Aaron Farris, Founder of Fieldborne
Aaron Farris, Founder of Fieldborne

Businesses are no longer using AI just to draft text or answer questions. It's increasingly being employed to write code, handle transactions, and inform decisions. Deloitte's State of AI in the Enterprise 2026 report found that 74% of leaders expect their organizations to use AI agents at least moderately by 2027. Yet only 21% have a mature governance model for agentic AI. That gap may matter more than the technology. The issue might be less about what AI can do and more about who verifies it before it acts.

Written policies might not always mean governance. In EY's September 2026 survey of senior AI executives at large US companies, almost all reported formal AI governance policies. Yet 26% of those using agentic AI said their organization could not detect unauthorized AI agents operating internally. When an agent acts on a flawed output, the error could pass straight into live operations before anyone notices.

The risk lies partly in how AI fails. AI systems can produce answers that sound confident but are wrong, invent information, or depend on outdated sources. McKinsey's 2025 State of AI survey found that inaccuracy was the most common negative consequence organizations reported, cited by nearly a third of respondents. A single unchecked error could carry real financial costs, especially in software architecture, banking, accounting, or healthcare.

Repeated success can also breed complacency. A tool that has been right a hundred times may be trusted on the hundred-and-first, even when that output is flawed. Some businesses may hold back from AI altogether because they cannot predict how automated decisions will interact with established processes. Others might adopt it without fully understanding what governance means in practice.

Aaron Farris, founder of Fieldborne, is developing a bounded AI-governance pilot built around a control plane, a layer that sits between AI-generated work and its implementation. Farris explains that the thinking grew out of his independent scientific research, where long-term projects needed experiments tracked and tested against defined outcomes. "I built the control plane to keep my work on track," he says.

Farris notes that when he shared his work publicly, engineers saw potential value for the wider AI industry. Some parts of the software have passed engineering reviews, and others are still being checked. Reviewing the second control plane, linking the two, and building the full product are later steps. The pilot is not presented as production-ready or as a guarantee of compliance. "I'd rather show what has been reviewed than promise what hasn't," he says.

Farris argues that governance should apply to the AI tools making decisions, not only to company policies and people. In his approach, AI outputs pass through a security gate with defined boundaries intended to reduce the chance of false or poorly informed decisions taking effect. "It's not about regulating the business," he says. "It's about regulating the tool that's regulating the business."

Farris compares the approach to a manager reviewing an employee's work, with a person examining each AI output and signing off before it is implemented. The system is designed to record the AI's reasoning, the reviewer, and the approval, linking each decision to evidence. "You should know who approved it and why," he says. Such records could help an organization explain a decision if it is later questioned.

Farris believes the first benefit may be efficiency, with less time spent proofreading and repairing AI work after the fact. He notes that no system is perfect and that a flawed output may still get through, which is why the human layer matters. Fieldborne is inviting organizations to bring one consequential AI workflow for a bounded pilot discussion. "Start with one decision that matters, and govern it properly," he says.

As AI shapes more business decisions, Farris explains that unchecked outputs could affect finances, livelihoods, and reputations. AI can also produce answers that sound right or pleasing without being accurate. "It shouldn't be about pleasing the user," Farris says. "It should be about giving people what they need to act safely." That may mean treating verification as a foundation of AI adoption rather than a fix after failure.

The choice for businesses may not be between rejecting AI and trusting it blindly. Farris sees potential in human-governed, evidence-linked AI workflows that could support more accountable operations. He also believes the infrastructure behind them may grow best when tied to measurable development milestones. "Govern your AI before it makes a decision you can't explain," he says. The real measure of AI adoption could be whether people can show who decided what, and why.