Safety Model

Overview

Meta Ads Copilot is read-only by default. It can pull any report, analyze any data, and generate recommendations without asking. But it will never take action without explicit approval. This safety model protects you from unintended spend changes while enabling fast, autonomous reporting.

The Core Principle

Always ask before acting. Any operation that affects ad delivery or spend requires your explicit confirmation:

Pausing or resuming ads, adsets, or campaigns
Changing budgets (daily or lifetime)
Modifying bids or optimization goals
Creating or deleting campaigns

Reports, insights, and recommendations run automatically. Actions wait for approval.

Read-Only Operations (No Approval Needed)

The copilot can run these autonomously:

Pull insights and reports

social marketing insights campaign CAMPAIGN_ID --preset last_7d
social marketing campaigns --status ACTIVE
social marketing ads --adset-id ADSET_ID

All reporting commands are safe to run automatically.

Analyze performance data

Calculate efficiency metrics (CPA, ROAS, CPC)
Identify trends (CTR declining, frequency climbing)
Compare against benchmarks
Flag bleeders and winners

Generate recommendations

“Pause this bleeder — CTR 0.4%, spend $127”
“Shift $100/day from Retargeting to Product Launch”
“Resume this winner — previously 3.2% CTR, $18 CPA”

Log findings to learnings.md

Append insights and patterns to workspace/brand/learnings.md for future context.

Reports and analysis are always safe. The copilot can explore data freely without risk.

Approval-Required Operations

These actions require explicit user confirmation:

Pause/Resume Actions

# REQUIRES APPROVAL
social marketing pause ad AD_ID
social marketing resume campaign CAMPAIGN_ID

Why: Pausing affects spend immediately. Resuming restarts delivery. Approval flow:

Agent: "Pause ad 'Summer Sale' (ID: 120212345678)? This will stop delivery immediately."
You: "yes"
Agent: ✓ Paused. Logged to learnings.md.

Budget Changes

# REQUIRES APPROVAL
social marketing set-budget adset ADSET_ID --daily-budget 10000

Why: Budget changes affect spend rate and delivery. Approval flow:

Agent: "Increase 'Product Launch' budget from $450/day to $550/day?"
You: "yes"
Agent: ✓ Budget updated. Logged to learnings.md.

Campaign/Ad Creation

# REQUIRES APPROVAL
social marketing create campaign --name "New Campaign" --objective CONVERSIONS

Why: New campaigns create new spend. Approval flow:

Agent: "Create new campaign 'Q1 Product Launch' with $200/day budget?"
You: "yes"
Agent: ✓ Campaign created (ID: 120212345690). Logged to learnings.md.

Approval Workflow

Here’s how the always-ask-before-acting principle works in practice:

Agent identifies action opportunity

From reports or user request, the copilot determines an action is needed.Example:

Found bleeder: "Flash Sale Banner" — $89 spend, 0.6% CTR, frequency 4.1

Agent presents recommendation with data

Recommendation: Pause "Flash Sale Banner"
Reason: CTR 0.6% (target: >1%), frequency 4.1 (max: 3.5), $89 spent with 0 conversions
Impact: Frees $89/day budget for reallocation

Agent requests explicit approval

Pause ad "Flash Sale Banner" (ID: 120212345680)?
This will stop delivery immediately.

Type 'yes' to proceed, or 'no' to skip.

User approves or rejects

You respond:

"yes" or "proceed" or "do it" → Action executes
"no" or "skip" or "not now" → Action skipped
Ask for more info → Agent provides additional context

Agent executes and logs

On approval:

social marketing pause ad 120212345680

Then logs to workspace/brand/learnings.md:

## 2026-03-04 — Pause Action
- Paused "Flash Sale Banner" (ID: 120212345680)
- Reason: CTR 0.6%, frequency 4.1, $89 spend, 0 conversions
- User approval: Confirmed via daily check

Batch Approvals

When multiple actions are recommended, you can approve all at once:

Agent:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🩸 BLEEDERS (3 found)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. "Flash Sale Banner" — $89 spend, 0.6% CTR
2. "Testimonial Carousel" — $124 spend, 0.3% CTR
3. "Holiday Promo" — $67 spend, 0.5% CTR

Recommendation: Pause all 3 bleeders.
Total budget freed: $280/day

Proceed with pausing all 3?

You: "yes"

Agent:
✓ Paused "Flash Sale Banner" (ID: 120212345680)
✓ Paused "Testimonial Carousel" (ID: 120212345681)
✓ Paused "Holiday Promo" (ID: 120212345682)

All actions logged to learnings.md.

Audit Trail

Every action is logged to workspace/brand/learnings.md with:

Timestamp — When the action occurred
What changed — Ad ID, name, old/new values
Why — Data justification (metrics, thresholds)
Who approved — User confirmation method

Example Log Entry

## 2026-03-04 — Ad Management Actions

### Paused Ads

**Ad:** "Flash Sale Banner" (ID: 120212345680)
- **Reason:** CTR 0.6% (target: >1%), frequency 4.1 (max: 3.5), $89 spend, 0 conversions
- **Impact:** Frees $89/day budget
- **Approved by:** User confirmation via daily check (10:23 AM)
- **Command:** `social marketing pause ad 120212345680`

**Ad:** "Testimonial Carousel" (ID: 120212345681)
- **Reason:** CTR 0.3%, frequency 3.8, $124 spend, 1 conversion at $124 CPA (target: $25)
- **Impact:** Frees $124/day budget
- **Approved by:** User confirmation via daily check (10:23 AM)
- **Command:** `social marketing pause ad 120212345681`

### Budget Changes

**Campaign:** "Product Launch" (ID: 120212345683)
- **Old budget:** $450/day
- **New budget:** $550/day
- **Reason:** CPA $18.20 (28% below target), ROAS 4.2x (40% above target), scaling winner
- **Approved by:** User confirmation after budget optimization analysis (11:05 AM)
- **Command:** `social marketing set-budget campaign 120212345683 --daily-budget 55000`

This audit trail provides full transparency and accountability for every action taken.

Preventing Accidents

Confirmation Requirements

The copilot recognizes these as approval:

"yes"
"proceed"
"do it"
"confirm"
"go ahead"

Any other response is treated as “no” or “need more info”.

Double-Check for High-Risk Actions

For particularly risky actions (e.g., pausing an entire campaign), the copilot may ask twice:

Agent: "Pause campaign 'Product Launch' (3 adsets, 12 active ads)?
        This will stop ALL delivery in this campaign.

        Are you sure? Type 'yes' to confirm."

Undo Warning

Most actions cannot be automatically undone. Pausing is reversible (resume), but budget changes take effect immediately and affect historical pacing.

If you accidentally approve an action:

Pause/Resume: Immediately reverse by resuming/pausing
Budget change: Set budget back to previous value (check learnings.md for old value)
Campaign creation: Pause the new campaign immediately

Customizing Safety Settings

You can configure approval requirements in ad-config.json:

{
  "safety": {
    "require_approval_for_pause": true,
    "require_approval_for_resume": true,
    "require_approval_for_budget_changes": true,
    "require_approval_for_campaign_creation": true,
    "double_confirm_campaign_pause": true,
    "auto_log_to_learnings": true
  }
}

Recommendation: Keep all approval gates enabled. The copilot is designed to make recommendations fast — approval takes 2 seconds and prevents costly mistakes.

Daily Workflow with Safety Model

Here’s how the safety model enables efficient, safe ad management:

Morning Routine (Automated)

Cron triggers daily check at 9:00 AM
Copilot pulls insights, identifies bleeders/winners (read-only, no approval needed)
Copilot sends summary to Telegram/Slack with recommendations
You read over coffee (2 minutes)

Action Phase (Approval Required)

You reply: "Pause the bleeders"
Copilot confirms: "Pause 3 ads? yes/no"
You reply: "yes"
Copilot executes, logs to learnings.md
You get confirmation: "✓ 3 ads paused. $280/day budget freed."

Reallocation (Approval Required)

You ask: "Shift that budget to the winners"
Copilot calculates, recommends: "Increase Product Launch budget by $280/day?"
You reply: "yes"
Copilot executes, logs to learnings.md
You get confirmation: "✓ Budget updated. New daily budget: $730."

Total time: 5 minutes. No Ads Manager clicking. Full audit trail. Safe, explicit approvals.

Next Steps

Pause and Resume — Stop bleeders, restart winners with approval
Budget Optimization — Shift budget safely with data-backed recommendations

Get Started

Core Concepts

Skills

Configuration

Reports

Actions

Advanced

Overview

The Core Principle

Read-Only Operations (No Approval Needed)

Approval-Required Operations

Pause/Resume Actions

Budget Changes

Campaign/Ad Creation

Approval Workflow

Batch Approvals

Audit Trail

Example Log Entry

Preventing Accidents

Confirmation Requirements

Double-Check for High-Risk Actions

Undo Warning

Customizing Safety Settings

Daily Workflow with Safety Model

Morning Routine (Automated)

Action Phase (Approval Required)

Reallocation (Approval Required)

Next Steps

​Overview

​The Core Principle

​Read-Only Operations (No Approval Needed)

​Approval-Required Operations

​Pause/Resume Actions

​Budget Changes

​Campaign/Ad Creation

​Approval Workflow

​Batch Approvals

​Audit Trail

​Example Log Entry

​Preventing Accidents

​Confirmation Requirements

​Double-Check for High-Risk Actions

​Undo Warning

​Customizing Safety Settings

​Daily Workflow with Safety Model

​Morning Routine (Automated)

​Action Phase (Approval Required)

​Reallocation (Approval Required)

​Next Steps

Overview

The Core Principle

Read-Only Operations (No Approval Needed)

Approval-Required Operations

Pause/Resume Actions

Budget Changes

Campaign/Ad Creation

Approval Workflow

Batch Approvals

Audit Trail

Example Log Entry

Preventing Accidents

Confirmation Requirements

Double-Check for High-Risk Actions

Undo Warning

Customizing Safety Settings

Daily Workflow with Safety Model

Morning Routine (Automated)

Action Phase (Approval Required)

Reallocation (Approval Required)

Next Steps