As enterprises move from GenAI experimentation to deploying autonomous AI agents across core business workflows, the risks of weak foundations are becoming harder to ignore. In this conversation with Tech Achieve Media, Dr. Rania Khalaf, Chief AI Officer, WSO2, explains why integration, identity, API governance and observability need to evolve alongside AI adoption. She highlights the risks of agents operating with human or shared credentials, the growing role of agent-visible APIs and MCP, and why enterprises must treat non-human identities as first-class actors. Her message is clear: AI at scale cannot be built by simply adding intelligence to existing architectures; the underlying enterprise fabric must also be ready to support, govern and secure autonomous action.
Also read: Harry Ault Named CEO at WSO2
TAM: Many enterprises rush into GenAI applications while ignoring underlying integration and identity debt. At what point does foundational neglect turn an AI initiative from a productivity driver into an operational liability?
Dr. Rania Khalaf: Most GenAI pilots survive on borrowed trust, where a human checks the outputs. In these cases, weak integration and identity underneath don’t show up as risk yet. I’ve seen agents provisioned through a shared service account with broad access just to get a demo working, and it holds up fine right up until that same agent is processing refunds or updating a customer record in production.
The shift happens the moment the agent starts acting instead of a person reviewing what it suggests. Integration and identity debt that sat quietly under a human-in-the-loop process turns into unattributed actions in live systems. This can easily become catastrophic. For example, we have seen agents delete production databases even when told not to. That could have been avoided if those agents had their own identity with limited write access instead of the full credentials of the engineers that created or deployed them. Greenlighting agents on top of that foundation isn’t moving fast, it’s borrowing against a bill that comes due the first time something goes wrong in production, with real consequences.
TAM: How does WSO2 shift enterprise mindsets from viewing APIs as mere plumbing to recognizing them as the primary neural pathways for autonomous AI agents?
Dr. Rania Khalaf: APIs have always carried the information. We just built them for people and deterministic systems: a developer reading docs, a dashboard someone glances at once a day. Agents don’t read docs or dashboards. They read structured, well-described capabilities and act on them directly, increasingly through MCP that lets an agent discover what a system can do without a person wiring it up by hand.
That changes what the API layer has to be. It’s no longer integration glue, it’s the surface an agent senses and acts through. Enterprises need to govern it the same way they’d govern any production system, with the same rigor around access and change control, just visible to the agent calling it not only the human maintaining it.
TAM: What is the biggest blind spot executive leadership teams exhibit when attempting to unify identity, access management, and AI workflows on a large scale?
Dr. Rania Khalaf: The biggest blind spot is assuming an identity stack built for humans can just be force-fit onto agents. It can’t. A human logs in a handful of times a day; a workflow agent orchestrating a single purchase across five systems needs five separate, scoped grants for that transaction window alone, not one master credential that happens to work everywhere. The same applies really for the full platform stack: AI implications come in at every layer from API platforms, to integration to identity. These systems have to be redesigned with agents, generative AI models, tools, and knowledge bases as first class abstractions and not as an afterthought that is bolted on.
People will often tell you they have a unified identity strategy and mean the human one, while agents are handed a static API key because it’s faster than doing it properly. Closing that gap means building an identity system that actually models non-human actors, their own credentials, their own lifecycle, their own audit trail. Most teams haven’t closed that gap yet.
TAM: As autonomous AI agents move from pilot projects into core enterprise workflows, they frequently run into security vulnerabilities by borrowing human credentials. How is WSO2 architecting “agent-first” identity framework to prevent catastrophic privilege escalation and forensic blind spots?
Dr. Rania Khalaf: We give every agent its own Agent ID, not a repurposed service account or a human’s OAuth token. For agents operating on their own, that’s a straight machine-to-machine flow with no redirect. For agents acting on a person’s behalf, the delegation is explicit and asynchronous rather than the agent quietly inheriting whatever the human is entitled to. Access itself is scoped tightly rather than one broad grant, and it’s time-bound with automatic rotation, so an agent has exactly the permissions it needs for that transaction window and nothing standing afterward.
That combination is what closes both problems, privilege escalation and forensic blind spots, at once. Catastrophic privilege escalation gets harder because there’s no broad, long-lived credential to escalate from. The forensic blind spot closes because every action logs back to a distinct agent identity, kept separate from human activity, instead of disappearing into “the service account did it.
TAM: For CIOs and engineering leaders bogged down by decades of technical debt who feel pressured to jump straight into generative AI, what is your definitive “stop-doing” list to regain the architectural agility required for true AI readiness?
Dr. Rania Khalaf: Stop treating every agent as a one off and design for agents running through your core systems instead of sitting next to them. Practically, that means embedding dedicated agent identity, access, and observability that works with the stack you have, instead of stringing together agent governance across a patchwork of components that were never built for it.
I still see teams greenlighting agents with no governance plan, provisioning agents through a human’s credentials or shared service account, or spinning up MCP servers without the security controls agents actually need, all in the name of moving fast. Speed and control should not be a tradeoff here. Put the controls in first in a way that blends with your current workflows and systems, and they become an accelerator instead of a blocker.
TAM: If you look beyond the current AI hype, what three digital capabilities would you say an enterprise absolutely cannot compromise on to make AI work at scale?
Dr. Rania Khalaf: My top three: First, a governed, agent-visible API and tool layer, increasingly sitting behind an AI gateway, so an agent has something real and well-described to act on. Second, an identity and access model that treats agents as distinct actors, with their own credentials and their own scoped and time-bound access. And finally, observability and evaluation built to handle and effectively record non-deterministic behavior. I’m seeing a huge push for agent control planes that cover all three, so your controls scale with consistency across different models and frameworks (which continue to evolve) instead of adding technical debt.
Predicting where AI goes next is a guessing game, but no matter which model or framework an enterprise bets on, the foundational controls it needs remain essential. These three areas make up the fabric an agentic enterprise needs to run on without compromising quality, control or security. They are not independent, but are intertwined.
