AI in Government: How OBO Chains Solve the Accountability Problem
AI in Government: How OBO Chains Solve the Accountability Problem
Category: Government & Regulated Industries | Reading time: 5 min
The Accountability Requirement That Stops Deployments
More government AI deployments are blocked by the accountability question than by any technical challenge. The question is simple: "When an AI makes a decision that affects a citizen, who is accountable?"
Legal frameworks in most jurisdictions require that consequential government decisions be attributable to a human official with legal authority to make them. An AI cannot hold office. It cannot take an oath. It cannot be held legally responsible for a decision. Yet in a workflow where an AI agent processes a benefit claim, evaluates a permit application, or flags a document for review, the decision that matters happens at machine speed, often without any human seeing it.
The On-Behalf-Of (OBO) chain solves this by making every AI action legally attributable to the human official who authorised the agent to act.
The Legal Principle: Agency in Public Administration
Administrative law in most democratic systems recognises the concept of delegated authority — a senior official can delegate decision-making authority to a subordinate, but the senior official remains accountable for decisions made within the scope of that delegation.
When an AI agent acts in a government workflow, it acts as a delegated agent — but only if the delegation is documented, scoped, and attributable. An AI that "just processes things" is not a delegated agent; it's an unaccountable black box.
The OBO chain makes the delegation explicit:
- Official X (with legal authority) initiates the workflow
- The AI agent acts within the scope Official X authorised
- Every AI action is recorded as executed "on behalf of Official X"
- Official X is accountable for every action within that scope
This is the administrative law concept of delegated authority, implemented as a technical artifact.
A Benefits Processing Example
A social services department deploys an AI agent to process disability benefit applications. The workflow:
- Case worker initiates review for applicant #A-88421
- AI agent extracts and validates application data
- AI agent checks eligibility criteria against policy
- AI agent flags for human review or recommends determination
- Case supervisor reviews flagged cases and makes final determination
The OBO chain for this workflow:
{
"initiator": "caseworker.johnson",
"authority_scope": "disability_benefit_review",
"jurisdiction": "state_social_services",
"delegations": [
{
"principal": "caseworker.johnson",
"role": "case_worker",
"action": "initiated_benefit_review",
"case_id": "A-88421",
"timestamp": "2024-03-15T09:00:00Z"
},
{
"principal": "ai_eligibility_agent",
"action": "evaluated_eligibility_criteria",
"policy_version": "disability_policy_v12",
"result": "requires_supervisor_review",
"authorised_by": "caseworker.johnson",
"timestamp": "2024-03-15T09:00:01Z"
},
{
"principal": "supervisor.rodriguez",
"role": "case_supervisor",
"action": "reviewed_and_approved_determination",
"determination": "eligible",
"timestamp": "2024-03-15T09:45:00Z"
}
]
}
If the applicant challenges the determination: "Who made this decision?" — the OBO chain answers: Case Worker Johnson initiated the review, the AI evaluated under policy version 12, Supervisor Rodriguez made the final determination. Three accountable principals, fully documented, cryptographically verifiable.
High-Stakes Decisions: The Human Review Gate
For high-stakes determinations — large benefit amounts, appeal cases, cases involving medical evidence — the Rego policy requires human review before any determination is finalised:
package taos.gov.benefits
# High-stakes cases require supervisor review
require_supervisor_review {
input.benefit_amount_annual >= 50000
}
require_supervisor_review {
input.case_type == "appeal"
}
require_supervisor_review {
input.evidence_type == "medical"
input.medical_review_complete == false
}
action = "require_approval" { require_supervisor_review }
required_role = "case_supervisor" { require_supervisor_review }
The AI evaluates, extracts, and prepares. For cases matching these rules, a human supervisor makes the call. The policy is clear about which cases require which level of oversight — and it's enforced by the kernel, not by the case worker's judgment in the moment.
Freedom of Information: The Audit Trail as Public Record
Government agencies in many jurisdictions are subject to Freedom of Information (FOI/FOIA) requests. When citizens request documentation of AI-assisted decisions affecting them, the Taos audit trail provides:
- Every step the AI took in processing their case
- The policy version that governed the AI's evaluation
- Which specific eligibility rules applied to their case
- Whether any human reviewed the AI's work, and who
- The complete delegation chain from the initiating official to the final determination
This is not just a compliance response — it's a demonstration of accountable government. The citizen can see exactly how their case was handled, what rules applied, and who was responsible at each step.
Preventing Scope Creep: The ART Token
A persistent concern with government AI is scope creep — an AI system deployed for one purpose gradually being used for another. A benefits processing AI shouldn't be accessing immigration records. A permit evaluation AI shouldn't be querying financial data.
The Taos ART (Ambient Runtime Token) bounds the AI's capability at workflow initiation:
{
"workflow": "disability_benefit_review",
"allowed_tools": [
"fetch_application_data",
"check_eligibility_criteria",
"flag_for_review",
"record_determination"
],
"prohibited_data_sources": [
"immigration_records",
"criminal_records",
"financial_records"
]
}
The AI cannot access resources outside its allowed tool list, regardless of what the LLM decides to attempt. The scope limitation is enforced at the kernel level — it's not a prompt instruction that can be overridden.
The Bottom Line
Government AI doesn't fail at the technical layer — AI can process benefit claims, evaluate permits, and analyse documents with high accuracy. It fails at the accountability layer: when no human can be named as responsible for a consequential decision, the deployment is legally and politically untenable.
The OBO chain makes every AI action attributable to an authorised official. The policy engine makes the rules explicit. The audit trail makes the record verifiable. Accountability isn't a barrier to deploying AI in government — it's the foundation that makes deployment possible.
Tags: government AI, accountability, OBO chain, public administration, AI governance, FOIA, delegated authority