The autonomy dial: How much should businesses delegate to AI?

Yet the ability to execute a task efficiently does not mean that the associated decisions are safe to delegate in every context. Where context and judgement matter, an error can affect customers or colleagues, interrupt operations, cause financial loss or create legal exposure. Incorrect invoices, gaps in diagnosis or treatment, mishandled insurance claims, and poor customer service all point to the same conclusion: autonomy must be calibrated to the consequences of each action.Zero-touch processes redesigned with AI can deliver business benefits, but accountability remains with the organisation and its named decision owners.

Agentic systems therefore need clear triggers for human review and escalation. During solution design, leaders should define who owns each decision, what agents may and may not do, who can approve or override an action, how exceptions will be escalated, and who can stop execution. These permissions and controls should be reassessed as evidence accumulates and operating conditions change.A framework for setting autonomyBusiness and technology leaders should treat autonomy as a dial, not an on-off switch.

Two questions determine its setting: What are the consequences if an agent makes an error? How easily can the action be reversed? Leaders must answer both questions for each action because a single workflow or agentic system may contain decisions with very different risk profiles.Consequence measures the effect of an error on people, finances, operations, and legal obligations. Reversibility measures how quickly, affordably, and completely an action can be undone.

Reversing an action does not necessarily repair the harm it caused. Together, these two dimensions produce four settings for enterprise autonomy.The four settings imply distinct operating models. High-consequence, low-reversibility actions require human authorisation. High-consequence, high-reversibility actions may permit bounded execution, but only when exposure limits, detection, and rollback are robust.

Low-consequence, low-reversibility actions require narrow permissions and pre-action checks. Low-consequence, high-reversibility actions are the strongest candidates for autonomous execution within tested limits. This framework directs scarce human attention to the decisions that materially affect outcomes. It separates execution capacity from decision authority: AI can perform more of the work, while broader decision rights are granted only when the consequences, reversibility, and operating controls justify them.Consequence and reversibility define the quadrants for the autonomy dial; capability and operational readiness determine how far autonomy can increase, while economics helps identify the most effective setting.

Leaders should measure the total cost of each accepted outcome, including AI tokens, platform charges, and human effort for review, retries, rework, exception handling, and recovery. Maximising token use is not inherently valuable, but minimising the cost of an AI interaction can undermine reliability. The foundation leaders must buildBefore increasing autonomy, business and technology leaders must examine the foundations beneath the dial.

In enterprises that have evolved over decades, data may be fragmented, siloed, inaccessible or largely unstructured, while critical work may depend on informal knowledge that agents cannot access. Even low-consequence, high-reversible actions require reliable context, tested boundaries, and documented exceptions. Human reviewers also need sufficient evidence and authority to assess an action and intervene.Enterprises must ensure that automation and a human+AI system are designed to operate as one unit.

Leaders can begin with a single workflow: document the exceptions recognized by process owners, beneficiaries, and executives; identify where critical knowledge resides; assess the consequence and reversibility of each action; and assign decision owners. They should then measure quality, human review time, exceptions and total cost per accepted outcome. Autonomy should expand only when evidence shows that the system can operate reliably within its defined boundaries.

This is how the autonomy dial becomes a repeatable business discipline rather than a race to delegate.The writer is CTO, Wipro Disclaimer: The views expressed are the author's own and do not reflect the official policy or position of Financial Express.