Is this task ready for an AI agent?
Evaluate one candidate task against the five TRACE criteria (Traceability, Reversibility, Acceptance Criteria, Compliance, and Escalation) before any vendor selection or architecture decision begins.
Listen to this briefing
The TRACE AI Pre-Flight Checklist
0 of 10 questions answered
Name the task you are evaluating
TRACE evaluates one specific task at a time. Be precise: "process customer refunds" is too broad; "validate and approve refund requests under $500 in the CRM system" is evaluable.
Be specific enough that someone unfamiliar with your operations could understand what this agent would do.
Traceability of Inputs and Outputs
Every input consumed and output produced must be logged, timestamped, and attributed to specific data sources. Tasks drawing on unverified sources should be deprioritised until provenance controls are in place.
Can every input to this task (data, triggers, parameters) be logged, timestamped, and attributed to a specific source?
Can every output this task produces be recorded, attributed, and audited against its source inputs?
Are data provenance controls currently in place for the sources this task relies on?
Reversibility and Risk Tier
Tasks are classified by the consequence of an erroneous output: reversible with no harm, reversible with cost, or irreversible. Irreversible-consequence tasks require human-in-the-loop checkpoints before any agent autonomy is granted.
If this agent produces an erroneous output on this task, what is the worst-case consequence?
Acceptance Criteria Definition
Success criteria must be measurable and pre-agreed independent of the agent's output. A task lacking an external ground truth or validation method is not ready for deployment regardless of technical feasibility.
Does this task have measurable, pre-agreed success criteria that exist independently of what the agent outputs?
Is there an external ground truth or independent validation method that can confirm whether the agent's output is correct?
Compliance and Regulatory Mapping
Applicable regulatory regimes, sector codes, and governance policies must be identified before architecture decisions. This is where a task's exposure to regimes such as the EU AI Act, GDPR, and the NIST AI Risk Management Framework gets scoped. For government and regulated environments, this mapping is a prerequisite to scoping.
Have the regulatory regimes (for example the EU AI Act, GDPR, or NIST AI RMF), sector codes, and data protection requirements applicable to this task been identified?
Have applicable internal governance policies and approval processes been mapped before any agent architecture decisions?
Escalation and Override Pathway
A clear, tested mechanism must exist for a human operator to intervene, override, or halt the agent mid-task. Absence of this pathway disqualifies a task from autonomous execution regardless of model performance.
Does a clear, documented mechanism exist for a human operator to intervene in, override, or halt this agent mid-task?
Has this escalation and override pathway been tested in a realistic scenario?
Traceability
Reversibility
Acceptance Criteria
Compliance
Escalation
This tool produces a self-assessment against the TRACE framework based on the answers you gave. It does not constitute a certification, an accredited audit or legal advice. The guidance above is the starting point for your own governance review, which remains your responsibility. Your answers stay in your browser and are neither sent to Terence Kok nor reviewed by anyone.
What each criterion tests and why it comes before architecture
TRACE is an original framework and does not implement any external standard. Its five letters loosely map onto the instrumentation OpenTelemetry's GenAI semantic conventions now define for agent traces (orchestration, tool-call, model and memory spans), and the mapping is conceptual with no claim of formal conformance.
TTraceability of Inputs and Outputs
Without full logging of what data entered the task and what output was produced, you cannot audit errors, satisfy regulators, or improve the system. Traceability is the precondition for accountability.
When this fails: Any erroneous output becomes unattributable. You cannot tell what caused it, cannot correct the root issue, and cannot demonstrate due diligence to auditors or affected parties.
RReversibility and Risk Tier
The risk classification determines how much oversight is required before granting autonomy. Low-reversibility tasks demand human review before any consequence propagates. Committing to autonomous execution on an irreversible task without a human checkpoint is not a risk tolerance decision. It is a governance gap.
When this fails: Irreversible errors occur without human review. A single mistaken output in a high-stakes domain can generate costs or harms that no amount of post-incident process can fully reverse.
AAcceptance Criteria Definition
An agent that grades its own output is not evaluated. It is trusted. Acceptance criteria must exist independently of the agent, defined before deployment begins. Without an external ground truth, you have no basis for judging whether the agent is performing correctly, improving, or degrading.
When this fails: You cannot distinguish a performing agent from a malfunctioning one. The first indication of systematic error is typically a downstream incident, because no monitoring alert exists to raise it earlier.
CCompliance and Regulatory Mapping
Discovering applicable regulations after architecture decisions have been made forces redesign, delays, or deployment with known compliance gaps. The mapping must precede design, not follow it. For most agent tasks this means checking exposure to the EU AI Act, GDPR, and the NIST AI Risk Management Framework at minimum, alongside any sector code. For cross-jurisdiction deployments, identify the most restrictive regime first and design to it.
When this fails: The deployment violates a regulatory requirement that was not identified until after go-live. Remediation requires suspension, redesign, and regulatory engagement that could have been avoided with a pre-deployment compliance inventory.
EEscalation and Override Pathway
A tested override mechanism is the non-negotiable minimum for any autonomous agent deployment. "Tested" is the operative word: a pathway that exists only on paper has unknown failure modes. The escalation pathway is what converts autonomous execution into accountable execution.
When this fails: When an agent behaves unexpectedly, no tested mechanism exists to halt it. The time spent constructing a manual stop while the system continues to run is the cost of this gap.
How to interpret the three outcomes
All five criteria pass. This task is structurally appropriate for autonomous agent deployment. Proceed to architecture design and vendor selection, keeping the first deployment to no more than three to five tasks under close monitoring.
One or more criteria are partially met. The task may proceed to scoping, but the amber criteria must be resolved before deployment begins. Identify whether the gap can be closed in 30 days; if not, treat the task as not yet ready.
Any criterion fails. The one exception is Reversibility failing on its own: that task is marked Human-in-the-Loop Required, and it cannot run autonomously until a human checkpoint is designed. The task should remain manual or assisted until the failing criteria are remediated. Deploying before remediation is not a calculated risk. It is an unmanaged one.
Answer points per criterion and the rule that sets the verdict
Thresholds as the tool applies them on this page. Traceability passes at 5 of 6 points and is conditional from 3; the two-question criteria pass at 4 of 4 and are conditional from 2. An irreversible task can reach autonomy only through a human checkpoint.
The full TRACE methodology is in the article
The article behind this tool sets out the full TRACE methodology, a leadership checklist for scoping your first task cohort, and why the task inventory belongs in front of leadership before procurement begins.
Once a task passes TRACE and an agent is designed for it, rate that agent on the Agent Risk Assessment Matrix to get its oversight tier and the ISO/IEC 42001 control set for the register.
TRACE came out of watching teams jump straight to picking a vendor before anyone had asked whether the task itself was even fit for an agent. Reversibility is the criterion that trips people up most: it forces you to name the worst case before you've built anything. A conditional or not-ready result is the cheapest information you will get on this task all year.
Last updated 29 Sept 2026
More in Agent labs
- Advanced AI Agent BuilderWhat changes when agents work as a team?
- Expert AI Agent BuilderWould our guardrails hold against a real attack?
