10.4 Human-in-the-Loop Workflows

Complete automation isn't always what you want. Sometimes you need human judgment, approval gates, or review steps before taking important actions. This section shows you how to design workflows that combine the speed of automation with the wisdom of human oversight.

What Is Human-in-the-Loop?

The Concept

Human-in-the-Loop (HITL) means placing humans at strategic points in automated workflows.

Not: Humans vs. Machines

Instead: Humans + Machines working together

⚡

Machines excel at

•Processing speed
•Consistency
•Handling volume
•Pattern recognition
•Data extraction

🧠

Humans excel at

•Judgment calls
•Nuanced decisions
•Creative solutions
•Edge case handling
•Risk assessment

HITL combines both strengths

Real-World Analogy: Airport Security

❌

Full automation (no humans)

Baggage → X-ray scan → Algorithm decides → Auto-allow/deny

Problem: Algorithm might miss threats or flag innocuous items

⚠️

Full manual (no automation)

Baggage → Human inspects each bag thoroughly

Problem: Too slow, inconsistent, expensive

✅

Human-in-the-loop (hybrid)

Baggage → X-ray scan → Algorithm flags suspicious → Human reviews flagged items → Decision

Result: Fast + Accurate + Safe

Automation handles 95% automatically, humans focus on 5% edge cases

In Taibles

Human-in-the-loop patterns:

Approval gates: Automation generates, human approves
Review and edit: Automation drafts, human refines
Exception handling: Automation tries, human fixes failures
Override controls: Automation decides, human can override
Selective execution: Human chooses what to run

Human-in-the-Loop Pattern

Automation stops for human review before continuing

Research

Auto

Automatic

→

Draft Email

Auto

Automatic

→

Human Review

Manual

→

Send Email

Manual

How it works:

1. System automatically generates email draft
2. You review the draft and make any edits
3. You check an "Approved" checkbox when ready
4. Send Email column only runs if approved
5. You manually click "Send" button for final safety

✅ Advantages

• Safe (human oversight)
• Can edit before sending
• Accountability tracked
• Prevents mistakes

⚠️ Considerations

• Slower (waits for human)
• Requires attention
• Need review process
• Not fully automated

🛡️ Best for: Sending emails, making purchases, or any action that can't be undone

Pattern 1: Approval Before Action

Use Case

Problem: Automation generates output that requires approval before executing critical action

Examples:

Marketing email drafted by AI → Review → Send
Lead qualifies automatically → Review → Create CRM record
Invoice data extracted → Review → Process payment
Social media post generated → Review → Publish

Key principle: Automation prepares, human approves, automation executes

Implementation Architecture

Column sequence:

Input data

Manual entry

↓

2-4

Automated processing

Run mode: Once

↓

Draft/Result

Output from automation

↓

Approval flag

Manual boolean toggle - HUMAN REVIEW

↓

Execute action

Run mode: Manual + Condition

Key components:

Automated draft column produces content
Approval toggle (checkbox) for human to check
Action column has condition + manual trigger

Example: Outbound Email Campaign

Business scenario: Generate personalized outreach emails, but require review before sending

Step 1: Create Input Columns

Column Name	Type	Purpose
Prospect Email	Text	Target recipient
Prospect Company	Text	Company name
Prospect Role	Text	Job title

Add 10-20 rows with prospect data

Step 2: Create Research Column

Column: Company Research

Company Research

LLM Column

Run mode

Once

Dependencies

Prospect Company

Prompt

Research this company and find:
- What they do
- Recent news or achievements
- Pain points they might have
- Why they might need our service

Company: [Prospect Company]

Provide 3-4 sentences of relevant information.

What this does:

Automatically researches each company when you add a new prospect. The system uses AI to find background information for personalization.

Result: Background research for personalization

Step 3: Create Email Draft Column

Column: Email Draft

Email Draft

LLM Column

Run mode

Once

Dependencies

Prospect EmailProspect CompanyProspect RoleCompany Research

Prompt

Write a personalized outbound sales email to:

Name: [Prospect Email]
Company: [Prospect Company]
Role: [Prospect Role]

Research: [Company Research]

The email should:
- Be 3-4 sentences max
- Reference specific company context
- Offer clear value proposition
- Include soft call-to-action
- Sound human and conversational (not salesy)

Subject line and body:

What this does:

Creates a personalized email draft using the research data. The AI writes naturally and includes specific company context to make each email unique.

Result: AI-generated personalized email draft

Step 4: Create Approval Column

Column: Approved for Send

Approved for Send

Boolean Column

Default value

Unchecked (false)

Manual entry

Yes

How users interact:

User reads Email Draft column

Reviews the AI-generated email

If satisfied, clicks toggle

In Approved for Send column

Toggle changes appearance

✗ Unchecked

→

✓ Checked

What this does:

This checkbox lets you approve emails one by one. Only emails with the checkbox checked can be sent. This prevents accidentally sending unreviewed emails.

Purpose: Human reviews email draft and toggles to approve

How it works:

User reads Email Draft column
If satisfied, clicks toggle in Approved for Send column
Toggle changes from ✗ (unchecked) to ✓ (checked)

Step 5: Create Send Email Column

Column: Email Sent

Email Sent

Send Email Column

Run mode

Manual

Dependencies

Email DraftApproved for Send

Condition (safety check):

Approved for Send = true

This condition ensures emails only send if the Approved for Send checkbox is checked.

Email Settings

To:[Prospect Email]

From:your_email@company.com

Subject:(extracted from Email Draft)

Body:[Email Draft]

⚠️ Important - Manual Mode:

This column will NOT run automatically. You must:

1. Select the rows you want to send
2. Right-click and choose "Run column"
3. The email will only send if approved

Important: Even after approval, this column requires manual execution

How Users Work With This

Step 1

Automation runs automatically

1. Add row with prospect data
2. Company Research executes →Done
3. Email Draft executes → Done
4. User receives notification or checks taible

Step 2

Human review

1. Click row to review
2. Read Email Draft column content
3. Decision point:
• Good email: Check Approved for Send toggle ✓
• Bad email: Leave unchecked, optionally edit draft manually
• Edit needed: Copy and modify (see Pattern 2)

Step 3

Manual execution

1. Filter view: Show only rows where Approved for Send = true
2. Select approved rows:
• Click checkbox on each approved row
• Or use "Select All" if filtered
3. Right-click selection
4. Choose "Re-run column" from menu
5. Select column: Email Sent
6. Click "Confirm"
7. Emails send to approved prospects only

Result:

✓ Emails sent only after approval
✓ Condition prevents sending to unapproved rows
✓ Manual mode prevents accidental bulk sending

Why Both Approval Toggle AND Manual Run?

Question: "If I toggle approval, why can't it auto-send?"

Answer: Two-layer safety

Approval Toggle

"This content is good"

✓Reviews individual email quality
✓Can approve multiple at once
✓Reversible (can un-toggle)

Content reviewed and approved

Manual Run

"Send it NOW"

✓Explicit action to execute
✓Conscious decision moment
✓Prevents accidental bulk execution
✓Allows batch control (send 10, wait, send next 10)

Run Column

User clicks to execute

🛡️ Combined Protection:

Both layers work together: You can approve 50 emails on Monday (Layer 1), then send them in controlled batches throughout the week (Layer 2). This prevents accidents and gives you control over timing.

Example scenario:

Monday morning: Review and approve 50 emails (toggle all)
Monday afternoon: Send batch 1 (select and run 10 rows)
Tuesday morning: Send batch 2 (select and run 10 rows)
Wednesday: Send remaining 30

Benefit: Stagger outreach, manage reply volume, test results

Pattern 2: Review and Edit

Use Case

Problem: Automation produces good drafts, but human wants to refine before using

Examples:

Blog post draft → Edit → Publish
Contract generated → Customize → Send
Response drafted → Personalize → Reply

Difference from Pattern 1: Human edits the output, not just approves

Implementation Options

Edit in Downstream Tool

Draft → Approve → Send to external tool → Edit there

Best for: Email clients, CMS

Manual Column

Draft → Copy to manual column → Edit → Use edited version

Best for: Full control

Smart Fallback

Draft → Optional override → Smart merge → Use best version

Best for: Efficiency

Option A: Edit in Downstream Tool

📝

AI Draft

Generate text

→

✓

User Approves

Review & approve

→

📧

Send to Tool

Email client, CMS

→

✏️

Edit There

Modify before send

Example:

• Email Draft generates text
• User approves
• Action copies to external tool (email client, CMS, etc.)
• User edits in that tool before final send/publish

Example:

Email Draft generates text
User approves
Action copies to external tool (email client, CMS, etc.)
User edits in that tool before final send/publish

Option B: Manual Column for Final Version

🤖

AI Draft

Auto-generated

→

📋

Manual Column

User copies & edits

→

🚀

Use Edited Version

Final action

Implementation:

Column 1: AI Draft

LLM

Run mode: Once

Purpose: Generates content automatically

Column 2: Final Content

Manual Text

Type: Text (multiline)

User action: Copy from AI Draft, then edit

Column 3: Published

Boolean

User action: Click when ready to publish

Column 4: Publish Action

Action

Dependencies: Final Content, Published

Condition: Published = true

Run mode: Manual

Uses: [Final Content] (edited version)

User workflow:

AI Draft generates content automatically
User copies content from AI Draft cell
User pastes into Final Content cell
User edits Final Content as needed
User toggles Published when satisfied
User manually runs Publish Action

Result: Human has full editing control

Option C: Smart Fallback Logic

🤖

AI Draft

→

✏️

Optional Manual Override

→

🎯

Smart Column

→

✓

Uses Best Available

Implementation:

Column 3: Final Content (Smart Fallback)

Custom Code

Logic:

If Manual Override has content:

→ Use manual version

Else if AI Draft has content:

→ Use AI version

Else:

→ No content

Benefits:

✓ Most rows use AI (fast, no manual work)
✓ Edge cases get manual override
✓ Single source of truth (Final Content)
✓ Downstream columns don't need to know source

Benefits:

✓ Most rows use AI (fast, no manual work)
✓ Edge cases get manual override
✓ Single source of truth (Final Content)
✓ Downstream columns don't need to know source

Pattern 3: Exception Handling

Use Case

Problem: Automation works 90% of the time, but failures need human intervention

Examples:

Data extraction → 90% clean, 10% needs correction
Lead scoring → 95% confident, 5% need review
Invoice processing → Most parse correctly, some need help

Principle: Automate the majority, escalate exceptions

Implementation: Validation + Conditional Review

Extract Data

↓

Validate

↓

Quality OK?

No ✗

↓

Human Review

↓

Correction

↓

Retry

Yes ✓

↓

Auto-Process

↓

Done

Example: Invoice Data Extraction

Scenario: Process invoices from various formats

Component placeholder - content will be added based on the pattern

User Workflow

Happy path (90% of invoices):

Component placeholder - content will be added based on the pattern

Exception path (10% of invoices):

Component placeholder - content will be added based on the pattern

Efficiency gain:

✓ 90% auto-processed (zero manual work)
✓ 10% exceptions flagged automatically
✓ Human focuses only on failures
✓ 10× faster than manual processing

Pattern 4: Selective Execution

Use Case

Problem: Not every row should be processed automatically

Examples:

Research data on some leads, not all
Expensive operation only for high-value items
Risky action only after review

Solution: Manual run mode + filtering

Implementation: Manual Columns with Smart Filtering

Example: Expensive Lead Research

Component placeholder - content will be added based on the pattern

User workflow:

Import 1,000 leads
Lead Score auto-calculates for all
Qualified auto-calculates for all
Filter view: Qualified = true (shows 150 leads)
Review filtered list, manually select 50 high-priority
Right-click selection
Run "Deep Research" column on selected rows only
Research executes for 50 rows ($250 total vs. $5,000 for all)

Benefit: Human judgment + automation = optimal cost/value

Common HITL Patterns Summary

Component placeholder - content will be added based on the pattern

Best Practices

Practice 1: Make Reviews Efficient

Bad: Review every single row individually

Good: Batch review with filtering

Component placeholder - content will be added based on the pattern

Efficiency: 20× faster review process

Practice 2: Provide Context for Review

Bad: Just show generated output

Good: Show input data + reasoning + output

Component placeholder - content will be added based on the pattern

Benefit: Human can make informed decisions faster

Practice 3: Learn from Corrections

Pattern: Track what humans override

Component placeholder - content will be added based on the pattern

Use aggregated data to:

Improve prompts
Adjust thresholds
Train better models
Identify edge cases

Practice 4: Set Default to Safe State

Component placeholder - content will be added based on the pattern

Balancing Automation and Control

Finding the Right Balance

Component placeholder - content will be added based on the pattern

Decision Framework

When to require human approval:

✓ Yes (add approval gate):

Actions with external consequences (emails, payments, posts)
Expensive operations (>$1 per execution)
Regulatory/compliance requirements
Brand reputation risk (public communications)
Legal implications (contracts, agreements)

✗ No (fully automate):

Data extraction and enrichment
Scoring and classification
Internal data transformation
Low-risk calculations
Read-only operations

Summary: Human-in-the-Loop Workflows

You now understand human-in-the-loop workflows:

✓ Core concept: Blend automation with human judgment

✓ Four main patterns:

Approval gates (toggle + condition + manual run)
Review and edit (draft + manual column + fallback logic)
Exception handling (validate + escalate failures)
Selective execution (manual mode + filtering)

✓ Implementation techniques:

Boolean toggles for approval
Conditional execution on approval
Manual run mode for safety
Custom code for validation
Filtering for efficient review

✓ Best practices:

Batch review workflows
Provide decision context
Learn from corrections
Default to safe state

✓ Complete example: Outbound email campaign

AI drafts personalized emails
Human reviews and approves
Manual batch execution
Two-layer safety (toggle + run)

Next Steps

Now that you understand human-in-the-loop workflows, Section 10.5 will cover:

Integration patterns with external services
Bidirectional synchronization
Multi-tool enrichment with fallbacks
Building robust integrations

Let's master external integrations!

10.4 Human-in-the-Loop Workflows ​

What Is Human-in-the-Loop? ​

The Concept ​

Machines excel at

Humans excel at

Real-World Analogy: Airport Security ​

Full automation (no humans)

Full manual (no automation)

Human-in-the-loop (hybrid)

In Taibles ​

Pattern 1: Approval Before Action ​

Use Case ​

Implementation Architecture ​

Example: Outbound Email Campaign ​

Step 1: Create Input Columns ​

Step 2: Create Research Column ​

Company Research

Step 3: Create Email Draft Column ​

Email Draft

Step 4: Create Approval Column ​

Approved for Send

Step 5: Create Send Email Column ​

Email Sent

How Users Work With This ​

Automation runs automatically

Human review

Manual execution

Why Both Approval Toggle AND Manual Run? ​

Approval Toggle

Manual Run

Pattern 2: Review and Edit ​

Use Case ​

Implementation Options ​

Edit in Downstream Tool

Manual Column

Smart Fallback

Option A: Edit in Downstream Tool ​

Option B: Manual Column for Final Version ​

Column 1: AI Draft

Column 2: Final Content

Column 3: Published

Column 4: Publish Action

Option C: Smart Fallback Logic ​

Column 3: Final Content (Smart Fallback)

Pattern 3: Exception Handling ​

Use Case ​

Implementation: Validation + Conditional Review ​

Example: Invoice Data Extraction ​

User Workflow ​

Pattern 4: Selective Execution ​

Use Case ​

Implementation: Manual Columns with Smart Filtering ​

Common HITL Patterns Summary ​

Best Practices ​

Practice 1: Make Reviews Efficient ​

Practice 2: Provide Context for Review ​

Practice 3: Learn from Corrections ​

Practice 4: Set Default to Safe State ​

Balancing Automation and Control ​

Finding the Right Balance ​

Decision Framework ​

Summary: Human-in-the-Loop Workflows ​

Next Steps ​

10.4 Human-in-the-Loop Workflows

What Is Human-in-the-Loop?

The Concept

Real-World Analogy: Airport Security

In Taibles

Pattern 1: Approval Before Action

Use Case

Implementation Architecture

Example: Outbound Email Campaign

Step 1: Create Input Columns

Step 2: Create Research Column

Step 3: Create Email Draft Column

Step 4: Create Approval Column

Step 5: Create Send Email Column

How Users Work With This

Why Both Approval Toggle AND Manual Run?

Pattern 2: Review and Edit

Use Case

Implementation Options

Option A: Edit in Downstream Tool

Option B: Manual Column for Final Version

Option C: Smart Fallback Logic

Pattern 3: Exception Handling

Use Case

Implementation: Validation + Conditional Review

Example: Invoice Data Extraction

User Workflow

Pattern 4: Selective Execution

Use Case

Implementation: Manual Columns with Smart Filtering

Common HITL Patterns Summary

Best Practices

Practice 1: Make Reviews Efficient

Practice 2: Provide Context for Review

Practice 3: Learn from Corrections

Practice 4: Set Default to Safe State

Balancing Automation and Control

Finding the Right Balance

Decision Framework

Summary: Human-in-the-Loop Workflows

Next Steps