Human Accountable
for the Loop
The accountability framework for agentic systems.
As AI systems move from generating outputs to taking actions, governance must move from reviewing decisions to governing the systems that produce them.
The shift
Human-in-the-Loop was designed for assistants. HAL was designed for systems.
When AI generated outputs, you could ask: was a human involved? When AI takes actions thousands of times an hour, that question stops working. HAL asks a different one.
Who is accountable for the system that made the decision?
Three models of involvement
From reviewing decisions to owning systems
Human-in-the-Loop
A human reviews each decision
A person sits between the model and the action. Nothing happens until they approve it. Designed for assistants that produce outputs.
- +
- Works when volume is low and a human can meaningfully review every case.
- −
- Breaks at scale: the reviewer becomes a rubber stamp.
Human-on-the-Loop
A human supervises the system
The system acts; a person monitors and can intervene. Oversight shifts from every decision to the aggregate.
- +
- Works when monitoring is real and intervention is fast.
- −
- Breaks when supervision is nominal and no one truly watches.
Human-Accountable-for-the-Loop
A human owns the system
Execution is delegated to the system within bounded authority. A named human remains accountable for what it does.
- +
- Scales when accountability is built into the system itself.
- −
- Requires real ownership, limits, evidence, and review to hold.
The framework
Eight domains of accountability
HAL evaluates a system across eight domains. Together they determine whether an organisation can responsibly delegate action while retaining accountability.
Autonomy multiplier
The more a system can do, the higher the bar
The required accountability scales with autonomy. An advice tool and an autonomous actor are not held to the same standard.
Advice System
Produces outputs for a human to use. Takes no action of its own.
- Research
- Drafting
- Summarisation
Recommendation System
Influences decisions by ranking, scoring, or classifying. A human still acts.
- Risk scoring
- Classification
- Triage
Execution System
Takes actions in systems of record, within bounded authority.
- Workflow triggering
- Record creation
- Notifications
Autonomous System
Takes consequential, externally-facing actions with real-world effect.
- Regulatory actions
- Client communications
- Contract execution
Where HAL fits
A different question
HAL does not replace the standards you already follow. It answers the question they leave open: when delegation is responsible.
- NIST AI RMF
- helps organisations understand AI risk.
- ISO 42001
- helps organisations build AI management systems.
- EU AI Act
- defines regulatory obligations.
- HAL
- answers when an organisation can responsibly delegate decisions and actions to AI while retaining clear human accountability.
Could you prove who is accountable for your agents?
Take the assessment to get a HAL Score across all eight domains in about five minutes.