vision service, so matching happens per line rather than per total.
Matching and tolerance engine
Two- and three-way matching with tolerances by supplier, category and amount, partial deliveries handled, and duplicate detection on content fingerprints.
Exception queue for AP
A work queue showing the invoice image beside the failing check, with one-click actions: accept variance, request credit note, reassign coding, reject. Approvals route by amount.
AP performance dashboard
Straight-through rate, exceptions by cause, days to post, discounts captured, duplicates stopped and cost per invoice, reported by supplier and by month.
Guardrails we build in
Accuracy thresholds: extraction targets per field, measured on a labelled sample of your invoices before go-live; totals that do not reconcile with line items never post. Human-in-the-loop: straight-through posting applies only to invoices passing every check, with amount-based approval routing for the rest, and a weekly sample of posted invoices audited by AP. Audit trail: invoice image, extracted values, match results, coding, approver and posting reference stored together for the retention period your auditors require. PII and payment data: bank details and personal data in invoices are masked in logs and processed in your region; bank detail changes are always reviewed by a person. The full control set is described under AI security and governance, including AI data privacy.
Typical rollout
Shadow the AP team, then let the clean invoices through
Sample and rules (2 weeks)
A representative invoice sample per supplier group, the coding rules and tolerances AP applies today, and the ERP integration design.
Extract and match (3–4 weeks)
Capture, extraction, matching, duplicate detection and coding, evaluated on the labelled sample until field targets and match rates are met.
Shadow run (2–3 weeks)
Live invoices processed in parallel with the AP team. Differences reviewed daily, tolerances tuned and the exception queue refined with the clerks.
Go live (1–2 weeks)
Straight-through posting enabled for the first supplier group, approval routing switched on, dashboard live and AP trained. Further supplier groups follow in weeks, not months.
Illustrative example
What a first release usually looks like
Illustrative scenario, not a client case
A distribution business processes several thousand supplier invoices a month through a small AP team and an ERP with template OCR configured for its largest suppliers only. Everything else is keyed by hand, and month-end regularly slips. The first release covers PO-backed invoices from all suppliers. In the shadow run, the majority of invoices from regular suppliers match within tolerance and would have posted without a person; the exceptions are mostly partial deliveries and price changes that the clerks already know how to resolve. After go-live, the team works the exception queue in the morning and spends the afternoon on supplier queries and discount negotiations. Non-PO invoices, which need a different approval flow, are added in the second phase.
For finance teams in regulated environments, see how we approach fintech engagements.
Template OCR works for the suppliers whose layouts you have configured and breaks on everyone else. A vision-capable model reads any layout, including line-item tables, and the matching, coding and duplicate checks are built around your rules rather than a fixed template. The review queue only shows what failed a rule.
Does it post invoices automatically?
Only when every check passes: fields extracted with confidence, PO and receipt matched within tolerance, no duplicate, valid supplier and coding. Everything else goes to an AP reviewer. Which invoices qualify for straight-through posting is your decision and can be tightened by supplier, amount or category.
Which ERPs and finance systems do you connect to?
Any with an API or a supported file interface, through our AI integration practice. Where purchasing lives in Salesforce or a CRM-connected system, see Salesforce integrations. We do not replace the ERP; it remains the ledger of record.
What about fraud and duplicate invoices?
Duplicate detection compares supplier, number, amount, date and line-item fingerprints, not just the invoice number. Bank detail changes, round-amount patterns and new-supplier first invoices are flagged for review. For broader transaction monitoring, see fraud and anomaly detection.
How long does it take?
Six to ten weeks for the first supplier group: two for sampling and rules, three to four for build and evaluation, and two to four for a shadow run against the current AP process before straight-through posting is enabled.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.