01The Chain of Custody Problem in AI Systems
In forensic investigations, chain of custody refers to the documented, unbroken sequence of possession and handling of evidence. If the chain is broken — if evidence passes through undocumented hands or is stored without verification — it becomes inadmissible. The record cannot be trusted.
AI systems create an analogous problem at operational scale. A customer inquiry arrives in Zendesk. An AI agent generates a response. The response is routed through an internal quality engine. A human reviewer approves it with edits. The final output is delivered to the customer and logged in Salesforce. Five systems. Three actors. One outcome. And in most enterprises, no single record that reconstructs this sequence with verifiable integrity.
The chain of custody problem in AI is not about whether work was performed. Vendor dashboards confirm that. The problem is whether the enterprise can independently reconstruct the complete lineage of a specific task — from initiation through completion — and verify that the record has not been altered after the fact.
Without this capability, enterprises operate on assertion rather than evidence. The vendor asserts the AI resolved the inquiry. The CRM asserts the ticket was closed. But no tamper-evident record proves the full sequence of events, identifies every actor involved, or confirms the output matched what was actually delivered.
02Why Logs Do Not Provide Chain of Custody
Traditional logging systems capture discrete events: an API call was made, a function returned, a database row was updated. These events are valuable for debugging and system monitoring. They do not provide chain of custody.
Chain of custody requires continuity — an unbroken sequence of evidence from task initiation through verified completion. Logs are discontinuous by design. Each system logs its own events independently. A Zendesk log records that a ticket was updated. An OpenAI log records that a completion was generated. A Salesforce log records that a case was closed. No single log captures the relationship between these events or verifies that they describe the same task.
Logs also lack integrity guarantees. Most logging systems store events in append-only databases that can be modified by administrators. Log entries can be deleted, altered, or backfilled without detection. This is acceptable for operational monitoring. It is not acceptable for audit evidence.
The distinction matters when an auditor, regulator, or internal investigator asks: "Show me the complete, verifiable record of how this specific AI-generated output was produced." Logs require reconstruction. Chain of custody requires replay — of sealed, tamper-evident records that were captured at the point of execution.
03Chain of Custody for AI Work
Preserving chain of custody for AI-performed work requires capturing six elements at every step of the task lifecycle:
These requirements describe a verification infrastructure that operates across system boundaries. No single vendor's telemetry can provide it, because no single vendor sees the complete task lifecycle.
04Verifiable Work Records
The Trusted Work Unit is designed to preserve chain of custody for AI work. Each TWU captures the complete evidence chain for a single task — every actor, every system, every input and output, every timestamp — and seals the record with a cryptographic hash computed from the full evidence sequence.
This produces a record with three properties that logs cannot provide:
These properties make TWUs admissible as evidence in the forensic sense. An auditor can replay the evidence chain, verify the hash, and confirm that the record has not been tampered with — without trusting any single vendor or system operator.
05Security and Compliance Implications
Chain of custody for AI work has direct implications across three enterprise functions.
Incident Response
When an AI system produces a harmful, inaccurate, or unauthorized output, the first question is: what happened? Without chain of custody, incident response requires manual reconstruction from fragmented logs — a process that is slow, unreliable, and vulnerable to gaps. With sealed work records, the incident response team can query the work ledger, retrieve the relevant TWU, and replay the complete evidence chain within minutes.
Regulatory Investigations
The EU AI Act requires organizations to maintain records that demonstrate transparency, human oversight, and accountability for high-risk AI systems. State-level AI legislation in the United States is introducing comparable requirements. These regulations do not require logs. They require evidence — verifiable records that prove compliance controls were operating and human oversight was applied. Chain of custody infrastructure produces the evidentiary artifacts these frameworks demand.
Enterprise Governance
For organizations deploying AI across multiple business functions, chain of custody provides the operational foundation for governance infrastructure. It enables executives to answer: which AI systems are performing what work, with what level of human oversight, and with what verified outcomes. Without verifiable work records, these questions are answered with vendor reports and internal estimates. With TWUs, they are answered with sealed evidence.