Human in the loop
"We keep a human in the loop" is the most repeated sentence in enterprise AI, and the least
specified. In which step? Before or after the agent acts? Reviewing everything or a sample? With
authority to override, or just to observe? A loop that is not drawn is a loop that will not survive
contact with month-end pressure.
Three positions for the human
- In the loop. The agent prepares or recommends; a person approves before anything happens. Right for consequential, hard-to-reverse decisions.
- On the loop. The agent acts; people monitor dashboards, samples and alerts, and can intervene. Right for high-volume, low-impact, reversible work.
- Out of the loop. The agent acts unobserved. Legitimate only where impact is trivial, and even then with corrective controls behind it.
The patterns, drawn
On the canvas, human involvement is not a caption, it is structure. A hybrid step
puts the person and the agent on the same task, with the agent recommending and the person deciding.
A human gate blocks the flow until someone approves, the pattern for payments,
offers and customer communication. A detective control implements on-the-loop
monitoring: sampling, drift alerts, weekly reviews. The governance check tells you when a step's
risk demands a stronger pattern than the one you drew, one of the 11 rules of the
governance framework.
Keeping the loop honest
Two failure modes eat HITL designs. Rubber-stamping: the reviewer approves everything because the
queue is long, which detective sampling and workload numbers expose. And silent scope creep: an
agent that recommends today quietly starts acting tomorrow because nobody re-checked the design.
Version the design, re-run the check after every change, and the loop stays real. This is also why
autonomy levels live on each step rather than in a policy PDF.
🤖 Classify claim type and urgency
Humans 👤 Assess coverage
Hybrid (HITL) 🤝 Review AI recommendation
AI Agents 🤖 Extract claim data
Systems ⚙️ Send SEPA payment
The Hybrid lane is the human in the loop, drawn where everyone can see it.
Frequently asked questions
What does human in the loop mean?
Human in the loop (HITL) means a person participates in an AI system's decision cycle: reviewing, correcting or approving before the outcome takes effect. It contrasts with human on the loop, where a person monitors and can intervene but the AI acts by default, and human out of the loop, where the AI acts autonomously.
When is human in the loop required?
As a design rule: when the step's impact and likelihood of harm are high, when decisions materially affect people (hiring, credit, claims, disciplinary action), when regulation demands accountability, or when the model operates far from its training distribution. Low-risk, reversible, high-volume steps are usually better served by human on the loop with detective controls.
Does human in the loop kill the efficiency gains?
Only when applied indiscriminately. The pattern that preserves gains is recommend-then-approve on the risky fraction: the agent handles extraction, drafting and routing autonomously, and a person reviews the cases the agent itself flags, thresholds route the rest. Designs that keep humans on judgement and take them off transcription tend to win on both speed and quality.
How do I design HITL into a process instead of bolting it on?
Model it explicitly: hybrid steps where person and agent share a task, human gates before consequential actions, and an accountable person named on every autonomous agent. If your process map cannot represent those three, the loop exists only in good intentions.
Draw your loop before you deploy it
The studio is in private preview. Open it to request access, or walk through a real, read-only
example first.