By Vishy Swaminathan, Senior Principal Scientist at Adobe Research
Adobe believes that the best use of AI is to give people more control and free them to spend more time on the work they love, whether that’s creativity, analysis, or collaboration. At last month’s Adobe Summit, Adobe presented its vision for AI assistants and agentic AI, giving the public an early taste of how this approach to AI has the potential to enhance everyone’s work, and new agentic AI previews from Adobe showcased the many ways this will enable customer success.
Here at Adobe Research, we are advancing AI assistants and agents. Our researchers have already helped create the Acrobat AI Assistant and AEP AI Assistant, and now, we are building the foundational pieces for Adobe’s emerging new agentic AI framework. These efforts aim to enhance customer interactions with products and data, making complex tasks more intuitive and efficient. Summit showcased AI innovations that our team has helped create, emphasizing the importance of reasoning engines and conversational interfaces in transforming customer experiences. And our research publications show the depth of our breakthroughs in this area, which will superpower the creativity and productivity of Adobe customers.
These are our team’s key initiatives in our work on agentic AI:
- Model Development and Training: Creating and fine-tuning foundational models, including generating synthetic data that closely reflects customer workflows without using sensitive customer data. Ensuring models provide accurate responses based on synthetic data.
- Agentic Experiences, Evaluation, and Safety: Shifting from application-based experiences to agentic experiences to solve customer problems. Enabling agents to make well-reasoned decisions for customers or for the system. Building the necessary scaffolding through structured evaluation for the agents to prevent hallucination, be trustworthy, and safe.
- Attribution: Ensuring accurate and trustworthy responses by attributing sources correctly.
- Planning and Reasoning: Developing reasoning and planning capabilities to solve complex problems using APIs, tools, and agents. Ensuring the system can backtrack and replan if models do not meet performance needs.
- Conversational Data Science: Bringing data, document, and content formats closer to interactions to enable direct actions based on conversations between customers and agents.
- Experimentation and Audience Agents: Simplifying AB tests and audience segmentation through conversational interfaces. Hiding statistical complexity to make AB tests user-friendly.
- Data Insights: Creating visualizations based on questions to rethink dashboards and interact with data dynamically. Data insights agent working on customer journey analytics data.
These themes collectively aim to enhance customer interaction with products and data, improve model accuracy, and simplify complex tasks through conversational interfaces.
The transition to agentic interfaces is poised to revolutionize how individuals interact with systems, simplifying human-computer interactions, boosting productivity, and enhancing creativity. Historically, systems have been designed based on presumed user intent. Now, we can dynamically map individual expressions of intent far more precisely with agents than traditional software interfaces.
As users increasingly expect these capabilities due to their experiences with AI, Adobe Research’s work will be pivotal in meeting and exceeding these expectations, ultimately transforming user experiences. While many challenges remain unsolved, the research community has an important role to play in these advancements. For instance, seamlessly integrating natural language with complex systems like databases and statistical models is an intriguing yet unresolved issue. Successfully addressing this and other research questions will empower people in profound ways.
Learn more about our work on Agentic AI by viewing some of our research publications in this area here.
