AI’s Impact on Data Silos and Knowledge Hubs

Extraction, Classification, and Augmentation

ML techniques are also helpful for extracting metadata from—and classifying—enterprise content that can be made available across disparate business units. Although this application doesn’t necessarily involve mitigating silos, it still makes information available across the organization in a way that’s not possible were it locked in silos. According to Pava, this use case is a subtler form of employing AI techniques to share content across the enterprise for something as pervasive as onboarding a new employee.

KM systems, for example, can ingest documents specifying what information is needed to perform the tasks for adding a new employee. Specifically, those systems parse documents containing the steps to be completed for new hires, extract the relevant information from them via ML, and make it accessible across business units. For example, the newcomer might “need to get security set up for the facilities, or need to get their badge from HR, or their laptop setup from IT, or get different roles for the manager to schedule onboarding orientation meetings and other tasks,” Pava remarked.

In addition to parsing relevant content to extract the necessary information from it, contemporary systems employ ML to supplement the information gleaned from the documents. Examples of augmentation tasks might include translating the extracted information into a different language. “It’s not just one-to-one extraction,” Pava pointed out. “I can actually augment that with additional information, like, ‘If it’s in this office, this is the additional information we need.’”

Language-Model Infused Agents

Dynamic agents that enable language models to perform actions similar to those of humans are also viable for minimizing the effects of silos while increasing the worth of knowledge hubs. In some use cases, agents perform actions that span silos—such as assembling enterprise knowledge, data, code, and other information to modernize legacy applications or build new ones. In other use cases, they access tools from different clouds, departments, and even geographic regions to coordinate instances of question answering, reporting, or business intelligence.

The capability for agents to access various tools from different locations helps counteract the effects of data silos. According to Bixby, progressive organizations are able to minimize cost while boosting performance for facilitating the tool selection process that is necessary to benefit from such multi-agent architectures. “For our larger clients, we’re expecting agents will grow to have hundreds, potentially many hundreds, of tools available,” Bixby commented. “Passing all those tools to the agent every time and saying, ‘Hey, which is the right tool to use?’ doesn’t feel like an effective way to do it.”

At present, many organizations utilize the approach Bixby referenced for the tool selection process—which consumes tokens going to the models powering agents and increases costs. However, by creating vector embeddings of use cases for which tools are the most helpful or widely used (which can involve the history of agent interactions), organizations can minimize the costs, the tokens, and even the time required for agents to select tools.

“We’ve learned the vector database and a vector query is the perfect way— in literally a couple of milliseconds—to find the right tool based on the query from the user,” Bixby observed. What’s important about this process is that the user can either be an agent or a human. According to Bixby, this method results in “massive gains in terms of performance, as opposed to passing all the tools, and all the tool definitions, as part of a prompt and burning a bunch of tokens. A lot of agent frameworks to this day work that way.”