Aug 14, 2026

Write-Access Matrix for AI on Meta and Google

A write-access matrix is a permission framework that assigns AI automation levels to specific ad account functions based on risk tolerance, spend volume, and channel type. Tiers range from always-human (AI recommends, human executes) to propose-only (AI executes with human review window) to kill-switch (AI executes, human can stop).

Three Access Tiers Defined

Always-human: AI generates recommendations, analysis, or creative variants. Humans approve before any account change. No API write access. Used for high-risk decisions or low-frequency actions.

Propose-only: AI executes changes with a mandatory review window (typically 2-4 hours). Changes are live but can be reverted by human operator. Used for medium-risk, repeatable optimizations.

Kill-switch: AI executes immediately. Human monitors and can pause campaigns or disable AI automation without approval delay. Used for low-risk, high-frequency actions with clear rollback paths.

Channel-Specific Assignments

Meta (Facebook/Instagram) typically permits higher automation due to platform speed and audience targeting depth. Google Search requires tighter control because bid changes directly affect cost-per-click and daily spend variance is higher.

Meta campaign budget allocation: Propose-only (48-hour review window)
Meta audience expansion or lookalike seed changes: Always-human
Google Search bid adjustments (device, location, time): Kill-switch (daily spend variance < 15%)
Google Search campaign pause/resume: Propose-only (4-hour window)
Google Search new keyword additions: Always-human
Creative rotation or ad copy swaps (both platforms): Propose-only (24-hour window)

Spend Threshold Decision Rule

Access tier should tighten as daily spend increases. A brand spending $500/day can tolerate more automation risk than one spending $5,000/day. Use this threshold ladder:

< $1,000/day total: Propose-only for budget and bid changes; kill-switch for creative rotation
$1,000 - $5,000/day: Always-human for budget allocation; propose-only for bids and creative
$5,000 - $20,000/day: Always-human for budget, bids, and audience changes; propose-only for creative only
> $20,000/day: Always-human for all write access except pre-approved creative swaps (kill-switch with daily spend cap)

Account Structure and Segmentation

Separate accounts or campaigns by risk profile. High-performing, stable campaigns can run at kill-switch tier. New campaigns, experimental audiences, or seasonal products should start at always-human and graduate only after 2-4 weeks of stable performance.

Evergreen core campaigns (> 90 days, stable ROAS): Propose-only or kill-switch
New product launches (< 30 days): Always-human
Seasonal or promotional campaigns: Propose-only (24-hour review)
Test campaigns (< $500/day budget): Kill-switch with daily spend cap
Holdout or control groups: Always-human (no AI writes)

Implementation Checklist

Before assigning any write access, audit account structure, define review windows, and document rollback procedures.

Map all campaigns and ad sets to tier (always-human, propose-only, kill-switch)
Set daily spend caps per campaign; cap should be 1.5x average daily spend
Define review window duration (2-4 hours for propose-only; document in Slack or email alert)
Assign human reviewer(s) for each tier; ensure coverage during business hours
Create rollback SOP: who can pause AI, how to revert changes, escalation path
Log all AI writes to a shared audit sheet (timestamp, change type, reviewer, outcome)
Test with a pilot campaign at each tier before full rollout
Review matrix quarterly or after major account changes (new products, budget shifts, team changes)

Common Pitfalls

Overestimating human review capacity. A single operator cannot review 50+ propose-only changes per day. Reduce scope or move to kill-switch with tighter spend caps.

Ignoring platform API latency. Meta and Google can take 15-30 minutes to reflect changes. A 2-hour review window may be too short if changes are already live.

Mixing tiers within a single campaign. Assign one tier per campaign; tier changes require explicit decision and documentation.

No rollback plan. Kill-switch access without a clear pause procedure creates liability. Always define who can stop AI and how.

Monitoring and Escalation

Set up daily alerts for anomalies: spend variance > 20%, ROAS drop > 15%, or CTR change > 25%. Route to human reviewer immediately. If AI is propose-only or kill-switch, pause automation and investigate before resuming.

Daily spend variance alert: If actual spend vs. forecast > 20%, pause AI budget changes
Performance regression: If ROAS or conversion rate drops > 15% vs. 7-day average, escalate to always-human tier
Frequency cap breach: If AI increases audience size or bid without approval, trigger manual review
Quarterly audit: Pull all AI-executed changes; spot-check 10% for correctness and alignment with strategy

Questions

FAQ

Can a campaign move between tiers?

Yes. Campaigns should graduate from always-human to propose-only after 2-4 weeks of stable performance (ROAS within 10% of target, no anomalies). Demotion to always-human should happen immediately if performance drops > 15% or spend variance exceeds threshold. Document all tier changes in a shared log.

What if the human reviewer is unavailable during the review window?

Propose-only changes should not execute if no reviewer is available. Set up a backup reviewer or extend the review window to the next business day. For kill-switch access, define a maximum time the AI can run unmonitored (e.g., 4 hours); pause automation outside business hours.

Should creative changes (ad copy, images) be treated differently than budget or bid changes?

Yes. Creative changes have lower financial risk and faster rollback time. A brand can move creative to kill-switch or propose-only (24-hour window) while keeping budget and bid changes at always-human or propose-only (4-hour window). Test creative variants in a small segment first.

How do I know if my spend threshold is correct?

Start conservative: use the threshold ladder as a baseline, then run a 2-week pilot at each tier. Measure human review time, approval rate, and performance impact. If human reviewers are overwhelmed or AI changes are frequently rejected, tighten the tier (move toward always-human). If performance is stable and approval rate > 90%, consider loosening (move toward kill-switch).

Want this on your account?

Thirty minutes. Bring the number that keeps you up.

Book a demo