What is a decision record?
Logs tell you what your system did. A decision record tells someone else why the decision stands. It is the artifact you hand over when a decision is questioned — months later, by a person with no access to your systems.
A decision record is the durable, self-contained document of one AI decision. It carries the reasoning as structured arguments, the evidence as it existed at decision time, the policies that were evaluated, and the identity of the person who approved it. It is designed for the moment a decision is challenged: readable without a login, verifiable without trusting the vendor who produced it, and complete enough that a reader who was not present can follow how the conclusion was reached.
A log is not a record
A log entry answers "what happened, and when." A decision record answers "why is this decision defensible." The difference matters at the moment of challenge: a log tells a reviewer that an agent produced an output; it cannot tell them what alternatives were weighed, which evidence was actually available, or whether a person exercised judgment. Observability tooling is excellent at the first question and was never designed for the second — a distinction covered in depth under AI audit trails.
What belongs in a decision record
Four things, and all four have to survive independently of the system that produced them. The reasoning — the arguments for and against, structured rather than narrated, so a reader can see what was weighed. The evidence — snapshotted as it existed at decision time, because evidence that can silently change afterwards proves nothing. The policy — which rules were evaluated and which failed, failures first. And the authorisation — who approved it, and on whose behalf. Together these are what turn a trace into compliance evidence.
Why it has to be readable in ten years
A decision record's whole purpose is to be examined long after the decision. That rules out anything requiring a live account, a running service, or a vendor still being in business. A record that can only be read inside a portal is not a record; it is a view. This is why format matters as much as content — an archival document that carries its own machine-readable data as an attachment stays legible when everything around it has changed. It is also what regimes with multi-year retention duties, such as EU AI Act Article 12 record-keeping, implicitly assume.
Tamper-evidence, and what it does not prove
A decision record can be sealed so that any later alteration is detectable, and signed so a reader can verify it without trusting the storage it came from. That establishes integrity: this is the record as sealed. It does not establish that the content was true when written. No cryptography can prove that an agent reported its reasoning honestly, or that a human deliberated rather than clicked — which is why human oversight under Article 14 deserves its own structural treatment. Being precise about that boundary is what makes the rest of the claim credible.
What a decision record contains
Four elements, each answering a question a reviewer will actually ask.
Structured reasoning
The arguments for and against, as a graph rather than prose, so a reader can see what was weighed rather than what was summarised.
Evidence as it stood
Sources and values snapshotted at decision time, so the record reflects what was actually knowable then, not what is known now.
Policies evaluated
Which rules were checked and which failed, failures shown first, so exceptions are visible rather than buried.
Authorisation
Who approved the decision, including on-behalf-of delegation, recorded so that a machine approval can never be mistaken for a human one.
Together these turn "the agent decided X" into a record a reviewer can assess without reconstructing the moment from scratch.
Decision records: common questions
How is a decision record different from an audit log?
An audit log is a chronological stream of events across a system; a decision record is a bounded document about one decision. The log answers "what happened here"; the record answers "why does this decision stand." Most organisations need both, and they are not substitutes.
Does a decision record prove the decision was correct?
No, and any tool claiming otherwise is overselling. A decision record establishes what was decided, on what basis, and by whom — the inputs to a judgment about correctness, not the judgment itself. Whether the decision was right is a question for the reviewer.
Can a decision record be altered after the fact?
It can be sealed so that alteration is detectable. That is a meaningfully different guarantee from prevention: the record does not become physically immutable, but a reader can tell whether what they are holding is what was sealed.
Who actually reads these?
In practice: internal audit, external assessors, regulators responding to a complaint, and colleagues inheriting a decision. Each needs the same thing — enough context to evaluate a decision they were not present for.
Do we need a decision record for every AI decision?
No. The obligation attaches to decisions with consequences — those affecting a person's rights, access, money or safety. Recording everything at the same depth is expensive and makes the important records harder to find.
What format should a decision record be in?
Something archival and self-contained, that also carries machine-readable data. Human readability matters because people assess decisions; machine readability matters because systems need to search and verify them. A format that offers only one leaves the other audience stranded.
How does this relate to the EU AI Act?
Article 12 concerns record-keeping and Article 14 concerns human oversight, and both are ultimately evidenced by records of individual decisions. A decision record is one way to satisfy those obligations — it is not itself compliance, which depends on the whole system and how it is operated.
Related AI governance topics
AI Governance
The umbrella discipline: how organizations keep AI agents accountable, observable, and compliant — start here.
AI Observability
Seeing what your AI systems do in production — metrics, traces, and logs.
LLM Observability
Monitoring prompts, tokens, latency, and quality of large language model calls.
AI Traceability
Reconstructing the full lineage of an AI output — inputs, steps, and decisions.
LLM Traceability
End-to-end traces of multi-step LLM and prompt chains.
AI Agent Observability
Observability for autonomous, multi-step agents — tool calls, plans, and decisions.
Agentic AI Governance
Governing autonomous agents: policy, oversight, and accountable autonomy.
AI Audit Trail
Append-only, tamper-evident records of what an AI system decided and why.
AI Agent Monitoring
Real-time monitoring of agent behavior, drift, and decision quality.
Explainable AI (XAI)
Making AI decisions understandable to the people accountable for them.
AI TRiSM
Gartner's framework for AI trust, risk, and security management.
Decision Retrieval
GraphRAG for agents — retrieving past decisions as bounded, auditable packets.
AI Compliance Evidence
What auditors actually ask for, and why policy documents are not evidence.
AI Conformity Assessment
How an AI system is checked against the rules, and what that check consumes.
Decision Tracing
Capturing the structured reasoning behind every AI decision — AI Agentree's category.
AI Precedent Systems
Letting agents learn from past decisions as searchable precedent.
Decision Audit Trails
How human teams record why a decision was made — the deliberation counterpart to an AI audit trail.
Transparent AI
Making model reasoning inspectable, and what changes when several models are compared against each other.
Multi-Agent Simulation
Running many AI personas against one scenario to surface risks before a decision is taken.
See a decision record
The clearest way to understand the format is to look at one — the reasoning, the evidence, the policy trail and the approval, in a single document.
Explore the platform