I School Professor Launches New Company Centered Around “Agentic AI”
UC Berkeley School of Information assistant professor Niloufar Salehi has partnered with University of Southern California professor Afshin Nikzad and entrepreneur Steven Mih to launch Across AI, a startup pioneering agentic AI for complex enterprise workflows. The startup recently came out of stealth, announcing it had raised $5.7 million in a seed round of funding co-led by venture capital firms Cota Capital and Village Global.
Across AI aims to harness the power of agentic memory to support enterprise sales teams and chief revenue officers in their work. Agentic AI, refers to the ability of artificial agents to plan, take action, and store and recall information about past interactions, experiences, and knowledge, allowing it to make informed decisions and adapt to changing circumstances. Across AI’s system works with internal and external data sources such as customer relationship systems, communication and collaboration tools, and calendars to identify and quantify new sales opportunities, spot risks, and suggest questions to ask customers. The team plans to release a web app and chatbot that will provide in-time assistance when needed.
“The system actively tracks, timestamps, and monitors information updates, recognizing when data becomes outdated or conflicts with new information,” said Mih in a TechCrunch article on the company. “Unlike traditional AI systems that treat all data equally, our agentic memory system prioritizes information based on contextual importance. Where possible, the apps keep the inferences up to date themselves. Where ambiguity exists, determinations are escalated to a relevant person, such as a sales manager or product manager.”
For years, Salehi’s research into reliable AI systems has centered the people who use or are impacted by the technology. Years before launching the company, Salehi conducted research with then-members of Google’s Ethical AI team on human-centered machine translation and how to design more safe and reliable systems. These experiences became a crucial part of Across AI’s approach to AI, promising users privacy and data security.
“You need to have control over what the agents do and you need to have transparency into what they’ve done,” explained Salehi. “A cornerstone of the product is that whenever there’s an output, you can always click on it and dig deeper to see the steps of reasoning that were taken, as well as the source documents…All of those things enable the verification, transparency, and explainability of the models, which is a really important part of making them useful and reliable in the real world, building trust with users.
“My vision for Across AI is for us to be the place that cracks the code on how to make AI agents useful and reliable for people. I want us to be at the forefront of developing a human-centered AI ethos and developing the best practices that…center the people who will be working with these agents and augmenting them.”
— Niloufar Salehi
Additionally, Salehi was inspired by other human-computer interaction research she conducted while at the I School. In one specific paper, she worked with then-Master of Information Systems and Management student Eva Yiwei Wu (‘20) to explore algorithmic systems and the roles that people envision them having. For example, when students studied the YouTube algorithm, they began assigning roles or personas to the algorithm, such as “gatekeeper” or “personal talent agent,” even though they knew that the YouTube algorithm is ultimately a combination of different machine learning algorithms. By assigning these personas, users could decide how much they could rely on the algorithm, what to do when it was wrong, and how to work with it or against it. This research also applies to how people may view and interact with autonomous agents, which is the core of Across AI’s technology.
Currently, Salehi is on leave from the I School and serves as the Chief Product Officer at Across AI, where she leads a team of ex-Google employees, UX researchers, and product designers to build the technology.
“We have to invent a whole new set of ways for people to interact with AI and with these models,” she elaborated. “Now you’re at a point where [these models] will be doing some of the work for you and they’ll be stochastic…and people are not used to that. Part of what makes this work really exciting is that you have to develop the UX and the AI at the same time.”
“My vision for Across AI is for us to be the place that cracks the code on how to make AI agents useful and reliable for people,” Salehi added. “I want us to be at the forefront of developing a human-centered AI ethos and developing the best practices that…center the people who will be working with these agents and augmenting them. Trust, reliability, and transparency are what I would like our mark on the industry to be and what we could be seen as a leader for.”