AI solutions
Document Processing Automation
Most of the invoices, delivery notes and customs documents that reach your company still enter your systems because someone reads them and retypes them, line by line. We build the layer that does that work: it captures each document where it arrives, extracts the fields, checks them against your purchase orders and vendor records, and escalates to a person when something does not add up.
Where document traffic actually jams
A supplier invoice lands in the finance inbox as a PDF and has to be matched to a purchase order, except the supplier used their own item codes and the mapping to yours lives in one person's head. The delivery note at goods-in says twenty cartons, the order says eighteen, and whoever notices raises it in a chat message rather than in the ERP.
What these tasks share is that each one only takes five minutes. The cost is not the five minutes; it is the interruption repeated dozens of times a day, the absence of any record of who looked at what, and the fact that mistakes only surface after a payment term has lapsed.
- The same data is typed twice: once in the supplier's system, once in yours
- The only way to learn a document's status is to ask someone
- Discrepancies get resolved in email threads, not in the system of record
- Month-end close turns into a batch run of everything that piled up
- The process stalls when one key person is on leave
Which documents we handle
The right place to start is the document that is touched by hand most often and whose structure is most stable. Supplier invoices and delivery notes meet that description in almost every company; importers add the customs file, manufacturers add certificates of analysis.
Adding a new document type does not mean starting over. The same extraction pipeline receives a new field set, new validation rules and a new target record definition. What decides whether a type is a good candidate is not how messy it looks, but whether its fields can be defined consistently and checked against a source of truth.
- Procure-to-pay: supplier invoices, proformas, delivery notes, order confirmations, expense receipts
- Import and export: customs declarations, bills of lading and CMRs, packing lists, certificates of origin
- Production and quality: certificates of analysis, subcontractor progress claims, material issue notes
- Commercial and legal: contracts, price lists, statement-of-account letters, insurance policies
- Field and warehouse: goods receipt notes, proof of delivery, count and damage reports
How the system works
The flow starts wherever the document arrives: a shared mailbox, a network folder, a scanner, a supplier portal, a photo taken in the yard. The incoming file is first cleaned up for orientation, resolution and page boundaries, because a single PDF containing twelve invoices scanned back to back is routine rather than an edge case.
Classic OCR alone does not get you there; the hard part is reading line items correctly. The raw output is normalised: date formats, currency, VAT, carton-to-unit conversion, and the mapping from the supplier's item code to yours. Validation comes last, and it is what makes the automation trustworthy: do the line totals reconcile, does the invoice match the purchase order and goods receipt, does the unit price deviate from the contracted price.
- Capture: email, folder, scanner, portal, mobile photo
- Pre-processing and document boundary detection
- Document type classification
- Header field and line item extraction
- Normalisation and code mapping
- Arithmetic, three-way match and contract price validation
- Posting to the target system, image attachment and audit trail
How it connects to your existing systems
The word integration usually assumes a modern REST API. Reality is less generous. An ERP that has been running for a decade may have no API at all, or an API that cannot create the document type you need. We evaluate, in order: an API, controlled writes at the database layer, the ERP's own bulk import interface, file-based exchange, and UI automation as a last resort.
Whichever route is chosen, three rules hold: data is staged before it is written so a problem on the ERP side leaves the document queued rather than lost; every write is safely repeatable; and everything is traceable.
- API, database layer, native import interface, file exchange or UI automation
- Staging and queueing so nothing is lost when the target system is unavailable
- Duplicate protection and safe retries
- End-to-end audit trail
- An installation that does not compromise your ERP's support or upgrade path
Marketplace, shipping, accounting and e-invoice connections are covered by our integration service.
Misreads, exceptions and human control
No document capture system reads every document correctly, and any setup that promises otherwise is unsafe. The useful question is not whether there will be errors but what happens when there are. Each field carries a confidence value; fields below the threshold do not reject the whole document, they drop into a verification queue where a reviewer sees the extracted value beside the image.
Alongside that, we define rules for cases that must never pass automatically regardless of confidence. Corrections are captured, and recurring corrections become a supplier-specific rule, so the system stops asking the same question over time.
- Field-level confidence and selective human review
- A verification screen showing image and extracted data side by side
- Hard stops: new supplier, changed IBAN, price deviation, invoice without a PO
- Corrections that feed back into mappings and rules
- An owned exception queue for unprocessable documents, with no silent failures
What it looks like by sector
The pipeline is the same across sectors; the place it pays off is not. In manufacturing the bottleneck usually sits in procurement; in logistics the weight is on import files and freight cost; in retail the volume comes from a long tail of suppliers.
What all three share is that document volume scales with business volume, while the team cannot scale the same way.
- Manufacturing: raw material invoice matching, subcontractor claims, CoA-to-lot linkage
- Logistics and distribution: import file consolidation, freight audit, proof-of-delivery archive
- Retail and e-commerce: long-tail supplier invoice traffic, receiving discrepancies, returns and damage reports
How we start and how we measure
We start with discovery: where each document comes from, who touches it, which system it ends up in, and where it currently gets stuck. We do this against real examples rather than assumptions, asking for a representative sample including the awkward ones. A pilot is then scoped to a single document type and run in parallel against live documents.
To measure the outcome we record baseline figures before the pilot begins, because after means nothing without before. We do not open with a promised saving; we fix the measurement method up front and keep measuring the same way in production.
- Discovery: document inventory, sources, target records, current blockers
- A representative sample of real documents
- A pilot scoped to one document type, run in parallel
- Baselines captured before the pilot starts
- Indicators tracked: human touches per document, arrival-to-posting time, straight-through rate, exception backlog, corrective entries
Frequently asked questions
Where do our documents and data go?
That is a decision we make with you at the start, and the architecture follows it. The system can run entirely on your own servers or in your own cloud tenant. If an external model service is used, we work under enterprise terms where your data is not used for training, retention is limited, and personal data is masked where required. Access control, logging and GDPR compliance are in scope either way.
Our ERP is old and has no API. Is this still possible?
Yes, and it is the situation we meet most often. An API is not the only route: controlled writes at the database layer, the ERP's own bulk import interface, file-based exchange or UI automation are all viable. Our only condition is that the route must not compromise your system's vendor support or upgrade path. We identify which options are genuinely open during discovery.
What happens if the system reads a document incorrectly?
Misreads are designed for, not hoped against. Every field carries a confidence value, and fields below the threshold go to a verification queue; the document is not posted without a person seeing it. On top of that we define rules for cases that must always be reviewed regardless of confidence: a new supplier, a changed IBAN, a deviation from contract price, an invoice with no purchase order.
How many document types do you support?
Quoting a number would be misleading, because what matters is not the count but whether a type's fields can be defined consistently, whether its volume justifies the work, and whether there is a source of truth to validate against. Implementations usually start with one or two types, prove themselves, and then expand.
Our documents contain handwriting and our scans are not clean. Is that a blocker?
Usually not, but expectations should be set correctly from the start. Printed fields, stamps and table structures are largely readable; handwriting varies with legibility and is typically configured to route into the verification queue. During the pilot we see concretely which document quality flows through untouched and which needs human eyes.
We already use e-invoicing, so our documents are electronic. Why would we need this?
E-invoicing solves the format of the invoice, not the process around it. Even with structured data you still have to match it to the purchase order and goods receipt, map the supplier's item descriptions to your own stock codes, catch price deviations and move it through an approval flow. And in most companies e-invoices are only a slice of inbound document traffic.
Let's look at your own documents
This is understood on your real paperwork, not in a slide deck. Share a sample set of recent invoices, delivery notes or customs documents, and we will work out together what flows through cleanly, where human review will be needed, and which route into your existing systems is actually available.