Document extraction demos are won on clean, digital PDFs. Production is won on the documents nobody wants to show in a demo: a phone photo of a supplier invoice, a stamped delivery challan, a receipt with a handwritten total, a purchase order scanned at an angle. Buy for the queue, not the demo.
This guide is for a finance or operations team that has decided to stop rekeying documents and wants to buy a workflow rather than a model. The troubleshooting guide to document extraction covers what goes wrong after launch; this page covers how to specify, compare and contract the work before launch.
Start with the documents, not the technology
Build a document register before you speak to vendors. For each type, record the source, typical condition, language mix, volume band and what the extracted data feeds. A supplier invoice and a courier receipt may look similar to a buyer and completely different to an extraction system.
- Document families: supplier invoices, purchase orders, delivery challans, expense receipts, bank statements, application forms.
- Condition: digital, scanned, photographed, stamped, handwritten, faded or mixed.
- Language: English, Bangla or a mixture, including names and addresses that do not transliterate cleanly.
- Volume: average per day, peak per day and month-end spikes.
- Destination: accounting package, ERP, spreadsheet, approval workflow or archive.
Count the exceptions before you count the pages. A process that handles most clean invoices and a minority of problematic ones is not solved in proportion to the clean share; the problematic set determines the support load. Ask the vendor to price exception review separately, because it consumes human time even when extraction is charged per page. Also record where each document arrives from — email inbox, courier, WhatsApp photo or shared drive — because an extraction engine only sees what reaches it.
Requirements to put in the brief
A model that reads text is not a workflow. The brief should describe the full path from arrival to posting, including what happens when confidence is low.
- Target schema: the exact fields, line-item structure and totals you need, with which fields are mandatory.
- Validation rules: arithmetic checks, tax logic, duplicate detection and matching against purchase orders or supplier masters.
- Confidence handling: a review queue with the source page, the extracted value and the reason for low confidence.
- Human approval: who can accept, correct or reject a document, and what evidence is stored.
- Audit trail: original file, extraction version, reviewer, changes and posting reference.
- Retention and deletion: how long documents and extracted data are kept, and how deletion is proven.
- Integration: API or file-based handover, error handling and reprocessing rules.
- Reporting: straight-through rate, exception reasons, correction time and cost per document.
Vendor categories and what they are good at
Cloud document AI services offer strong OCR and prebuilt invoice models with consumption pricing. They are fast to start and easy to over-consume. Document automation suites add review queues, approvals and connectors, which reduces integration work but increases platform lock-in. LLM-based extraction handles varied layouts and language mixtures well, but needs guardrails, evaluation and cost control. Local integrators bring implementation capacity and support in your time zone, with varying depth. Custom builds fit unusual approval or accounting rules but carry the highest ownership burden.
Most SMEs end up with a combination: a cloud extraction service, a review interface and an integration into the accounting system. The buying decision is therefore less about the model and more about who owns the seams.
Evaluation scorecard
| Criterion | What to test | Weight |
|---|---|---|
| Accuracy on your documents | Field-level correctness on a labelled sample, including poor scans | High |
| Exception handling | Queue design, correction speed, reason codes, reviewer roles | High |
| Audit and controls | Original file retained, version history, approval evidence | High |
| Integration fit | Test posting into your accounting or ERP environment | Medium |
| Security and data use | Residency, subprocessors, training on your data, deletion | High |
| Commercial model | Per page, per document, per field, minimums and reprocessing | Medium |
| Support and change | Who tunes rules, response times, release process | Medium |
Run a pilot that can fail
Select a sample that represents the real mix, not the cleanest folder. Label the expected values yourself, agree the accuracy measure in advance, and include duplicates, multi-page documents and at least one format the vendor has not seen. Run the pilot through the review queue and into a test ledger, not just through an extraction API. At the end, compare correction minutes per document and the number of documents that could not be processed at all.
Pricing questions that reveal the real cost
Ask what counts as a page, whether a failed or reprocessed page is charged, how minimum monthly commitments work, what happens when volume spikes at month-end, and whether human review is included or billed separately. Ask for the cost of adding a second document type and a second language. The cheapest per-page rate can lose to a higher rate that includes review and integration support.
Contract terms and red flags
The contract should state data ownership, permitted use for model training, subprocessor notification, breach notification, deletion timelines, service levels for availability and support, and an exit route for extracted data and original files. Treat vague answers on training data, unlimited price increases, no accuracy definition and no export path as red flags. A vendor who will not run a pilot on your documents is asking you to buy blind.
Where this fits in our work
Document workflows sit inside Business Automation, alongside approval and reporting work. The pricing page explains how a pilot is separated from production support. For post-launch symptoms and fixes, read the AI document extraction troubleshooting guide.
Sources
- NIST, AI Risk Management Framework, accessed 17 November 2026.
- ISO/IEC 42001, AI management systems, accessed 17 November 2026.
- OWASP Top 10 for Large Language Model Applications, accessed 17 November 2026.
- EU Artificial Intelligence Act, Implementation Timeline, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.