Parashift

More than thirty prebuilt connectors reach ERP, DMS, CRM and workflow systems, including SAP, Salesforce, SharePoint, M-Files, JobRouter, d.velop, Lobster, BSI and Octo by TCG Process, and they span RPA, iPaaS, ERP, DMS, CRM, email inboxes and the customer's own applications with no middleware needed. Invoices are matched two and three way against purchase orders and order confirmations and released straight for posting in the customer's ERP, and cases export automatically to the target system at the end of the pipeline.

SourceParashift, parashift.ai/en/platformread 2026-09-07

Documents move through a multistep pipeline from intake to a completed case.

Intake runs by email, scan, upload or API, multipage batches are split into individual documents automatically, each is classified against a catalog of 300 or more types, fields are extracted with a confidence score each, and case management follows with validation status, an audit trail and automated export to the target system.

Agents orchestrate parts of the pipeline within guardrails and escalate when a threshold is not met, and the two and three way matching module reconciles an invoice against purchase orders and order confirmations before releasing it for posting.

SourceParashift, parashift.ai/en/platformread 2026-09-07

A class catalog of 300 or more document types, each with its own extraction schema, is maintained for the customer, models are retrained directly from the customer's corrections in live operation through OneTouchLearning, and matching runs against the customer's own purchase orders and order confirmations, all of which persist and shape later runs. No index, graph or embeddings layer holds the customer's policies or reference material for agents to query, and the extraction schemas are configuration for reading documents, not a knowledge layer.

SourceParashift, parashift.ai/en/platformread 2026-09-07

Review is built into the platform in two parts. With human in the loop, cases below a set confidence threshold or showing anomalies go automatically to an embedded validation interface where a person makes the final call, with no frontend development needed, and the interface can be embedded in third party applications under an SLA.

With agent in the loop, agents run parts of processing only within predefined guardrails, and confidence thresholds, validation rules and approval workflows decide when an agent may continue on its own and when a case is escalated. Every extraction carries a confidence score against a threshold the customer sets, and the same guardrails apply whichever model is running.

SourceParashift, parashift.ai/en/platformread 2026-09-07

Parashift holds ISO 27001, SOC 2, BSI C5 and PCI DSS and publishes a Compliance Trustcenter at trust.parashift.io. Customer control comes from multitenancy, with centrally defined configurations that individual tenants inherit, dedicated tenants, end to end encryption in transfer and storage, and confidence thresholds and approval workflows the customer sets per agent. No single sign on, SAML, SCIM or named role model is shown, so access control rests on tenancy and configuration, not an identity integration.

SourceParashift, parashift.ai/en/platformread 2026-09-07

A complete audit trail records model decisions, confidence scores and human interventions across the whole pipeline, with traceability of every automated decision built in. Parashift presents end to end governance and an audit trail per confidence score and decision as what sets it apart from a classic IDP platform. The pipeline ends in a complete auditable case with validation status and audit trail, and every field carries its own confidence score whichever model captured it, so the record shows why a decision was made and by which model.

SourceParashift, parashift.ai/en/platformread 2026-09-07

Parashift keeps learning, not memory. OneTouchLearning retrains models from corrections during live operation, and DocumentSwarmLearning pools model learning across customer environments, so the result lives in model weights. No memory with a stated scope and lifetime, no session or case state carried between runs, and no way to delete such memory for one customer is described.

SourceParashift, parashift.ai/en/platform and parashift.ai/en/about-parashiftread 2026-09-07

Customers choose where their data is processed among three compliance zones, Switzerland in Swiss data centers, the EU in EU data centers, and Germany in a BSI C5 operating environment in German data centers. Document data never leaves the selected zone, and the platform can run so data is processed only in the chosen country. Separately, the Dokumentli model and its Node can run fully in the customer's own data center, including air gapped deployment, which covers the model, not the whole platform.

SourceParashift, parashift.ai/en/platformread 2026-09-07

A catalog of 300 or more document types ships out of the box, each with its own extraction schema, and customers adopt them, customize existing templates or build from scratch. The platform also ships named modules, the Matching Hub for purchase to pay two and three way matching, the embedded validation interface, automation agents and the Flow integration component. Each class and module does its own job when selected, and the rest stay whole when one is left out.

SourceParashift, parashift.ai/en/platformread 2026-09-07

Documents arrive by email, scan, upload and API, forming an inbound queue across four channels. An inbox is monitored, boundaries within it are detected automatically, and each document is routed to the matching workflow on arrival. Downstream, confidence thresholds and validation rules fire escalation events that route a case for review. No broader event trigger system is described.

SourceParashift, parashift.ai/en/platformread 2026-09-07

Customers choose the model, and Parashift presents this as its main differentiator. Under a bring your own model principle, the platform connects OpenAI API compatible models, including GPT, Claude, Gemini and Mistral, or Parashift's own Dokumentli Vision Language Model, and any LLM a customer wants can be connected.

The model can be swapped by use case, price or model generation without rebuilding the guardrail and automation logic, and a specialized and a general purpose model can run together on the platform, with guardrails, audit trail and human in the loop applying whichever model runs. Parashift contrasts this open approach with the hard wired engine of a classic IDP platform. Its own proprietary model approaches are technologies it uses, not choices the customer makes.

SourceParashift, parashift.ai/en/platformread 2026-09-07

Parashift publishes an open API through which every platform function is available, API first and platform independent, with documentation at docs.parashift.io. The validation interface can be embedded as a standalone component in a third party application, and software vendors can embed Document AI in their own products under OEM licensing.

SourceParashift, parashift.ai/en/platform and docs.parashift.ioread 2026-09-07

Every extraction passes a live quality gate. Each field carries a confidence score against a threshold the customer sets, results above it continue automatically and everything below goes for review, whichever model captured the data, and corrections retrain the model in live operation through OneTouchLearning. The gate checks each output, not a change under test. Customers can swap models, but there is no way to compare one model against another on their own documents, and no evaluation set, scored test cases, release path or holdout is described.

SourceParashift, parashift.ai/en/platformread 2026-09-07

The platform reaches outside systems through an open API and more than thirty prebuilt connectors and takes documents in by email, scan, upload and API, all programmatic routes. Separating, classifying and extracting pages means reading files at build time and run time, not driving a screen, and no browser, desktop or remote computer control is described.

SourceParashift, parashift.ai/en/platformread 2026-09-07