Governance playbook
Decide who can approve an ad change — before the AI proposes one.
The fastest way to lose a client account isn’t a bad change. It’s a change nobody signed off on. When AI is drafting moves across a dozen client accounts, you need a per-client answer to “who approves this, and where’s the record” — not a shared login and good intentions.
For agencies and freelancers managing multiple client ad accounts with AI in the loop.
Per-client approval matrix
A starting RACI you can adapt per client before you let an AI draft anything. It answers the two questions that actually stall agency work: who’s allowed to push this change, and is it written down. Set your own dollar thresholds; the example rows show the shape.
| Change type | Who can draft | Who must approve | Logged? |
|---|---|---|---|
| Budget shift under 10% | AI operator / strategist | Account lead | Always |
| Budget shift ≥ 10% or over your $ cap | AI operator | Account lead + client contact | Always |
| New campaign or creative | Strategist | Account lead + client | Always |
| Pause / major status change | AI operator | Account lead | Always |
| Bidding strategy change | AI operator | Account lead | Always |
- Decision rule: set the dollar threshold per client. A safe default — the client approves anything that changes total spend or pauses delivery; the account lead approves the rest.
- Edge case the matrix has to cover: an automated platform change (Smart Bidding, an Advantage+ shift) isn’t “approved” by anyone. It can’t go through the gate, so it must at least be logged and surfaced — silent automation is the gap a matrix alone misses.
- This is an operating template, not a contract. Map it to each client’s actual sign-off terms.
You started with two clients and a shared login. Now you have twelve.
Two clients and a spreadsheet of logins worked fine. Then you added ten accounts, three strategists, and an AI drafting changes across all of them — and now, for any given account, nobody can say who is actually allowed to push a budget change live.
So it goes one of two bad ways. Either everything funnels through you, and you’re the bottleneck on twelve accounts. Or changes go out that the client never saw, and you find out when they do.
- “Who actually needs to approve this?” is named one of the biggest agency pain points — the ambiguity itself causes confusion and delays.
- “Without clearly defined approval hierarchies, even the most advanced automation collapses under confusion and delays.”
- Only about 6% of companies fully trust AI agents to run core processes, and trust drops to around 20% when money is at stake — which is exactly why a human approval gate, not autonomy, is the design clients accept.
How a governed approval gate works
Per-client grants, not a shared login
Each client account gets its own grant: who can view, who can draft, who can approve. A strategist drafting on Client A can’t push live on Client B. The shared-login era ends here.
Every AI draft hits that client’s queue
The AI operator proposes; the change waits in the client’s approval queue. The approver your matrix names — not whoever happens to be logged in — sees the before→after and approves or rejects.
The approval is the record
Approve, and the change goes live with a signed entry: who approved which version, and when. That record is the same artifact a client asks for and an audit wants — governance and accountability are the same write.
Shared login vs. governed approval
Why this is a real category, not a feature
Approval governance is a named 2026 agency category, not invented demand.
2026 agency-stack writing explicitly bundles multi-client permissioning and role-based, multi-level approval, and names “who actually needs to approve this?” as a top pain point whose ambiguity causes confusion and delays.
octopusmarketing — governance & approval 2026Best practice calls for a record of who approved which version, when.
The same 2026 governance writing states an approval workflow should produce “a clear record of who approved which version and when” — which means the approval gate and the audit record are the same artifact, not two systems.
adlibrary — 2026 agency stackTrust collapses when AI touches money — so a human gate is the buyable design.
Only about 6% of companies fully trust AI agents to run core processes, and trust falls to roughly 20% for financial transactions. Selling “autonomous spend” fights that wall; an approval gate sells into it.
HBR — only 6% fully trust AI agentsQuestions
Why not just give the AI guardrails and let it run?
Because the people who pay you don’t trust it yet — only about 6% of companies fully trust AI agents to run core processes, and that drops to ~20% once money is involved. An approval gate lets you use AI without asking a client to accept autonomous spend, which most won’t.
How is this different from a shared client login?
A shared login means anyone with it can change anything, and nobody can prove who did. Per-client grants split view, draft, and approve, so a strategist on one account can’t push live on another, and every approval is attributable.
Where’s the record of who approved a change?
In the approval itself. When a change is approved and pushed, it’s written as a signed entry — who approved which version, and when. That’s the same record you’d hand a client or pull for an audit, so governance and reporting share one source.
Can different clients have different approval rules?
Yes — that’s the point of a per-client matrix. One client may want to sign off on any budget change; another may only care above a dollar threshold. The grants and thresholds are set per account, not globally.
Does the AI ever push a change without approval?
No. The AI drafts; a human approves with the two-step control before anything spends. The one thing a gate can’t catch is the platform’s own automation — so those changes are logged and surfaced rather than approved.
Keep reading
Put an approval gate on every client account.
Bring your client accounts to a 15-minute demo and see per-client grants, an approval queue, and the signed record that comes with each approved change — AI drafts, a human approves.