Practices / Agency

Agency AI Coordination Playbook

Coordination practices for agencies running AI agent teams that manage client advertising, call centers, project delivery, and sales pipelines. Battle-tested patterns from a 25-agent production deployment.

4 practices 10 categories

Data

Measured

Alert on Deviation, Not Absolutes

Pacing alerts should fire when spend deviates from budget by a percentage threshold (e.g., +30% overspend), not when spend crosses a dollar amount. A $100 overspend on a $500/month account is a crisis. A $100 overspend on a $50,000/month account is noise.

What goes wrong without this

Every small account triggers alerts constantly because the dollar thresholds are set for large accounts. Alert fatigue sets in. When a real overspend happens on a large account, nobody notices because alerts are ignored.

Observed

Cross-Reference Before Alerting

Before an agent fires an alert about a client issue, it must check at least 2 other data sources to confirm. The ad pacing agent checks project status (is the account offboarding?) and the exclusion list (is this account paused?) before alerting on overspend.

What goes wrong without this

The pacing agent fires an overspend alert for a client that is offboarding next week. The team wastes 30 minutes investigating a non-issue. This happens 3 times a week. Trust in alerts erodes.

Measured

Median Over Average for Client Metrics

When reporting aggregate client metrics (cost per lead, appointment rates, etc.), use median instead of average. One outlier client spending 10x the others will skew the average and hide real performance patterns.

What goes wrong without this

The portfolio "average CPL" is $45, which looks fine. But the median is $28, and one client at $180 CPL is dragging the average up. The expensive client's problem is invisible in the aggregate.

Rule

Single Source of Truth for Client Data

One agent owns each data domain. The performance analyst owns all ad performance data (Meta + Google). The project manager owns all project status data (Accelo). No two agents independently query the same data source for the same purpose.

What goes wrong without this

The sales agent and the retention agent both pull client spend data independently. They use different date ranges. Their reports contradict each other. The founder does not know which number is right.

Stay in the loop

Get weekly coordination intelligence updates. No account required.