Identification turns a structured page into typed values: named entities, key-value fields, table line items and contract clauses. Every value returns with its page number, bounding-box coordinates and confidence score, so nothing you put in a downstream system is unverifiable.
Accuracy is measured field-by-field against a human-adjudicated ground truth, not character-by-character, and always against a stated document mix. The full methodology and per-field breakdown are on the accuracy page.
A document is rarely a form. It is a form with a table in the middle, a stamp over the total, a handwritten correction in the margin and two pages of terms at the back.
People, organisations, locations, dates, monetary amounts, identifiers and reference numbers are recognised wherever they appear, including inside free-flowing narrative text.
Labelled fields are paired with their values using layout geometry rather than proximity guessing, so a value in the column to the left of its label on an Arabic form still binds correctly.
Invoice lines, statement rows, schedule entries and repair itemisations are extracted as typed rows from the reconstructed grid, not scraped from flattened text.
Clauses are found by what they do — indemnity, termination, governing law, liability cap, renewal, confidentiality — rather than by matching a heading you hoped would be there.
Define your own fields with names, types, validation rules and optional location hints. The extraction models are instructed against your schema at inference time, so there is no retraining loop.
Each field returns the page it came from, the bounding box it occupies and a confidence score of its own. Nothing is averaged into a single document-level number.
Field-level traceability is the difference between an extraction tool and a tool an auditor will accept. Every value DocxIntel returns keeps a pointer back to the exact region of the exact page it was read from, in the original document, months or years later.
This is what an internal auditor, a regulator or a court asks for: not the value, but where the value came from.
The built-in types available before you define anything of your own. Custom types extend this list without a model change.
Field coverage for a specific deployment is confirmed in writing during the Proof of Value against your actual document mix and your actual schema, not against a generic list.
The point of confidence scoring is not to eliminate review. It is to send your reviewers the two per cent that need judgement instead of the hundred per cent that do not.
Fields, types, validation rules and per-field accept and review thresholds are configured by your team. A policy number can demand near-certainty while a free-text remark is allowed to be approximate.
Because analysis has already resolved structure and reading order, extraction works on a grid and a block tree rather than a wall of text. Each field is returned with coordinates and a confidence score.
Checksums, format rules, date ordering and arithmetic checks run automatically. A total that does not equal the sum of its line items is flagged even when both values were read with high confidence.
Fields below threshold or failing validation are queued with the source region pre-highlighted. The reviewer sees the pixels and the proposed value side by side and confirms or corrects in one action.
Every correction is recorded against the field, the document and the reviewer, giving you an audit trail and your own labelled dataset. Nothing is sent anywhere to be trained on.
Per-page pricing is not just a cost question. It quietly decides which documents you can afford to look at, and how often you are willing to change your schema.
Where fields are extracted
Adding a custom field
Coordinates and per-field confidence
Re-extracting 2M documents on a new schema
Arabic entity formats
Where reviewer corrections live
True air-gapped operation
Comparison reflects publicly documented behaviour of metered extraction APIs as of 2026, including credit-based per-page pricing, preset extraction charged at higher rates and self-hosting gated to enterprise plans. Verify against current vendor documentation before making a decision.
The objections that decide an evaluation, answered without hedging.
No. Fields are defined in a schema — name, type, validation rule, optional hint about where it usually appears — and the extraction models are instructed against that schema at inference time. Adding a field is a configuration change your own team can make in an afternoon, not a labelling project and a retraining cycle.
