Running an automated ad agency: what scales and what still needs humans
The promise of the automated ad agency is one operator managing a hundred accounts. The reality is closer to one operator managing twenty accounts with help from scripts that break silently. The gap between those two numbers is where structural automation — not just bid rules — makes the difference.
The real bottlenecks in agency operations
Bid management is largely solved. Most agencies run Smart Bidding or a third-party bid layer on every account. The bottlenecks are upstream: monthly client reports that take four hours each, weekly optimization reviews where someone logs into sixteen accounts and manually checks pacing, and onboarding calls that require a human to translate client goals into campaign structure.
These are language and reasoning tasks, not bid math tasks. They are exactly what AI clients do well — if they have read access to the accounts.
What an MCP server changes for an agency
With SpendSignoff connected to a client account, an operator can ask "summarize last week's performance and flag anything that needs attention" and get a structured answer across all connected platforms in one response. No dashboard login, no manual export.
For structural changes, the operator types the instruction and reviews the draft before it goes live. The client account never gets touched by an unreviewed action.
Example operator workflow
# Morning account review across 8 client accounts
# In Claude: "Check pacing for all accounts. Flag any over 105% of daily budget."
# SpendSignoff reads each account, returns ranked list of pacing alerts.
# Operator reviews, picks the ones needing action, says:
# "Draft a 15% budget cut for the Search campaign in account X."
# Draft appears in approval queue. Operator approves. Done.The agency approval model
In a white-label agency setup, the approval chain runs: AI proposes → operator reviews → client is notified (optional). The key is that the AI never touches live settings before a human has seen the change.
This matters for client trust as much as for safety. A client who sees a clean before→after changelog of every optimization decision has more confidence in the agency than one who gets a monthly PDF summary.
No model can approve on behalf of a client
mcp.approve scope does not exist for AI clients. Approvals require human credentials. For agency setups, that means the operator approves, not the model — even if the change was AI-proposed.What still requires humans
- Client strategy sessions — the AI can summarize account data, not define business goals.
- Creative direction — brand voice, visual identity, and offer framing require human judgment.
- Escalation calls — when a client's account has a major issue, a human needs to own the conversation.
- New platform strategy — the first time you run TikTok for a client, someone needs to set the strategic parameters.
Realistic scale
An agency operator using SpendSignoff for read/draft workflows can realistically handle 35–50 accounts with the same time budget a non-automated operator spends on 15–20. The multiplier comes from eliminating the dashboard-switching and report-assembly work, not from removing human judgment.
FAQ
- Can clients see the AI-proposed changes before they are approved?
- Yes. The approval queue can be configured to notify clients of pending changes. They see the before→after diff and can comment or flag before the operator approves.
- Does the white-label mode change how SpendSignoff connects?
- The MCP connection and OAuth setup are the same. White-label mode adds multi-tenant account isolation and operator-level access controls so one operator can manage multiple client accounts under one login.
- Is SpendSignoff suitable for agencies with enterprise clients?
- The KMS-signed audit log and role-based access controls are designed for enterprise compliance requirements. The append-only log gives enterprise clients the tamper-evident record most require.
Connect an account read-only and watch the operator work.
Reads are free on every plan. Nothing spends without your two-step approval.
Related reading