Architecture and security guidelines for Document Automation with Generative AI capability

Here's a list of privacy, security, and typical use cases for your consideration for using Document Automation with Generative AI capability.

Functionality

Note: Generative AI models can produce errors and/or misrepresent the information they generate. It is advisable to verify the accuracy, reliability, and completeness of the content generated by the AI model.
What is different about the manner in which Document Automation processes unstructured documents and shipment documents compared to invoices?
Automation Anywhere incorporates (individually modeled) large language models (LLM) in the Document Automation product to assist with processing unstructured documents and shipment documents.
What document types can be processed using Generative AI?
Any unstructured and semi-structured documents, including pre-trained document types such as: invoices, bill of lading, waybill, arrival notice, or packing list.
Are there any limitations on the field types supported by Document Automation?
No, we support form and table fields with GenAI capability.
What are the supported languages?
Officially English is supported, but other languages would also work.
What is the pricing structure for Document Automation with Generative AI capabilities?
Automation Anywhere charges per page for Document Automation, and the OpenAI cost is included in the price.
Could a customer call their own Azure endpoint for Microsoft Azure OpenAI LLM?
Yes, we support bring your own license (BYOL) for Microsoft Azure OpenAI and a customer-defined LLM endpoint use case is supported.
Is the Generative AI with Document Automation feature available On-Premises in a customer’s private cloud?
Yes, integrated generative AI is now available for use for the On-Premises and Cloud versions of Document Automation.
What OCR engine can be used for unstructured document types?
Currently we support Google Vision OCR and ABBYY OCR.
What OCR engine can be used for shipment documents?
We recommend using ABBYY OCR and Google Vision OCR for shipment documents.

Security

Which LLM does Document Automation send the queries to?
Document Automation currently uses Microsoft Azure OpenAI services for LLM use. We currently use GPT3.5-Turbo and ADA models.
In which countries are the Document Automation documents submitted for processing?
Automation Anywhere has selected the EU and US regions for processing the documents.
Does Automation Anywhere permit vendors to train their LLMs on data from documents submitted to Document Automation?
No, Automation Anywhere does not use or allow any third-party vendors to use documents or data submitted by customers or users for training purposes.
Does Automation Anywhere train its own instance of any LLMs on data from documents submitted to Document Automation?
No, Automation Anywhere does not use or allow any third-party vendors to use documents or data submitted by customers or users for training purposes.
Do any additional vendors have access to prompts or documents which customers submit to Document Automation?
Yes, please see Exhibit C of our DPA (link at the end of this paragraph) for a list of vendors who have access to data submitted by customers to Document Automation. The applicable vendors will say “Document Automation” in the second column labeled Applicable Services. Link to DPA.
Does Automation Anywhere store documents submitted to Document Automation or data extracted from the documents?
Beside Automation Anywhere Cloud hosting providers, the Automation Anywhere vendors do not store any documents, as documents exist in a transitory state only. However, to enable customers validate the documents, Automation Anywhere stores the document images and the OCR output of the document text in the Cloud instance.
What kind of data is shared with the LLM services?
Documents, content from documents, and prompt or search queries.
Will Automation Anywhere enter into a Business Associate Agreement (“BAA”) for the processing of Protected Health Information (as defined by C.F.R. § 160.103) (“PHI”) with respect to PHI, which a customer submits to Document Automation?
Yes. Typically, our BAA is automatically incorporated into our Cloud Automation Agreement. Please reach out to your Customer Success Manager (CSM) or Sales Executive (SE) to check whether you have a BAA in place, and if not, they will get you a copy of our BAA.
Will Automation Anywhere enter into a Data Processing Agreement (“DPA”) for the processing of Personal Data which a customer submits to Document Automation?
Yes. Typically, our DPA is automatically incorporated into our Cloud Automation Agreement. Please reach out to your Customer Success Manager (CSM) or Sales Executive (SE) to check whether you have a DPA in place, and if not, they will get you a signed copy of our DPA. If you would like to see the terms in our DPA , it can be found here: Automation Anywhere DPA Agreement
Important data usage details from Google Vision OCR Service (Source: Google).
Google Vision OCR Service Data Usage

Architecture diagram

Document Automation with GenAI solution architecture

The extraction process consists of several steps:
  1. Authentication:
    • Extraction process starts by running the Document Extraction package on a bot runner device
    • (1) Configuration of the extraction process is defined in a learning instance. If the learning instance uses Generative AI-driven extraction, the package connects to Control Room to fetch information about the Proxy Gateway host location and authentication token.
  2. OCR:
    • The next step is to retrieve a text layer from the document. The system uses OCR engines for the task
    • (2) If Google Vision OCR is selected in the learning instance, the package sends a document to Proxy Gateway
    • (3) Proxy Gateway forwards this request to the Google Vision APIendpoint for OCR, and (4) sends the results back to the package
  3. Data Extraction:
    • (5), (6), (7) To process large documents, the package creates embeddings for different chunks of the document using the ADA model
    • (8), (9) Next, the package identifies the most relevant chunk of the document for the provided search query(s) using embedding and sends that chunk along with a prompt to the GPT model through the Proxy Gateway
    • (10) Finally, the package receives responses from GPT and converts them into document extraction results

For Automation Anywhere Cloud Generative AI data security information, see: Data security for generative AI - FAQ