Changelog

Today we're announcing second opinion validation: a new way to cross-check AI predictions before they move downstream.

In high-volume document workflows, there is always a trade-off between automation and accuracy. Manually validating every prediction improves confidence, but it slows everything down. Fully automated processing is faster, but only works if you can trust the data.

Second opinion validation adds an extra layer of certainty. By comparing predictions against a second model, Cradl AI can flag the cases that need review and let confident predictions pass through automatically.

That means less manual validation, without compromising on data quality.

We’ve significantly improved performance when reviewing documents with many lines or large tables.

Review now feels faster and smoother, especially for high-volume documents like invoices, order forms, price lists, statements, and other table-heavy files. You should notice quicker loading, more responsive navigation, and less friction when working through complex documents.

This makes it easier to validate detailed documents without slowing down your review workflow.

This month we are shipping a native Cradl AI node for n8n.

Setting up production-grade document extraction in n8n used to mean building and maintaining a lot of custom logic, API calls, validation, error handling, retry mechanisms, and more. Not anymore.

The new Cradl AI node brings the full power of the Cradl AI platform into your n8n workflows. Drop it in, connect your account, and you get accurate data extraction out of the box, along with everything that makes it production-ready: human-in-the-loop review for edge cases, model training to improve accuracy over time, and built-in validation to catch errors before they reach your systems.

The result is a document extraction pipeline that gets smarter the more you use it.

To get started, search for Cradl AI in the n8n node library. Full documentation is available in our docs.

This month's update focuses on reliability, ensuring your document workflows keep running smoothly, even when the unexpected happens.

LLMs are in high demand, and latency or availability issues are a reality. To keep your agents running without interruption, we have introduced automatic fallback support for LLM models. If your primary model encounters an issue, Cradl AI will now seamlessly retry and, where needed, route to a fallback model — all without any manual intervention. This means fewer failed runs, less downtime, and greater confidence that your workflows will complete reliably, even during peak demand periods.

This update introduces a new state-of-the-art model and a suite of validation tools designed to make your document workflows more accurate and automated than ever before.

We have officially integrated the Qwen 3.vl LLM into our platform. In our extensive testing, Qwen 3.vl has consistently outperformed Claude Sonnet 4.5 in data extraction accuracy and visual understanding. Because of this superior performance, Qwen 3.vl is now the default LLM for all new agents, ensuring you get the highest quality results from day one.

Data consistency is critical for downstream automation. Our new "Match and Map" validator allows you to ensure extracted values align perfectly with your existing data. You can now match extracted text against a set of predetermined values, automatically mapping them to the correct format for your ERP, CRM, or database.

To further reduce manual review, we’ve added a comprehensive list of new specialized validators. These tools automatically verify the format of extracted data, flagging errors before they reach your systems. The new validators include:

• Financial & Tax: IBAN, BIC/SWIFT, VAT ID, Tax ID.

• Identity & Contact: Passport Number, Phone Number, Email.

• Logistics & Transport: Freight Container ID, License Plate.

• Web: URL

Alongside these features, we have implemented several general stability improvements and addressed minor bugs to ensure a smoother, faster experience across the entire Cradl AI dashboard.

All documents are now automatically rotated to the correct orientation before processing. This means you no longer need to manually rotate your files, saving you time and effort.

We have revamped our role system to provide more granular control over your workspace. You can now invite members to your workspace with the following roles:

• Admin: Full access to the workspace, including billing and workspace settings.

• Developer: Can edit workflows, integrations, and AI models.

• Reviewer: Can only review runs that have been flagged for review.

This new role system allows you to collaborate more effectively and securely with your team.

Reviewing documents is now faster and more accurate with our new Smart OCR feature. When reviewing a file, you can simply click on a word in the document to instantly copy it to the corresponding field. This eliminates the need for manual typing and reduces the risk of errors, making your review process more efficient.

We’ve added a brand-new Insights section to help you better understand how your agent performs. Get detailed metrics on accuracy, throughput, and intervention rates to identify where your workflows can improve.

You can now enable email notifications to alert reviewers whenever new documents require human validation, ensuring faster turnaround and smoother collaboration. Prefer other channels? You can also integrate notifications with Slack, Teams, or any other platform using webhooks.

You can now trigger workflows directly in Microsoft Power Automate when key events occur in Cradl AI. This makes it easier than ever to connect your document automation to other business tools and streamline end-to-end processes.

We’ve made several improvements and bug fixes in the Review section for a smoother and more reliable validation process. Expect faster loading, better navigation, and improved annotation handling.

Updated: Excel Integration

Our Excel integration has been enhanced for improved import/export reliability and better compatibility with complex spreadsheet formats.