Resources / Product Brief

Human-in-the-Loop, Product Brief.

Confidence thresholds, escalation triggers, pre-send review queues, override controls, and full audit trail. Every AI decision is observable, reviewable, and overridable, by design.

Book a demo See the pillar →

What Human-in-the-Loop does

Ephanti’s Human-in-the-Loop (HITL) framework ensures that AI autonomy is calibrated, not assumed. Every conversation type, every customer tier, and every channel can have different oversight rules.

HITL is not an emergency stop bolted on top of an autonomous system. It is built into the MEVA reasoning loop. Every AI decision is scored for confidence, checked against business rules, and either acted on autonomously, queued for human approval, or immediately escalated, based on thresholds you configure.

For regulated industries, enterprise procurement teams, and organisations where AI failure carries brand or legal risk, HITL is the governance layer that makes AI deployment possible.

How it works

1. AI reasons

MEVA generates a proposed response or action with a confidence score.

2. Rules check

Score checked against configured thresholds. Above threshold: act. Below: queue or escalate.

3. Human reviews

Agent approves, edits, or rejects from the review queue. Context and reasoning shown.

4. Logged

Every decision, override, and escalation logged to the audit trail with full metadata.

Key capabilities

Confidence thresholds: per channel, per intent type, per customer tier
Supervised mode, all AI responses queued for human approval before send
Autonomous mode, AI acts directly above confidence threshold
Escalation triggers: sentiment, intent type, conversation value, tier, SLA breach
Pre-send review queue: agent sees AI draft, can edit or approve with one click
Override controls: suppress, edit, or replace any AI-generated response
Override feedback loop, overrides feed back to improve AI accuracy over time
Full audit trail, every AI decision, human override, and escalation with timestamp and agent ID
Audit export: for compliance, legal review, or SOC 2 evidence
Per-agent supervised mode: new agents start supervised, graduate to autonomous on performance

Configuration options

Confidence threshold per intent category (support, sales, compliance, sensitive topics)
Supervised vs autonomous mode per channel, per queue, per agent
Escalation trigger definitions, which signals route to human immediately
Review queue assignment, which agents review which conversation types
Audit retention: default 13 months, configurable on Enterprise
Override notification, alert supervisor when an agent overrides an AI response

Tier availability

Capability	Starter	Pro	Enterprise
Basic escalation triggers	✓	✓	✓
Pre-send review queue	✓	✓	✓
Confidence thresholds	Fixed	✓	✓
Supervised mode per channel	,	✓	✓
Audit trail	90 days	13 months	Custom
Audit export (compliance)	,	,	✓
Per-agent supervised mode	,	✓	✓
Custom escalation rules	,	,	✓

FAQs

How does Ephanti decide when to escalate to a human? ⌄

Escalation triggers include: AI confidence below your configured threshold, negative sentiment detected, customer tier flagged for human touch, conversation value above a set amount, specific intent types (complaints, legal, medical), and manual agent request.

Can I review and override an AI response before it is sent? ⌄

Yes. In supervised mode, all AI-drafted responses are queued for human approval before send. Agents see the draft, the confidence score, and the reasoning. They can approve with one click, edit, or suppress.

What does the audit trail capture? ⌄

Every AI decision, human override, escalation trigger, and sent message is logged with timestamp, agent ID, confidence score, intent classification, and the version of the knowledge base used at decision time.

Is HITL available on all tiers? ⌄

Basic escalation and pre-send review are available on all tiers. Configurable confidence thresholds, per-channel supervised mode, and compliance audit export require Pro or Enterprise.