By Tarang Puranik, EVP, Global Service Offering Head (SAP) & Head Europe, Infosys
AI agents are moving into production faster than governance frameworks can keep up. In 2025, fewer than four in ten organisations were assessing the security of their AI tools, as per WEF. Now, enterprise AI is expanding at a pace traditional security often cannot match – with the average Fortune 500 company set to deploy over 150,000 AI agents by 2028, up from fewer than 15 in 2025, as per a recent Gartner report.
Because each agent acts on its own while holding key access to sensitive systems, security models built for a small handful of digital assistants often become insufficient at this scale.
When governance gaps surface after a production incident rather than before, agents may already have accessed data outside their authorized scope or taken actions without traceable records.
Three foundational controls close that gap. Establishing a verifiable identity for every agent comes first, because it determines what access is possible before any action is taken. Access governance, calibrated to each agent’s actual scope and autonomy level, comes second. And finally, real-time auditability, built into agent operations from initial deployment, makes every action reconstructible when it matters.
Establish a verifiable identity for every AI agent
Every AI agent deployed in an enterprise should have a registered identity and defined credentials. The lifecycle governing that identity, from creation through decommissioning, needs to be managed with the same rigor applied to human accounts and service accounts. Static scripts typically follow fixed paths, but dynamic agents adapt on the fly, often exposing the limits of rigid access keys.
Traditional identity frameworks, already extended to cover machine actors and automation credentials, must extend to AI agents with the same discipline.
There is a huge growth in the number of non-human identities in most enterprise environments. AI agents are the newest and fastest-growing category in that count. An agent operating without proper identity controls can access systems beyond its designated scope or take actions that are difficult to attribute during post-event audits.
Apply access controls proportional to autonomy
Not all AI agents carry equal risk. One that reads and summarizes documents operates differently from one that is authorized to initiate transactions or communicate externally on behalf of the organisation. Applying identical controls to both is either too restrictive for simpler agents or dangerously permissive for more autonomous ones. A tiered access structure lets automation move quickly, while keeping permissions strictly tied to verified safety boundaries.
The principle of least privilege, which has always been foundational to enterprise security, applies with the same logic to non-human identities.
As per Gartner, by 2027, four in ten enterprises are projected to scale back or decommission autonomous AI agents due to governance gaps revealed by unforeseen operational risks. Agents deployed without proportional controls will have caused access-related security incidents that appropriate governance would have prevented. Classifying agents by autonomy level and calibrating access accordingly is where proportional governance begins.
Build in auditability from the start
AI agents operate at machine speed. When something goes wrong, two questions dominate every review: what the agent did, and when. For agentic systems without audit trails, those questions often have no clear answer. Obscure event records can quickly turn minor operational glitches into critical operational visibility gaps. At scale, retrofitting auditability after deployment is not a realistic option.
As per a Gartner research, by 2030, half of all AI agent deployment failures are projected to stem from insufficient runtime governance enforcement. That points to the same root cause: agents operating outside defined parameters with no mechanism to detect or halt the deviation before consequences accumulate. Building auditability means embedding continuous monitoring and defining automated guardrails that trigger when agents approach the boundaries of their authorized scope. Tracking the end result is no longer enough; audit trails need visibility into the reasoning behind each step.
In conclusion
The proliferation of AI agents is not slowing. But without foundational controls, each agent represents an unregistered identity and an ungoverned access point. Identity, access, and auditability are the conditions under which agentic AI scales with confidence rather than risk. Enterprises that build them in now will not be starting from scratch when the time comes.