How to move ad budget safely with an AI agent
Move budget from losers to winners in steps a learning phase can absorb, with every change previewed and confirmed before it touches the account.
IntentReallocate with discipline: right sizes, right order, previews first, and a rollback path.
- 5
- Steps
- 8
- Tools
- 3
- Rules
- 3
- Failure modes
When to use it
Use it when
- After a performance review names winners and losers.
- Seasonal shifts planned in advance.
- Spend caps forcing prioritisation.
Do not use it when
- Mid-learning-phase on the winners: feeding a campaign that has not stabilised measures nothing.
- Reacting to fewer than 7 days of data, where attribution lag owns the signal.
What it needs first
Required
Optional
Preconditions
- A source ranking of winners and losers with evidence, typically from cross-platform-performance or weekly-performance-review.
- Write access on the plan, and the user present to confirm.
Inputs
movesrequired- The intended shifts, or the ranking to derive them from.
step_cap- Maximum change per campaign per step.Default: 20% of current daily budget.
Procedure
In order. Every tool named is one the gateway ships, and links to its reference.
Read current budgets and statuses for every campaign involved.
Compute target budgets under the step cap, in account currency, and show the full plan before any call.
Preview each change with Safe Writes: the tool returns the exact diff and applies nothing.
- google_ads_update_campaign_budget
- meta_update_campaign_budget
- meta_update_adset_budget
- tiktok_update_campaign_budget
- pinterest_update_campaign_budget
First call without confirm: the response is the preview.
After human confirmation, apply, re-read the budgets, and record before/after.
Schedule the check-in: results judged only after the conversion lag has passed.
Decision rules
The observable condition, what it lets you conclude, and what takes the conclusion back.
When
A proposed increase exceeds the step cap
Conclude
Split it across steps days apart; large jumps reset learning and burn the measurement.
When
The loser funds brand or remarketing terms that protect other campaigns
Conclude
Flag the dependency before cutting; savings there can reappear as losses elsewhere.
When
The winner's efficiency was measured during a promo period
Conclude
Discount the ranking and say so; post-promo efficiency is the real baseline.
Evidence every conclusion must carry
- The plan shows every campaign's current budget, target, step and date.
- After applying, the re-read budgets are quoted as confirmation, not assumed.
Where the agent stops
- Every change previews and waits for explicit human confirmation: that is Safe Writes, not a courtesy.
- The skill never originates a move the source ranking does not support.
How it goes wrong quietly
The cases where the analysis is wrong and still looks right. Read them before trusting a number.
- Shared budgets on Google make per-campaign moves illusory; read the budget object, not the campaign, first.
- Meta ad set budgets vs campaign budgets (CBO) change where the lever is; confirm which mode each campaign runs.
- Cutting a loser that feeds the winner's retargeting pool moves the loss downstream a week later.
What the answer contains
- 01The plan
- Every move: from, to, step, date, evidence link.
- 02Previews
- The gateway's own diffs, verbatim.
- 03Confirmation record
- Applied changes with before/after budgets re-read.