Your next employee might be an AI Agent. Does your access policy know that?

Every access policy your enterprise runs today makes one assumption: whoever is logging in is human. That assumption no longer holds true. Fixed roles, predictable behaviour, and access reviews once a quarter: the whole architecture of identity and access management was built for people. And people, for all their unpredictability, are still easier to govern than many of the identities operating in your system now. An AI agent decides, acts, and interfaces with other agents across systems, without waiting for anyone to sign off in real time.

Indian enterprises are moving faster on agentic AI than most global peers expect. According to Nasscom’s AI Adoption Index 2.0, the Indian AI market is projected to grow at 25-35% CAGR through 2027. Clearly, this is no longer a pilot conversation, and the pressure to move fast is now colliding with identity infrastructure that was never built for AI.

According to a survey of 250 India-based IT decision-makers actively using AI in their environments, 46% reported that their AI initiatives have already been delayed by identity friction, the highest figure recorded globally. 79% said they had accepted standing access under operational pressure and 68% said they simply lacked alternatives to standing access altogether. For a market being pushed to move quickly on AI, this is a quiet but serious drag on velocity.

That tension is familiar to anyone who has sat through a security review during a product launch. The instinct under deadline pressure is to grant broad access now and clean it up later. With human employees, that habit is risky but survivable, since people change roles and eventually leave, forcing some natural churn in permissions. Agentic AI does not go through that cycle. An agent granted standing access will keep using it, exactly as configured, until someone actively revokes it.

Part of the difficulty is that AI agents do not act as ordinary machine identities. A service account or API key does exactly what it is programmed to do, in a fixed predictable way. An AI agent decides, in part, how to act, based on the instructions it is given and what it encounters in the moment, including other agents. Given that unpredictability, the instinct is to try and keep a running record of every agent in play, but that’s a harder ask than it sounds. Agents get spun up faster than any list can track, and even a fully current one wouldn’t stop access that’s already slipped through via manipulation, a technical failure or plain human error. For CISOs, the sharper focus is on what data an agent is reaching for, how sensitive that data is, and setting permission levels accordingly, rather than assuming an agent is safe just because it’s been accounted for somewhere

There is also an accountability gap sitting underneath this too. As agents increasingly talk to other agents, often across different business units, it is rarely clear who owns the risk when something goes wrong at that boundary. Is it the team that built the agent, the one whose data it touched, or security by default. Left unanswered, this is quietly becoming the new version of an old problem, unmanaged and undiscovered privileged access, just under a different name.

Retrofitting governance onto a system that is already live rarely works. It means untangling which permissions an agent actually needs from what it was given by default, usually under time pressure. The fix must start earlier, with identity architecture designed for non-human actors from the point an agent is first provisioned.

There is also a regulatory dimension specific to India. Under the DPDP Act and CERT-In’s six-hour breach reporting window, an unmanaged agent identity with standing access to sensitive data is a compliance exposure with a clock attached. This is pushing CISOs, chief risk officers, and governance functions to align far more closely than before, and boards evaluating agentic AI investment can no longer treat identity governance as a downstream IT decision.

The shift required is architectural. Least privilege applied to every agent, not just every person. Access granted just-in-time rather than standing by default. Credentials that expire automatically instead of sitting unused. Policy decisions evaluated in real time against context and risk, rather than trust granted once at login and assumed to hold.

None of this is a call to slow down. India’s enterprises are clearly not waiting, and the productivity case for agentic AI is real. But the organisations that capture that value safely will be the ones that build identity governance into agent deployment from day one, and that close the ownership gap before an incident forces the question. India’s focus should be on whether their organisations can confidently say exactly who or what has access to their most sensitive systems at any given moment, be it human or otherwise.

(The author is vice president-APJ at Delinea. Views are personal.)