Boomi is gearing to deliver before the end of this year a set of logic gates that ensure that actions being performed by an artificial intelligence (AI) agent or human are validated against a set of known facts about the business before they are permitted to be executed.

Designed to be embedded with the Boomi integration platform-as-a-service (iPaaS) environment, those logic gates are a critical element of a set of governance capabilities that can be applied to agentic AI workflows, says Boomi CEO Steve Lucas.

Digital CxOs to one degree or another are struggling with the same fundamental issue. Most workflows are deterministic in the sense that they need to be executed the same way every time. Generative AI technologies, however, are probabilistic. Tasks are rarely executed the same way twice. Logic gates enable organizations to define a set of business rules that AI agents must follow to significantly reduce the likelihood of unexpected rogue activity, says Lucas.

In general, more business and IT leaders are starting to better appreciate the inherent risks AI agents represent. Designed to execute tasks by any and all means available, there is a significant chance that at some point an AI agent is going to execute a task in a way that was unanticipated. The end result is a series of actions that, instead of benefiting the organization, winds up inflicting serious harm that could, for example, lead to significant fines being levied in a highly regulated industry.

Additionally, more organizations are also starting to better appreciate the implications of sharing data with large language models (LLMs) that are accessed via a cloud service, notes Lucas. While many of the providers of these models make commitments to not using customer data to train their models, they are still collecting metadata that provides them insights into how business processes operate and, crucially, how the underlying code that drives them was constructed, he adds.

Those insights could then be used at a future date to launch a competitive service using insights provided by employees that are sharing corporate data with models without much appreciation for its true value, says Lucas. As the providers of LLMs come under increased pressure to turn profitable, they are already launching offerings targeting multiple vertical industries that were largely generated using insights provided by end users accessing their models. As organizations start to appreciate this potential existential threat to their businesses, many of them are now looking to establish AI moats that are built around open-weight AI models that they can customize, and just as importantly, ensure that access is limited, notes Lucas. In some cases, organizations will go so far as to require end users to only use models provided by the company as a term of employment, he adds. “It will be a term of employment,” says Lucas. “A provider of proprietary AI models will be viewed as a competitor.”

The end result will be a massive wave of AI investments that will be directed toward everything from lower-cost open-source and private small language models to world models that add an ability to understand the consequences of an action, says Lucas.

Longer term, those AI investments will ultimately lead to the “great decomposition” of enterprise software as existing legacy enterprise resource planning (ERP), human resources and customer resource management (CRM) applications all become a set of headless backend services that AI agents are invoking either via an application programming interface or some type of integration framework, adds Lucas.

Ultimately, the only thing that differentiates one organization from a rival is the value of the data it collects to drive innovation. The more that data resides outside the organization, the greater the chances become that the business will not survive the AI era.