Verifying someone's identity is too high-stakes and too human to hand off to an AI agent.

The structural steps (check credentials, compare signals) are consistent, but each instance involves unique contextual factors — device history, behavioral anomalies, risk level — that require fresh judgment every time. Repeatability of structure does not translate to repeatability of decision quality.

Success criteria sound simple — is this the right person? — but in practice the answer is probabilistic and contested. There is no crisp, machine-readable ground truth available at decision time, making it impossible for an agent to know with confidence when the task is correctly done.

Meaningful identity verification requires access to government ID databases, biometric systems, behavioral baselines, device fingerprints, and fraud intelligence feeds — most of which are not available to a general-purpose AI agent. Even specialized systems only approximate the full picture.

A false positive grants an attacker access to an account, potentially enabling fraud, data theft, or harm to others. A false negative locks out a legitimate user and damages trust. Both errors can be irreversible in their downstream consequences, and liability exposure is real.

Genuine identity verification often requires interpreting ambiguous signals — unusual but legitimate travel, a user in distress, edge cases not covered by rules — that demand human intuition, ethical discretion, and accountability. AI cannot bear the responsibility this decision requires.