Teach an AI agent your documents. It reads every file after.
Show it two or three samples, say in plain language what to pull out, and approve the answers. Textikraft learns that kind of document and turns every new one into clean, structured data, with the evidence and a confidence figure for each value.
See how it works on a real screen
Pick a ready-made agent or teach your own, see what it reads and how well, and let a person confirm only what is doubtful. Made-up example documents.
From samples to an API in an afternoon
Show it examples
Upload two or three real documents of one kind. Scanned PDFs and spreadsheets with colour coding are fine.
Say what you need
Chat in plain language. Textikraft proposes the fields; you correct anything, and approve the answers you trust.
Measure, then publish
It is tested against your approved answers and scored. Only versions that reach your accuracy bar go live.
Call the API
Send a document, get JSON back with accuracy, credits used and timing. Retry-safe, signed webhooks, batches.
Clean data, with the proof beside it
Each result is structured data in the format you define, with an accuracy figure. Each value carries the exact text it was read from, so a person can check it in seconds.
- Results as JSON, Excel, CSV or a webhook
- Arithmetic and rule checks flag what does not add up
- Every version of an agent is kept and can be rolled back
{
"status": "completed",
"result": {
"data": {
"invoice_number": "INV-2026-1002",
"invoice_date": "2026-09-22",
"total": 376833.00
}
},
"accuracy": { "estimated_percent": 97.3 },
"timing": { "total_ms": 5500 }
}Illustrative output. Your agent returns the fields you define.
Built to be trusted with real documents
It does not guess
Values are checked against the document text and your rules. Anything it cannot find is returned as empty with a reason, never invented.
Cheaper and faster over time
After training, the AI writes extraction rules that run with no AI at all. The model is only asked about what the rules cannot read.
An accuracy figure on every result
Measured against your approved answers while training, estimated per document in production, and exact once a person reviews it.
Versions you can roll back
Retrained and it got worse? See every version's accuracy, go back to the old one, or retrain from its exact samples.
Many layouts, one endpoint
Different suppliers send different layouts. One agent holds them all and routes each document to the right one.
Prepaid credits
1 credit is …. Tax invoices, budgets and a cost on every document, so there are no surprise bills.
Invoices, receipts, purchase orders and the paperwork around them
Invoices from fifty suppliers. Photographed receipts. Purchase orders in a dozen formats. Delivery notes, bank statements and contracts with conditions buried in clauses. If a person can read it, you can teach an agent to.
- PDF (text or scanned), images, Excel, Word, CSV, text, Markdown
- Spreadsheet cell colours are understood, not lost
- Arithmetic checks catch totals that do not add up
Your team stays in control
Roles decide who can build, review, see billing or manage people, and each person's menus can be narrowed further. Every important action is in the audit log.
- Sign-in with two-step verification
- Documents deleted after 30 days, or sooner on request
- Test keys that are free and keep no data
One product, any document-heavy team
Wherever people read documents from many senders and type the facts into another system.
Try it on your own documents
Tell us about your documents and we will set up a pilot on your own files, so you can see the results before you decide.