蘑菇短视频

As artificial intelligence evolves from simple chatbots into systems capable of conducting research, analyzing data and completing complex tasks, organizations face a growing question: How much of that work should depend on technology and infrastructure controlled by outside companies?

A faculty member in 蘑菇短视频鈥檚 College of Business is conducting ongoing research to examine the potential for locally operated artificial intelligence to give researchers and organizations greater control over their data, AI models, workflows and costs.

Dan Bradbury, an assistant professor of marketing, is conducting the research as artificial intelligence increasingly shifts from systems that answer individual prompts to autonomous agents capable of carrying out longer, multistep processes.

鈥淢ost people think about AI as a single prompt and response, but an AI agent can carry out a much more involved process,鈥� Bradbury said. 鈥淚t may search the web, read documents, write code, analyze data, retrieve files, revise its own work and repeat steps when a task fails. That creates questions not only about costs (given that typical pricing models charge per token) but also about where data are being processed, which models organizations depend on and how much control users retain over the entire workflow.鈥�

Bradbury worked on the study over the summer through financial support from the College of Business Summer Research Grants Program, which awards funding to up to four faculty members each academic year.

Faculty submit proposals during the spring semester, and a committee of fellow faculty members, typically made up of past grant recipients, selects the winning proposals for the summer. Grant recipients are awarded $5,000 each to support their research and are expected to submit a proposal for conference presentation and then produce a written product of the research upon completion.

Bradbury used his grant funding to purchase hardware capable of running large language models and developed a local AI research environment to run and test these models from the open-source AI ecosystem. Initial work has demonstrated that Apple silicon technology (M-series chips) can run certain compressed language models containing as many as 72 billion parameters locally, he said.

鈥淭hat capability matters because most of the leading AI tools people use today operate through systems hosted and controlled by outside technology companies,鈥� Bradbury said. 鈥淩unning models locally creates the possibility of keeping the model, the data it processes and the records of the research workflow under the researcher鈥檚 direct control.鈥�

The research is part of a broader question surrounding what is sometimes described as AI or digital sovereignty 鈥� the degree to which individuals and organizations retain control over the technology, data and infrastructure on which increasingly important AI-enabled activities depend.

For researchers and universities, one of the clearest implications involves privacy.

鈥淔aculty members increasingly use AI to assist with writing, feedback, data analysis and course preparation,鈥� Bradbury said. 鈥淲hen student submissions, grade comments, advising materials or sensitive research data are sent to an outside AI provider, universities have to consider privacy obligations and institutional policies, including requirements such as the Family Educational Rights and Privacy Act.鈥�

A local artificial intelligence workflow can reduce or eliminate the need to transmit those materials to an external model provider because the model and associated files can remain on hardware controlled directly by the researcher, he said.

Local AI also changes the economics of intensive AI use.

Leading commercial AI services commonly charge according to the amount of information a model receives and produces. Those charges can be relatively small for an individual question but become much more significant when an autonomous agent repeatedly processes long documents, tool outputs, instructions and conversation histories while completing a complex task.

鈥淪ome frontier models currently charge $10 per million input tokens and $50 per million output tokens,鈥� Bradbury said. 鈥淎 demanding agent workflow using 5 million input tokens and 1 million output tokens, for example, can cost $100 or more through a commercial application programming interface before considering search, computer-use or other tool charges.鈥�

For a single task, those costs may be manageable. But Bradbury said the economics can change when AI is used repeatedly for research, data analysis or other organizational workflows.

The issue is not simply whether local AI is cheaper, however, Bradbury said.

鈥淭here are tradeoffs,鈥� he said. 鈥淐ommercial frontier models can be extremely capable and convenient. The question is when researchers or organizations may benefit from operating capable models locally because they value greater control over their data, their research process, their costs or the models themselves.鈥�

The project also gives Bradbury an environment in which to study how open AI systems perform on research tasks and where their limitations remain.

鈥淭he larger goal is to understand when local AI can provide a practical alternative or complement to externally hosted systems,鈥� he said. 鈥淚 want to develop transparent workflows that help researchers and organizations understand the tradeoffs and retain greater control over their data, models, costs and research processes.鈥�

The work may ultimately produce both a scholarly article and an open toolkit for applying local AI to areas such as secondary-data research, literature synthesis and other academic workflows.

Bradbury will continue working on the project this fall, refining the results and developing preliminary findings for presentation at next year鈥檚 Appalachian Research in Business Symposium, hosted at Appalachian State University. He will also serve on next year鈥檚 selection committee for the 2027 round of College of Business Summer Research Grants.

Other 蘑菇短视频 College of Business faculty members who received $5,000 in summer research funding for 2026 are Gary Curnutt, assistant professor of finance; Aaron Heinrich, assistant professor of operations management; and Christine Kirkland, assistant professor of management.