Key considerations for businesses authorizing AI agents
IBM research shows only 11% of companies feel fully prepared for AI agent deployment.
1Corporate AI readiness
An IBM survey of 2,000 C-level executives reveals only 11% feel fully prepared for AI agent deployment. Two-thirds of CIOs and CTOs report accountability for AI systems they don't fully control, while 70% say teams deploy technology faster than IT can track. These findings indicate a growing control gap as AI adoption expands across enterprises. The data suggests most organizations lack adequate governance frameworks for autonomous AI decision-making.
2Capability vs. authority
PromptHalo founder Madhuri Chandoor emphasizes the critical distinction between what AI agents can do and what they're authorized to do. Her security infrastructure company examines why actions are performed, not just what's being done. This contextual approach evaluates user intent, assigned permissions, and environmental factors before system execution. Chandoor's model provides organizations with deeper visibility into AI decision-making processes and potential risks.
3The context imperative
Chandoor illustrates context's importance with a database management example. An AI agent optimizing application performance might autonomously alter production table structures, potentially affecting live transactions or dependent processes. "Technical conclusions may seem reasonable in isolation," she notes, "but context, circumstances, and downstream impacts must be evaluated." This perspective highlights how narrow AI optimization can create broader operational risks without proper safeguards.
4The refund loophole scenario
Chandoor demonstrates contextual risks through a refund exploitation example. If an AI agent can issue $50 refunds without human review, users might request ten $50 refunds instead of one $500 transaction requiring approval. While each $50 refund appears legitimate individually, the sequence suggests threshold avoidance. This scenario shows how atomic permissions without session-level context can create systemic vulnerabilities in AI agent deployments.
5Financial services inspiration
Drawing from 20 years in financial services, Chandoor adapts fraud monitoring principles to AI governance. She recommends developing behavioral profiles for autonomous agents alongside traditional identity and access controls. Her proposed framework monitors resource access patterns, tool usage, activity trends, and role-inconsistent actions. This approach mirrors financial institutions' transaction monitoring systems that detect suspicious behavior across accounts and activities.
6Authorization lifecycle
Chandoor advocates considering authorization during both AI system design and operation. Organizations should document accessible resources, access conditions, and potential downstream effects of AI actions. She proposes observability gates to review activity when requests become repetitive, unusually broad, or inconsistent with assigned purposes. These checkpoints help contain impacts and determine necessary additional controls for underlying systems.
7Responsible AI adoption
Chandoor positions her approach as enabling responsible AI adoption. She supports using agentic automation for analysis and workflows but recommends additional verification for critical decisions. "Trust but verify" encapsulates her philosophy - businesses should embrace AI innovation while continuously monitoring behavior. Her framework aims to balance automation benefits with appropriate oversight and accountability mechanisms.
8Accountability architecture
Chandoor stresses establishing clear accountability ownership for AI application security across organizations. "Operationalizing guardrails requires defined responsibility structures," she asserts. This accountability architecture helps businesses pursue AI innovation while maintaining necessary security and oversight. Her model provides a blueprint for governing autonomous systems as they assume increasingly consequential roles in enterprise operations.