Human in the loop: designing supervision for AI agents

HITL: supervision, not build

Human in the loop is deciding where humans stay in an automated system. Build (flow, tools, business KPI) is in how to create an AI agent. Here: queues, thresholds, mental load, capture.

Three intents often mix: safety (block a dangerous action), quality before customer send, learning (labels to improve the system). If you do not separate them, you over-validate everything or miss risk.

Autonomy is not a switch. It is a dial: from pure suggestion to auto action on a strictly defined subset. Many serious business agents start between 20 and 60% effective autonomy by volume.

Without the third brick (control), you just have a POC / demo.

Where to place humans in the flow

Place HITL where error cost × error probability peaks. A standard follow-up email is low risk. A price change, legal send, or VIP client is high risk.

Map each action: read (often auto), draft (often auto), send (HITL at first), CRM write (HITL then auto on green), delete / pay / sign (permanent HITL or forbidden).

Risk #1 is review fatigue. If 100% of outputs need validation, the team rubber-stamps. Only queue amber and red, with a clear diff: what changed, sources, score.

Same logic on code: AI review speeds the first pass; humans keep ownership on auth and payments (see AI code review).

Thresholds and business policies

A model confidence score is not enough. Combine it with deterministic rules: amount over a threshold, VIP client, uncovered language, missing document, PII detected, tool outside allowlist.

Calibrate on an annotated sample (50 to 200 cases). Measure precision and recall of the green class. Too low: flooded queue. Too high: leaked errors. Recalibrate after every prompt or model change.

Write the policy in plain language, readable outside eng: “The agent may auto-send J+7 follow-ups if the quote is under €5k and the client is not VIP. Otherwise, sales validation.”

Version the policy like code: date, author, reason, impacted metric.

Capitalize on validations

Every accept, edit, or reject is useful data. Store input, agent output, human action, comment, review time.

Use the log for corrected few-shot examples, recurring error classes, autonomy candidates (always accepted without edit), validator training.

GDPR: minimize PII in logs, set retention, restrict access. HITL does not excuse an ungoverned data lake.

Ship HITL in 30 days

Week 1: action map and risk class. Week 2: minimal validation UI and logs. Week 3: threshold calibration on real corpus. Week 4: metrics and first green expansion.

Roles: business owner (rules), tech owner (tools and logs), trained validators with a short checklist.

HITL health: median review time, % green, % edits, post-action incidents, mental load. If load explodes, shrink auto scope.

If you want clean HITL on a real process, we can scope in 20-40 minutes.