What Are Shadow Decisions in
Agentic AI?

A shadow decision is a consequential choice made by an AI agent that no one in the enterprise explicitly authorized, made by an agent operating exactly as configured.

The definition’s second clause is the important one. A shadow decision is not a malfunction, an error, or a misuse. Every component performs as designed. The failure is that real authority was delegated, and nobody remembers deciding to delegate it.

That distinction changes what the review should ask. An error review asks why the agent was wrong. A shadow decision review asks why the agent had the power. In my experience, organizations run the first review reflexively and miss the second entirely, because the second question implicates the organization rather than the technology.

How Do Shadow Decisions Form?

Shadow decisions usually form through three mechanisms, often in combination:

• Latitude accretion: an agent’s mandate widens through small tuning changes, each made to improve performance, none reviewed as a delegation decision, until the accumulated latitude bears no resemblance to anything leadership approved.
• Configuration as policy: parameters set by technical professionals for technical reasons quietly become de facto enterprise policy, because a threshold in a config file is a policy whether or not anyone calls it one.
• The assumption gap: every team in the chain assumes some other team approved the authority, and the assumption is load-bearing precisely because it is never tested.

A shadow decision review does not ask why the agent was wrong. It asks why the agent had the power, and most enterprises have never asked.

Why Do Shadow Decisions Stay Hidden?

Here is the uncomfortable part: shadow decisions hide behind good outcomes.

The most dangerous agent in your enterprise may be performing flawlessly, because flawless performance is exactly what prevents anyone from auditing the authority underneath it.

Reviews are triggered by bad results, and shadow authority produces bad results only when conditions change. At that point, the organization discovers under incident pressure that an agent has been exercising powers nobody mapped.

Waiting for the outcome to expose the authority is a strategy in the same sense that not checking the wiring until the fire is a strategy. The exposure scales quietly: a single unexamined authority, exercised across thousands of transactions a month, accumulates enterprise commitment at a volume no individual shadow decision would ever suggest.

What Do Shadow Decisions Look Like in Practice?

Consider an international airline’s disruption-recovery agent, which rebooked passengers and issued compensation during irregular operations.

Through one difficult storm season, the agent was tuned eleven times, always for the same defensible reason: customer satisfaction scores. Each change slightly widened what it could offer.

By season’s end, the agent was proactively issuing upgrades and vouchers at levels finance had never approved. The discovery came not from an incident, but from a cost anomaly in a quarterly review.

The audit finding was the instructive part: the agent’s effective mandate had roughly tripled through changes that were each individually reasonable, and not one had been reviewed as what it actually was — a delegation of financial authority.

The remedy was procedural, not technical: a decision-rights audit comparing what every agent could decide against what leadership knew it could decide, with the gap — the shadow authority — either ratified explicitly or withdrawn. Ratified is a legitimate outcome. Unknown is the only non-legitimate outcome.

Decision-rights audits, and the boundary redesign that follows them, are work we lead through Inteq’s Agentic AI consulting practice. The full white paper this post accompanies makes the broader operating model argument.