Human in the Loop

Pauses a workflow and routes content to assigned reviewers for approval before it continues

What does the Human in the Loop node do?#

The Human in the Loop node pauses workflow execution and routes content to one or more assigned reviewers for a decision before resuming. Each reviewer receives a review request (email + review page) where they read the content and either approve or reject it. The workflow only continues when the configured approval policy is satisfied — and any rejection stops the step.

Use it as a manual checkpoint at any critical step — to review AI-generated content, validate data, or approve actions before they run downstream.

Common use cases:

• Reviewing AI-generated articles or summaries before publishing them to a CMS.
• Validating extracted data before writing it to an external system (CRM, database, spreadsheet).
• Quality-checking email content or customer-facing messages before they are sent.
• Approving sensitive or irreversible operations before they execute.

Quick setup#

Follow these steps to add and configure the Human in the Loop node in your workflow:

Configuration parameters#

Required fields#

Optional fields#

Approval modes#

The Approval mode decides how many of the assigned reviewers must approve. A rejection is always an early exit — one reject stops the step immediately, in every mode.

Mode	Reviewers	Passes when	Fails when
Any of	≥ 1	The first reviewer approves	The first reviewer rejects
All of	≥ 2	Every reviewer approves	Any reviewer rejects
Quorum (N of M)	≥ 3	N reviewers approve	N approvals become mathematically unreachable, or any reviewer rejects

• Any of — the first decision wins. Best for “anyone qualified can sign off”.
• All of — unanimous approval required. Best for legal review or peer sign-off.
• Quorum (N of M) — a configurable number must approve. Resolves as soon as N approvals are reached, or fails as soon as N can no longer be reached. Set N with the Required approvals (N) field (1 < N < M).

When one reviewer’s decision resolves the step, the other pending review requests are automatically closed as “not needed”.

Scoring#

Scores let you attach automatic, rule-based ratings to the content so reviewers get an at-a-glance signal before deciding. Each score you declare:

• adds a dedicated input port on the node (wire it to the upstream value to rate), and
• renders a colored card (good / neutral / bad) on the review page.

Each score has a type and up to three rule bands, evaluated in order good → neutral → bad (first match wins):

Type	Operators	Bands
number / percentage	gt, gte, lt, lte, eq, between	good · neutral · bad
string	equals, contains, starts_with, matches (regex), with optional case sensitivity	good · neutral · bad
boolean	is	good · bad
enum	in, not_in (over an allowed-values list)	good · bad

• The good band is required; neutral and bad are optional.
• If a value is provided but no declared band matches, the card falls back to bad — a builder who set a “good” threshold expects anything outside it to read as a failure.
• A missing value, or a score with no declared rules, shows as Unrated.

Scores are informational — they guide the reviewer but never approve or reject on their own.

Batch review#

When a Human in the Loop node runs inside a Loop, each iteration creates a separate review request. These are grouped into a batch review page where a reviewer can work through many items efficiently:

• Preview — a one-line, templatable summary shown under the scores on each card, so a reviewer can identify an item without opening it.
• Filters for batch review — templatable dropdowns (e.g. by category, author, status) whose options come from the distinct resolved values across the batch.

Both use {{variable}} placeholders, and every referenced variable becomes an extra input port on the node.

What does the node output?#

The node outputs the content as approved by the reviewers. The workflow stays paused until the approval policy is met, then execution continues with that content as the next node’s input. If the step is rejected, the output does not flow and the branch stops there — pair with a Fail Node or a Conditional to handle rejection explicitly.

How to use the output#

In Draft & Goal, you don’t need to look up a complex system-generated variable name. To use the result:

1. Draw a connection from the Human in the Loop node’s Output port.
2. Connect it to the next node’s input.
3. In that next node, create and name your own variable (for example, approved_content). The reviewed value will be injected into it automatically.

Input ports#

The node always has a static Content port. It grows additional ports dynamically:

• one port per declared score, and
• one port per {{variable}} used in the Message to reviewer, Preview, or Filters.

Usage examples#

Example 1: Approve an AI-generated article before publishing to Notion#

Make sure an LLM-generated article is accurate and on-brand before it lands on your CMS.

Workflow:

1. Web Scraper — Fetch a source page used as raw material.
2. LLM — Generate a structured article from that source.
3. Human in the Loop — Content connected to the LLM output. Reviewers = your content lead. Approval mode = Any of. Message to reviewer = “Review this article for factual accuracy, tone, and completeness before publishing to Notion”.
4. Notion Database Writer — Receives the approved article and publishes it.

Example 2: Unanimous legal sign-off before sending a contract#

Require every reviewer to approve before an automated email leaves the system.

Workflow:

1. LLM — Draft a contract clause from a template.
2. Human in the Loop — Content connected to the draft. Reviewers = legal + account manager. Approval mode = All of. A score of type boolean wired to a compliance flag surfaces a red/green card to speed the review.
3. Email Sender — Sends the approved contract only after both reviewers approve.

Best practices#

• Place a Human in the Loop node just before any irreversible action — sending emails, writing to databases, publishing to a CMS, updating a CRM. A short human check is much cheaper than rolling back a bad write to production.
• Write a specific message: name the exact fields, tone, and facts to verify instead of asking for a generic “review”.
• Add scores for anything a rule can pre-check (length, presence of a CTA, a compliance boolean) so reviewers spend attention on judgment calls, not mechanical checks.
• Match the approval mode to the risk: Any of for routine sign-off, All of for unanimous decisions, Quorum for balanced committee review.
• Leave yourself out of the reviewer list on steps you shouldn’t be able to self-approve.

Common issues#

The workflow is stuck on the Human in the Loop node

Cause: The node is doing its job — it pauses the workflow until the approval policy is met. There is no automatic timeout (reminders are “coming soon” and not yet active).

Solution: Open the workflow execution panel and look for a pending review request on this node. A reviewer must open it and approve for the workflow to resume. In All of mode, every reviewer must act; in Quorum mode, N reviewers must approve.

No reviewers can be selected / the field is empty

Cause: Reviewers are drawn from the users in your organization. If the picker is empty, no eligible users exist in the org yet.

Solution: Invite or assign at least one user with a reviewer-capable role in User Management, then reopen the node settings.

The reviewer doesn't know what to check

Cause: The message to the reviewer is empty or too vague, so they only see raw content with no acceptance criteria.

Solution: Fill the message with a precise instruction listing what to verify (specific fields, tone, facts, formatting). Reference upstream values with {{variable}} to make it concrete.

No content reaches the reviewer

Cause: The Content input is not connected, or the upstream node produced an empty output.

Solution: Verify the Content port is connected to the correct upstream node. Run the upstream node alone to confirm it produces non-empty output before reconnecting.

Quorum mode won't accept my N value

Cause: Quorum requires 1 < N < M. N = 1 is identical to Any of, and N = M is identical to All of, so both are rejected. Quorum also needs at least 3 reviewers.

Solution: Add reviewers until you have at least 3, then set N strictly between 1 and the reviewer count.

How does it fit into a workflow?#

Human in the Loop typically acts as an approval gate between content generation and any irreversible downstream action. Here’s a typical integration pattern for AI-generated content that ends up in an external system:

graph LR
    Source[Web Scraper / Data source] --> LLM[LLM generates content]
    LLM --> HITL[Human in the Loop
<br/>reviewers approve]
    HITL --> Publish[Notion / Email / CRM writer]

Related nodes#