Finance process automation
Automate the steps of a finance process that never needed a person, and keep the one that does.
Most finance processes are not slow because the decision is hard. They are slow because the information arrives as scans, PDFs, spreadsheets and phone photos, gets re-keyed, gets checked by a second person, and then has to be explained months later. We automate that path end to end: the intake reads itself into your schema, every field is validated against arithmetic and reference data, the pipeline scores its own confidence, and only the weak items reach a reviewer. Every correction lands in a record you can replay.
Where it fits
The problem
The work that eats the month is the work nobody chose to do. Analysts re-key figures out of bank statements and invoices, then a second person checks the typing. Volume peaks at close. Templates break the day a vendor changes its layout, and the audit question about how a number reached the ledger has no answer beyond a folder of PDFs.
What we build
A pipeline that classifies what arrived, extracts fields with layout and text models, and runs validation rules over the output before anything moves. Totals have to reconcile, dates have to fall in range, counterparties have to match your master data. Failures carry a reason code and the cropped region they came from, so the exception is faster to settle than the original typing.
How we engage
Senior engineers build it with your own documents and your own process in the room, and it runs inside your environment where the files already live. Weights, rules and the review interface ship as source you own. Your team can add a document type or retune a threshold without a change request to us.
What we build
Six stages of an automated process a controller can defend in an audit.
Ingest and classify
Documents land from mail, upload or a scanner drop and get deskewed, split into logical documents and typed before extraction runs. A phone photo of a payslip and a multi-part loan pack take different routes. Unrecognised types park in a queue rather than being forced into the nearest template.
Field extraction
Layout aware models read tables, stamps and handwriting alongside plain text, so a figure keeps its row and column meaning. Output maps to your schema with the page, coordinates and source text for each field. Nothing is inferred silently. A field the model cannot find comes back empty with a reason.
Validation rules
Extraction is checked before anything reaches your system. Line items have to sum to the stated total, IBANs have to pass checksum, dates have to sit inside the statement period, and names have to match counterparty records. A rule that fails names the field, the expected value and the document region.
Confidence scoring
Each field carries a score from model probability, validation results and agreement between passes. One global threshold sends either too much to review or too little. Thresholds sit per field, because a tolerance that suits a description column is wrong for a payment amount. Documents clear straight through, queue for review, or get rejected with the reason recorded.
Review queue
Reviewers see the cropped image beside the extracted value and correct it in one keystroke, which replaces the second pass of re-typing. The queue sorts by value at risk and by deadline. Every correction is stored as a labelled example, which feeds retraining and shows where a template or a rule needs work.
Audit trail
For each processed document the pipeline keeps the original file hash, model version, rule set version, field level scores, who reviewed what and when. An auditor asking how a figure reached the ledger gets the chain in one query instead of a folder search and a conversation.
See extraction, confidence and review on sample statements
The sample runs use synthetic documents, including deliberately poor scans, so you can watch low confidence items route to review.
How we work
Discover
Systems, constraints, and the regulation you operate under get mapped before implementation starts, so the design accounts for what already runs and for what examiners will ask about.
Architect
A design that fits your stack. Integration-first, self-hostable, and built to change as rules, regulation, and volume do.
Build
Senior engineers ship in tight increments, each tested and reviewed as it goes, so the system is reviewable at every step instead of only at the end.
Harden
Security, compliance, and load-testing run inside the build, so controls, audit trails, and peak-volume behavior are proven before launch.
Run
The handover includes clean, documented systems, with the option to keep the same engineers operating them once they are live.
Why Oxagile
Built for messy inputs
Clean sample PDFs prove very little. We test against creased scans, photos taken at an angle, multi-currency statements and layouts a vendor changed last quarter, and we set the confidence thresholds from that evidence rather than from a demo day. Analysts stop re-keying only when the awkward cases hold.
Straight through with a human path
Automation runs where the numbers reconcile and the score holds. Everything else stops for a reviewer, and anything outside the trained document types is rejected with a reason. Nobody has to trust a black box to post entries into the ledger.
Your documents stay with you
Documents are processed in your environment, so scans and payslips stay inside your boundary. Retention windows and redaction rules are configuration your privacy team reviews. Models, rules and the review interface come as source code with the licence terms written down.
Questions
What accuracy can we expect?
We will not quote a number before seeing your documents. The measurement comes from a labelled sample of your own files, split by document type, reported per field rather than as one headline figure. You then choose thresholds by consequence, so a payment amount can require review far more often than a memo line. Those numbers get rechecked as formats change.
What happens when a document type we have never seen arrives?
It stops. Classification returns low confidence, the document parks in an unknown queue, and a reviewer either handles it manually or tags it as a new type. Once enough examples exist, extraction gets extended and evaluated like any other type. The pipeline does not guess a mapping to make the throughput figure look better.
Can this work with our existing systems?
Yes. Output goes where you already post, through the API or file interface your ERP, core banking or accounting system expects. We keep the pipeline behind an interface, so a change on either side stays contained. Idempotency keys and reconciliation reports mean a retried batch cannot post twice. We also replay a historical batch during acceptance so both sides agree on the numbers.
Do we still need the review team?
Yes, with different work. Volume through review drops as thresholds settle and templates stabilise, and the remaining queue is the part where judgement matters. Reviewers also become the labelling source that keeps the models current. We size the queue with you during the pilot and agree who owns the exception path.
Part of AI in Finance
AI put to work in finance operations, grounded, governed, and running in production.
Get AI into production,
not just a demo.
Have a workflow begging for AI, whether documents, copilots or forecasting? Tell us the use case and we'll take it to production, governed for risk.