Botminds IDP Intelligent document processing that reads anything — and proves every value.

Botminds IDP is the intelligent document processing technology inside the Botminds platform. It reads any document or record — scans, handwriting, complex forms, tables, spreadsheets, emails, photos, call transcripts and video, in the language they were written in — extracts what matters, checks it against rules and other documents, and hands a cited result to a person or a system. Every value links to the page and region it came from.

Intelligent document processing (IDP) is software that turns documents into checked, structured data and decisions. It sorts what arrives, extracts the fields and passages that matter, validates them against rules and other documents, and passes the result on, with a person reviewing what the system is unsure of. Our buyer’s guide explains IDP in depth.

OCR (optical character recognition) turns printed text in an image into characters. ICR (intelligent character recognition) does the same for handwriting. IDP uses both as early steps, then understands the layout, classifies the document, extracts the meaning, validates it and routes the result. OCR gives you text; IDP gives you checked data you can act on.

Agentic document processing gives each job on a file to an AI agent with its own instructions, tools and limits. In Botminds IDP an intake agent sorts and splits the file, extraction agents read each document by meaning, a standardization agent puts values in your formats, a quality agent checks the reads, a cross-document agent compares the file, and redaction, answering and drafting agents do the rest. They stop at a person where judgment is needed, every step is recorded and every value is cited, so the work can be audited.

Each agent has a written boundary: the collections it may read, the tools on its allow-list, the rules it must follow and the cases that need a person. An agent whose tool settings are missing or broken gets no tools at all. Text inside a document is treated as evidence, never as instructions, so a page cannot widen what an agent may do. Tools that change something can wait for a named person’s approval, and every change to an agent is tested on a golden set before it publishes.

A collection is the home for one kind of work: its document types and fields, its review stages and the automations on each stage, the agents that run when a document arrives, the policy for when AI may skip the reviewers, and who may see and act on it. A mortgage program, for example, is a loan-file collection with a child collection for the documents in each file, a structured collection for income and asset lines, a knowledge collection of guidelines, a work queue of conditions and a golden set for testing.

Yes. Handwriting is read from the page image alongside printed text, with a confidence on every word. Ticked boxes are detected from the ink on the page, not guessed from text. Anything below the confidence threshold goes to a person, who sees the original image next to the reading.

Scanned pages are put upright and read by OCR, then the layout model rebuilds each table as a grid — header cells, rows, merged cells and a confidence for every cell. Extraction agents join tables that continue across pages and read each cell into a typed value with its units. Totals are then re-added, so a misread number shows up as a sum that does not foot.

Native and scanned PDFs up to 1,500 pages each; images (TIFF with every frame, JPEG, PNG, GIF, BMP, WebP); faxes; Word, PowerPoint and legacy Office files; Excel and CSV, kept whole; email (EML, MSG) with its attachments; HTML, text, JSON and XML; ZIP archives, unpacked safely; call and meeting transcripts; and video, sampled frame by frame. Documents are read in the language they were written in and cited to the original words.

We measure accuracy field by field on a golden set of your own documents and report the numbers together: the share of fields that pass automatically, the precision of those automatic passes, and how well confidence is calibrated. On held-out real financial statements — mostly French, a quarter of them scanned — fields above the confidence threshold passed automatically at 98% measured precision, and the rest went to a reviewer. Our target for what leaves the system after review is 99% field accuracy, and we run the test on your documents before you sign.

Botminds IDP scales out with the queue: a deployment sized for the wave reads a million pages of mortgage files in 30 minutes — about 2,000 loan files of 500 pages, or roughly 556 pages a second. That is a sizing figure for a scaled-out deployment; the measured runs are single files. Pages, documents and sections fan out across workers that are added as the queue grows, model calls are pooled across several deployments, and intake is governed so nothing is dropped. Measured on single files, a 1,000-page loan file was read end to end in under 14 minutes, and a 607-page packet was split into 61 documents in about 10 minutes with no errors. In production, deployments have read 100,000 scanned invoice pages a day.

Every extracted value must point to the page and region it came from, and the confidence score checks that the value actually appears on that page. Values are re-read and compared, checked by a second reviewer agent, and tested against rules and other documents. Low-confidence values go to a person, and every change to a model or prompt re-runs the evaluation set before it ships.

Yes. Agents find personal, financial, health and privileged data and propose each redaction with the rule behind it; a reviewer approves before anything is removed. In PDFs the text is taken out of the page, not covered by a box, and in images the pixels are overwritten, so nothing can be copied back. Recordings and video are masked on playback. Every removal is logged with its reason.

In the region you choose — in the Botminds cloud, a dedicated environment, your own cloud subscription on Azure, AWS, Google Cloud or Oracle Cloud, or on your premises. Original documents, page images, extracted values, indexes and run history stay in that region, with backup copies in its paired region in the same geography; model calls go only to the model deployments you approve. Each core service runs as at least two copies, workers scale with the queue, and the database has point-in-time restore.

Access is set by role and narrows from the platform to the workspace, the project, the collection, the review stage and the single action — see, correct, approve, export or configure. Sign-in goes through your identity provider with SSO and SAML; a workspace can be locked to it and require two-factor sign-in. Chat and search answer only from collections the asker may see, agents are scoped to the collections they are attached to, and one view shows any person’s effective access.

Botminds holds SOC 2 Type II and ISO 27001 and supports GDPR. Data is encrypted with TLS in transit and AES-256 at rest, isolated per customer, and never used to train shared or third-party models; reviewers’ corrections improve only your own deployment. Every action by a person or an agent is in the audit trail, redaction removes data rather than covering it, and four-eyes review puts a second named person on the decisions that matter.

Yes. Forensic checks look for mixed fonts, white-out overlays, editing-tool metadata and running balances that do not add up; they need no AI model, take well under a second, and hold a suspect document at intake. Exact duplicates are caught by fingerprint, and agents look for contradictions across documents. Each signal is shown for an investigator to judge.