Adobe Acrobat AI Assistant for enterprise: Our commitment to data governance and security

Today, we announced the general availability of Adobe Acrobat AI Assistant for our enterprise customers. Acrobat AI Assistant allows you to interact with your documents in Acrobat for quick answers and one-click summaries to create impactful content and improve productivity. It brings together our deep knowledge of PDF documents with our proprietary technology for document processing.

The Adobe difference: Responsible AI innovation

Organizations around the world trust Adobe with their most important digital documents. As organizations harness the power of AI to boost enterprise productivity and create impactful content, we continue to prioritize advanced, enterprise-grade security and privacy controls.

Just as Adobe Acrobat is tailored to meet the specific privacy and security requirements of enterprises, we have devoted significant consideration to developing Acrobat AI Assistant in line with those existing processes. All Adobe generative AI features go through Adobe’s AI Ethics governance process and are developed and deployed in alignment with Adobe’s AI Ethics principles of accountability, responsibility, and transparency.

Building trust in generative AI for PDFs

• Adobe’s agnostic approach to Large Language Models (LLMs) enables us to choose the best-in-class technology for the task at hand. For Adobe Acrobat’s generative AI features, we currently leverage Microsoft’s Azure OpenAI Service, which is contractually prohibited from manually reviewing or training its LLM on Adobe customer data.
• We leverage Microsoft Azure OpenAI’s content filtering service to moderate hate, sexual, violent, and self-harm content. To further ensure we provide ethically responsible responses, the Acrobat Generative AI Service also leverages a defined block list as an additional mechanism to filter hate and profanity. Responses are only provided to the user if they pass both checks.
• Adobe strictly limits who can access reported content, bugs, or vulnerabilities and user-provided feedback to a small number of trained Adobe employees. These employees use algorithmic and manual processes to review the session data and help address reported issues within the Acrobat generative AI features.

"Adobe has long focused on establishing a strong foundation of cybersecurity, built on a culture of collaboration, enabled by talented professionals, strong partnerships, leading-edge capabilities, and deep engineering prowess,” said Mike Mellor, vice president, Cyber Operations at Adobe. “We are continuously testing our generative AI technologies, learning from our global teams, the broader security researcher community, and partners to strengthen resiliency and improve our capabilities.”

Respect for customer data

• No Adobe customer data is used to train or fine-tune any LLMs. Also, for a document to generate summaries and insights, a user must click on the AI Assistant or generative summary buttons to activate the features. We also have guidance available on how users and enterprises can turn off the generative AI features in Acrobat, if so desired.
• The Acrobat AI Assistant only looks at the information presented in the document and no other external sources such as the web, email or stored content in other locations are referenced. LLMs have generally pre-trained knowledge that they can use to answer questions like ‘What is the capital of France?”. A lack of attributions or added qualifiers indicate the use of training data.
• Adobe’s custom attribution engine and proprietary AI generates clear citations so users can easily verify the source of answers from within the user-provided documents.
• Chat history (inputs and generated responses) remains in the user’s control for deletion or query continuation. Adobe Acrobat AI Assistant persists chat history to cater to the users’ desire to return to their prior session.
• Acrobat Generative AI Service extract content from the PDF to process the prompt. Except for the chat history retention mentioned above and any content reported by the user, the uploaded document, prompts and generated responses from a session are all automatically deleted from the Adobe cloud services after 12 hours.
• The AI Assistant responses are for individual consumption only. Even responses on shared documents on web are only accessible to that user, not shared across users.

Built for business

• AI Assistant in Reader and Acrobat provides enterprise-grade data security and information governance capabilities, including granular admin-level controls such as selecting users/user groups for access and enabling just-in-time provisioning of the product.
• To support data security, we have built robust testing and monitoring methodologies in pre- and post-processing and engineering processes.
• All user content, prompts, and responses are encrypted in transit. At rest, any data stored by the Acrobat Generative AI Service is encrypted using SHA-256.

We welcome feedback

The future of technology is exciting, but there can be implications if these innovations are not built responsibly. In addition to the initiatives mentioned, Adobe provides customers with multiple channels to provide feedback — from in-app and advisory boards to community groups and social listening programs — we encourage active conversations to help the company identify and address issues quickly. And as always, we value our customers and encourage feedback through your preferred channel to better serve you and your business requirements.

Learn more

Adobe's vision for Acrobat AI Assistant includes expanding insights across different document types, AI-powered authoring and editing, and AI-supported document collaboration. We are also exploring integrations with Adobe Express.

For a more efficient and intelligent document experience, learn more about Acrobat AI Assistant. For the Acrobat AI Assistant security factsheet, click here.