Use Cases

Where AI decisions have consequences

Headlights is built for environments where an AI agent's reasoning must be visible, auditable, and defensible — not just functional.

Energy / Utilities

Network Incident Triage

AI agents diagnose and remediate network faults in real time — routing around failures, adjusting load, escalating to human operators. When something goes wrong, operators need to see every step the agent took and why.

"The agent recommended failover to R-02 at 3:14am. Three minutes later we had a cascading fault. What was it reasoning from? What did it miss?"

Example Trace — Network Incident NOC-004Critical

Reasoning

Alert detected

Packet loss at 34% on SW-07. Threshold exceeded. Initiating triage.

120ms
success

Tool Call

Query telemetry logs

get_telemetry_logs({device: "SW-07", window: "15m"}) → anomaly confirmed

840ms
success

Failed

Run network diagnostic

run_network_diagnostic({scope: "full"}) → timeout after 3200ms

3.2s
failure

Retry

Retry with reduced scope

Upstream link degraded. Loss 34%. Downstream nominal.

1.1s
warning

Output

Recommend remediation

Failover to R-02. Window: 8 minutes before SLA breach.

90ms
success

Energy / Utilities

Grid Stability Monitoring

Agents continuously balance load across distributed energy assets — solar, wind, storage, demand response. Every automated dispatch decision needs to be traceable for regulatory compliance and post-incident review.

"The regulator is asking why we curtailed the wind farm at 14:32. The agent made that call. We need its reasoning, not just the outcome."

Example Trace — Grid Dispatch GD-2891High

Reasoning

Frequency deviation detected

Grid frequency at 49.7Hz. Below 49.8Hz threshold. Assessing response options.

80ms
success

Tool Call

Query available reserves

get_reserve_capacity() → Battery A: 42MW, Hydro: 120MW available

220ms
success

Tool Call

Dispatch battery reserve

dispatch_asset({asset: "Battery-A", mw: 40, duration: "5min"}) → dispatched

310ms
success

Output

Frequency restored

Grid at 50.01Hz. Reserve deployed. No curtailment required. Logged for AEMO.

60ms
success

Government

Automated Compliance Decisions

Government agencies are deploying AI to assist with permit approvals, benefit assessments, and regulatory monitoring. When those decisions affect citizens, every reasoning step must be auditable and explainable.

"Under the APS AI framework, we need to demonstrate that every automated decision was explainable and had a human review pathway. Can you show us the agent's reasoning chain?"

Example Trace — Permit Assessment PA-1140Warning

Reasoning

Application received

Environmental permit application for industrial site. Checking against 14 criteria.

140ms
success

Tool Call

Check environmental zone

lookup_zone({coordinates: "-33.8, 151.2"}) → Sensitive coastal buffer

480ms
success

Warning

Risk threshold exceeded

Site falls within 500m buffer. Automated approval not permitted. Escalating.

90ms
warning

Output

Referred to human review

Reason: coastal buffer zone. Assessor notified. Full trace attached to record.

70ms
success

Critical Infrastructure

Water Treatment Operations

AI agents monitor treatment plant sensors, adjust chemical dosing, and flag anomalies before they become safety incidents. The reasoning behind every dosing decision must be traceable to the sensor data and thresholds used.

"Chlorine levels spiked at 6am. The agent adjusted the dosing pump. We need to trace exactly what sensor readings it was acting on and whether the response was proportionate."

Example Trace — Dosing Anomaly WTP-0094Critical

Reasoning

Chlorine anomaly detected

Residual at 4.2mg/L. Safe range: 0.5–2.0mg/L. Immediate assessment required.

95ms
success

Tool Call

Read upstream sensors

get_sensor_array({zone: "dosing-A"}) → flow rate normal, dosing pump fault flag

340ms
success

Tool Call

Reduce pump output

set_pump_rate({pump: "CL-A2", rate: 0.3}) → acknowledged

210ms
success

Output

Levels normalising

Residual trending to 1.8mg/L. Operator alerted. Maintenance ticket raised.

80ms
success

Telecommunications

Network Fault Management

Telcos run AI agents that autonomously reroute traffic, isolate faults, and trigger field dispatch across tens of thousands of nodes. At scale, every automated intervention needs to be logged, traceable, and defensible to regulators and enterprise customers.

"We had 40,000 customers affected for 18 minutes. The agent rerouted traffic three times before escalating. The ACMA is asking for the reasoning chain behind each decision."

Example Trace — Outage Response OR-7741Critical

Reasoning

Core node failure detected

Node SYD-CORE-04 unresponsive. 40,200 downstream subscribers affected.

60ms
success

Tool Call

Assess alternate routing paths

get_routing_topology({region: "SYD"}) → 2 paths, capacity 68% and 91%

410ms
success

Retry

First reroute failed — capacity exceeded

Path A saturated at 103%. Retrying via Path B (91% capacity).

890ms
warning

Output

Service restored — 18 min outage

All subscribers online. Incident logged with full trace for ACMA reporting.

70ms
success

Mining & Resources

Autonomous Operations Safety

Mining operations deploy AI agents to manage autonomous haul fleets, monitor ground stability, and respond to safety events. When a vehicle stops or a zone is locked out, the reasoning behind that decision must be captured for safety audits and regulatory reporting.

"The autonomous fleet halted Pit 3 at 2am. The site manager wants to know exactly what the agent detected, what it ruled out, and why it chose a full stop over a partial restriction."

Example Trace — Safety Event SE-0312Critical

Reasoning

Ground movement anomaly detected

Slope sensor SM-14 reading 4.2mm displacement. Threshold: 3mm.

110ms
success

Tool Call

Cross-check adjacent sensors

get_sensor_cluster({zone: "pit-3-north"}) → SM-13: 3.8mm, SM-15: 4.1mm. Corroborated.

280ms
success

Tool Call

Halt autonomous fleet — Pit 3

halt_fleet({zone: "pit-3", reason: "slope_instability"}) → 14 vehicles stopped

190ms
success

Warning

Geotechnical engineer escalation

Displacement rate exceeds autonomous threshold. Human review required before resuming.

85ms
warning

Output

Zone locked — engineer notified

Pit 3 locked. On-call geotechnical engineer paged. Full sensor trace attached. WHS log created.

95ms
success

Government Department

Benefits & Services Decision Support

Federal and state agencies are deploying AI agents to assist case officers with eligibility assessments, fraud detection, and service routing. Under APS AI guidelines, every automated recommendation affecting a citizen must have a complete, explainable audit trail.

"A citizen has challenged an automated benefit decision through the AAT. We need to produce the full reasoning chain the AI used — every data point checked, every rule applied, every flag raised."

Example Trace — Eligibility Assessment EA-44821Warning

Reasoning

Application received for assessment

JobSeeker application. 14 eligibility criteria. Residency, income, activity test.

90ms
success

Tool Call

Verify income and assets

check_income_assets({tfn: "***redacted***"}) → Income: $0. Assets: $4,200. Within threshold.

640ms
success

Tool Call

Check activity test compliance

get_activity_log({id: "44821"}) → 3 job applications in period. Minimum 4 required.

380ms
success

Flag

Activity test shortfall — referral triggered

Below minimum job search requirement. Cannot auto-approve. Refer to case officer.

75ms
warning

Output

Referred to human case officer

Reason: activity test shortfall (3/4). Full trace logged. Applicant notified with reason code.

80ms
success

"Headlights isn't a debugger. It's the audit trail that answers the question every infrastructure operator will face: what was the AI reasoning when it made that call?"

Headlights — AI Governance for Critical Infrastructure

2026 Cohort

See it working on a real incident trace

The live demo shows a Network Incident Triage Agent reasoning through a critical fault — step by step.

View Live Demo → Analyse Your Spend → Book a Pilot