AI's Hidden Choices and Why Getting Them Right Avoids Lock-In
CIO, Tuesday, October 13th, 2026
Model choice matters less than the inference, compute and deployment decisions that keep AI agents affordable and governable.
Tushar Katarki argues that agentic AI turns enterprises from occasional token consumers into regular token producers, making inference cost a strategic concern.
Agents also act with real permissions at machine speed, which raises data residency, retention and security worries that stall deployments.
He says the inference layer determines whether organizations can see, control and cap token usage and enforce identity and policy consistently.
Models should be treated as a changeable portfolio within a stack rather than a single frontier-model bet, which helps avoid lock-in.