Autonomous Borders: Can AI Agents Replace Border Police?

Border control involves a variety of tasks, such as document verification, biometrics, visa and admissibility checks, travel analysis, and legal assessment. Some of these tasks are deterministic, some are probabilistic, and others require direct public authority and human judgment. Therefore, it is not correct to give a simple “yes” or “no” answer to the question, “Can AI replace border police?”

AI agents can automate a significant portion of document and data checks. However, entrusting decisions such as denial of entry, asylum or protection requests, and exceptional decisions affecting fundamental rights entirely to an autonomous model is not a safe approach.

What tasks can an AI agent handle?

• ePassport Passive Authentication and DS–CSCA chain verification
• MRZ, VIZ, and DG1 cross-verification
• Document validity, visa, and travel authorization checks
• Liveness, facial matching, and morphing risk signals
• Flight segment and transit consistency
• Queries regarding lost or stolen document services
• Presentation of multilingual standard questions
• Structured summary of evidence on the agent’s screen

Multi-agent architecture

Document Trust Agent
        │
Biometric Agent
        │
Visa / Admissibility Agent
        │
Journey Consistency Agent
        │
Policy Evaluation Agent
        │
        ▼
Evidence Aggregator
        │
        ├── CLEAR
        ├── MANUAL REVIEW
        └── CRITICAL ALERT
                 │
                 ▼
        Authorized Border Officer

Agents should be separated by task rather than consolidated into a single large model. While the Document Agent verifies digital signatures, the Visa Agent queries only the authorized data source. The Policy Agent evaluates current rules but cannot change them. The Evidence Aggregator consolidates the results; it has no authority to make legal decisions.

Common Outcome Agreement

{
  "caseRef": "border-case-pseudonym",
  "recommendation": "MANUAL_REVIEW",
  "checks": {
    "documentAuthenticity": "PASS",
    "biometricBinding": "PASS",
    "visaStatus": "REVIEW",
    "journeyConsistency": "PASS"
  },
  "reasonCodes": ["VISA_CONDITION_UNRESOLVED"],
  "humanDecisionRequired": true,
  "policyVersion": "border-policy-2026.07"
}

Sub-checks should be preserved rather than a single consolidated risk score. The case handler should be able to see which agent produced a result and based it on which source.

Example Orchestration with Python

from dataclasses import dataclass
from enum import Enum

class State(str, Enum):
    PASS = "PASS"
    REVIEW = "REVIEW"
    FAIL = "FAIL"
    UNAVAILABLE = "UNAVAILABLE"

@dataclass
class AgentResult:
    agent: str
    state: State
    reason_codes: list[str]
    evidence_refs: list[str]

def aggregate(results: list[AgentResult]):
    by_agent = {r.agent: r.state for r in results}
    reasons = [code for r in results for code in r.reason_codes]

    # Kritik güven kontrollerindeki hata otomatik geçişi durdurur.
    critical = {"document", "journey_authorization"}
    if any(r.agent in critical and r.state == State.FAIL for r in results):
        recommendation = "STOP_AND_REVIEW"
    elif any(r.state in {State.REVIEW, State.FAIL, State.UNAVAILABLE}
             for r in results):
        recommendation = "MANUAL_REVIEW"
    else:
        recommendation = "CLEAR"

    return {
        "recommendation": recommendation,
        "checks": by_agent,
        "reasonCodes": sorted(set(reasons)),
        "humanDecisionRequired": recommendation != "CLEAR"
    }

Advantages

• Speed and consistency in standard controls
• Reduced agent fatigue
• Multilingual passenger communication in a uniform format
• Comprehensive collection of decision evidence
• Allocation of specialized staff to complex cases
• Centralized management of policy and control versions

Disadvantages

• False alarms or dangerous misclassifications
• Biases in demographic and historical data
• Cascading propagation of agent errors
• The model’s misinterpretation of legal exceptions
• Uncertainty regarding liability among the institution, the official, and the provider
• Risk of cyberattacks, data poisoning, and prompt injection

Which decisions should remain with humans?

Decisions regarding entry denial, international protection, or asylum claims, as well as cases involving children, health issues, conflicting evidence, and high-impact profile assessments, should be referred to an authorized official. AI can summarize the evidence; it should not assume the role of providing binding legal interpretations or exercising discretion.

Conclusion

AI agents should not be designed to completely eliminate border police, but rather as a decision-support layer that accelerates document and data verification, organizes evidence, and makes uncertainty visible. The safest goal is not “autonomous border police,” but a human–AI team with defined roles and subject to oversight.

References

• Regulation (EU) 2024/1689 — AI Act
• NIST AI RMF Core
• Frontex Fundamental Rights Strategy