Overcome governance and trust issues to drive agentic AI

Fully autonomous agentic AI is still way off but AI agents are making inroads within enterprise software and workflows. Gartner predicts 40% of enterprise software will feature task-specific AI agents by the end of 2026 as the current trend for embedded AI assistants evolves.

Over the next two years, AI agents will only be deployed most likely across less critical functions, with businesses carefully monitoring their effectiveness. Quite rightly, trust in devolving decisions to AI agents is still shaky. A lack of transparency in AI agent thought processes, and issues with agents being able to consistently repeat outputs are real problems. Moving from formally structured scripted responses in applications such as chatbot assistants to more fully autonomous AI-based systems is a gamble many CIOs are not willing to take.

Internal governance is key

Agentic AI has the potential to upend workflows, organizational structures, supply chains, and entire industries. If autonomous agents can collect and track data, identify problems, and then deploy a solution with zero or minimal human interference, the implications are obvious.

The process of value creation in many industries will need to be reworked as inefficiencies are stripped out, costs reduced, and new revenue streams unlocked. As agentic systems share data and communicate across and between different enterprises via new protocols such as MCP and A2A, we can expect to see radical changes as many of the old ways of doing business disappear.

Before that can happen, however, systems of governance need to be established and agentic AI solutions need to incorporate better security technology. In fact, 62% of the 2,000 enterprises McKinsey surveyed said they were experimenting with AI agents with the redesign of workflows a core objective for many. However, moving from these experiments to actual deployments is less common, with two thirds of respondents saying they haven’t begun rolling out AI agents in any meaningful way.

Establishing centralized governance processes and maintaining humans in the loop to monitor accuracy are also identified in the McKinsey survey as vital to successful AI deployments, agentic or otherwise.

Then there’s Collibra’s survey of over 300 data managers and AI decision makers, which found that 86% believe agentic AI will generate positive ROI for their businesses in the medium term. And 60% believed providing governance and compliance training was a priority, even though less than half had established such processes. There’s clearly more work to be done on defining what an effective governance process looks like and how it should be deployed.

Vendor initiatives

On the supply side, vendors and solutions providers are increasingly emphasising the accuracy and security of their offerings. Salesforce puts the embedded security tools and guardrails in its Agentforce product range as central features. New entrant to this sector, MCP Manager, sees observability, governance, and security as key features of its plug-and-play agentic offering. Founder and CEO Michael Yaroshefsky believes customers need visibility into the servers and tools their agentic systems are calling on, as well as a clear picture of data flows and token cost estimates. This helps build confidence in the tools while inbuilt security measures, such as metadata locking to prevent tool poisoning, are important to instil trust.

The announcement of the formation of the Agentic AI Foundation (AAIF) earlier this month by OpenAI, Anthropic, and Block marks a major milestone in the work of vendors and developers to build trusted frameworks for agentic AI development. Anthropic has transferred ownership of MCP to the foundation alongside OpenAI’s Agents.md. Google, AWS, Microsoft, Cloudflare, and Bloomberg have also signed up to the AAIF, which will operate under the Linux foundation. Widely accepted and open standards will be essential if agentic AI is going to be deployed at scale.

Out of the starting blocks

The mainstream adoption of the internet was driven by the take-up of email and the web, and these were enabled by the open standards of SMTP, IMAP, POP3, and HTTP. Agentic AI won’t be a winner take all market or duopoly in the way that personal computing and smartphones have been, and the diverse nature of the different standards, protocols, data streams, and applications that will make up this transformative technology will see to that.

Ecommerce didn’t take off until the web was made more secure with the development of HTTPS, and cooperation from financial institutions enabled consumers and sellers to trust the internet with their data. With agentic AI, the starting pistol has been fired as if we were back in 1995 when the internet started gaining traction, and Amazon only sold books from its website.