AI agents are the new insider threat: Is your SaaS stack ready?
Imagine hiring someone who skips the interview, never signs a non-disclosure agreement, and walks in on day one with a master key to every system in the building. No manager would approve that hire. Yet, that is exactly what is happening inside enterprise SaaS stacks right now, except the new hire is not human. AI agents (or non-human identities) have quietly moved from answering questions to logging into systems, triggering workflows, and completing tasks with barely a glance from a human. By the end of 2026, they will sit inside 40 percent of enterprise software, up from under 5 percent a year earlier, according to McKinsey. Adoption is outrunning oversight, and the gap is widening fast.
The employee nobody onboarded
Every enterprise has a well-worn playbook for managing people like background checks, role-based access, exit interviews that kill every login on the way out. None of it was built for a colleague that never sleeps, never resigns, and multiplies overnight. Multiply they have.
Non-human identities, AI agents included, now outnumber human employees eighty to one inside the average enterprise. Governance for agentic AI remains immature almost everywhere. Per Deloitte, only about one in five organisations describe their governance model as mature enough for the task. Clear boundaries for what an agent can decide alone, real-time monitoring of its behaviour, and a usable record of what it did and why, for example, remain the exception rather than the rule. Unlike a human who might hesitate before touching data they should not have, an agent does not pause to ask. Compliance frameworks, still built around headcount, have not caught up to actors that move at machine speed across every connected tool at once.
The risk is not hypothetical anymore
Around 70% of organisations suffered at least one identity-related breach in the past year, with permissions gaps in external applications a frequent root cause. The stakes climb fastest in industries where the margin for error was already thin, from banking to healthcare, where AI agents are being handed real authority over risk, compliance, and fraud decisions. Picture an agent with standing access to a CRM or HRIS, quietly updating records or moving data between connected applications, with no trail left behind once it is done.
Give the ghost a badge
The fix requires the same discipline enterprises already apply to people, extended to the agents working alongside them through:
– Defined access – giving every AI agent a unique identity with role-based permissions that limit it to only the applications and data it genuinely needs.
– Audit trails and Continuous monitoring – Every action an agent performs should be logged in tamper-proof audit trails, allowing security teams to trace who initiated a task, what data was accessed, what decisions were made, and which downstream systems were affected
– Zero-trust design, and human oversight – Zero trust principles should ensure that agents continuously authenticate themselves and revalidate permissions instead of inheriting blanket access across connected SaaS applications.
– Watchdog / Oversight mechanism – Ensure all of the above is reviewed periodically through a control mechanism.
Just as importantly, enterprises need a live inventory of every deployed agent, the workflows it supports, the systems it can access, and the level of autonomy it has, so employees, engineers, and security teams know exactly where autonomous decision-making exists and where human approval remains mandatory. High-risk actions, such as modifying sensitive records, approving transactions, or accessing regulated data, should always require human review before execution.
Momentum is already building in this direction. A growing number of organisations have begun mandating human review before an agent is allowed to act alone, and identity platform vendors have started expanding their scope to cover human, machine, and agent identity together, treating provenance as a baseline requirement. The risk of standing still is structural. The upside of moving early is just as concrete. PwC found that organisations investing most heavily in governance capture 74 percent of all AI-generated economic value, meaning the companies treating this as a compliance checkbox are already leaving money on the table. Governance, in other words, is not the brake on AI adoption. It is the engine.
Govern now, or get breached later
Nothing here is slowing down. Agent-embedded software is set to nearly triple its footprint within the year. What is missing is not smarter AI. It is the identity discipline enterprises already know how to build, simply extended to a workforce that does not clock in. There are two paths forward. Either we treat agents as first-class identities inherited permissions, monitoring, and oversight built in from the start, or we continue letting them run as invisible features bolted onto existing tools and inherit the breach that governance failure is quietly setting up. The organisations that badge their agents today will be the ones still trusted tomorrow.
