Direct answer: AI should not move the next dollar to the channel with the highest average ROAS. It should compare the expected marginal gain from a small spend increase after outcomes, cohort maturity, uncertainty, and business constraints use the same definitions. Start in shadow mode, protect an experimentation reserve, require approval for material changes, and record how every move can be reversed. Platforms can keep optimizing inside their environments; the company must own the decision across them.
Google may recommend more spend, Meta may surface another opportunity, and retail media may claim a sale also visible in analytics. Each recommendation can be internally coherent while competing for the same money. A common layer does not replace platform algorithms. It establishes the economic denominator, hard limits, and acceptable uncertainty.
Use the Next-Dollar Decision Contract
- Common outcome — realized revenue, contribution margin, accepted opportunity, or another reconciled result; never rank clicks, leads, and sales together.
- Eligible envelope — movable budget, floors, ceilings, commitments, geography, inventory, and capacity.
- Maturity clock — minimum cohort age and data delay before comparison.
- Marginal curve — expected response to the next spend block, uncertainty range, and model version.
- Experiment reserve — protected money for learning, including channels the model may underestimate.
- Operating gates — margin, stock, frequency, sales capacity, brand safety, consent, and SLA.
- Approval and reversal — owner, maximum change, duration, alerts, and rollback condition.
- Evidence log — recommendation, inputs, human decision, execution, outcome, and correction.
The contract prevents three shortcuts: moving money because a dashboard is green, applying an old curve to a new market, and letting an agent optimize an objective finance does not recognize. Every decision names the denominator, cohort, increment, risk, and owner.
Average return is not marginal return
A channel may show high average ROAS because it harvested branded demand while its current spend is already saturated. Another may have lower historical ROAS but a better response to the next block. Robyn describes marginal response as the “next dollar” response at the current spend level and models carryover and saturation. Meridian optimizes constrained budget scenarios. Both are instruments: without calibration, adequate data, and constraints, a curve can be precise-looking rather than reliable.
Keep three values separate: observed outcome, attributed outcome, and estimated incremental outcome. The first comes from CRM, commerce, or finance; the second helps operate channels; the third requires a model or experiment and carries uncertainty. Never silently replace one with another.
Respect what platforms already solve
Performance Planner forecasts changes inside Google Ads and includes conversion delay for eligible scenarios. Meta can share part of spend across eligible ad sets. These capabilities are useful, but they do not observe the entire portfolio under a shared economic outcome. Treat native recommendations as decision inputs, not competing orders.
Climb the reallocation ladder
- 1. Repair cost, identity, conversion, and value data.
- 2. Optimize within each channel and document objectives and limits.
- 3. Simulate cross-channel moves without execution.
- 4. Compare recommendations with current decisions in shadow mode.
- 5. Release small reversible changes spaced by maturity.
- 6. Scale only after repetition, experimental calibration, and stable operations.
Do not jump from a report to an agent with write access. First produce a reviewable recommendation. Later, automate only moves inside pre-approved bands and retain a kill switch.
Run a six-week pilot
- Week 1 — choose two or three channels, one outcome, and a movable envelope; reconcile four weeks of cost and outcomes.
- Week 2 — define maturity, floors, ceilings, experiment reserve, and margin, stock, and capacity gates.
- Week 3 — estimate curves and uncertainty; compare simple rules, MMM, and native signals without hiding disagreement.
- Week 4 — generate shadow recommendations and log what would move, why, and under what rollback rule.
- Week 5 — execute one or two small moves with a comparator, decision window, and explicit approval.
- Week 6 — assess mature outcomes, volatility, error, operating load, and the decision to keep, correct, or stop.
Deliverables include an outcome dictionary, cost table, maturity rules, versioned curves, constraints, scenario simulator, approval queue, execution integration, logs, exception dashboard, and runbook. Cost grows with channel count, granularity, history, identity, markets, decision frequency, experiments, integrations, review, and operations.
Measure decision quality, not only ROAS
Track accepted recommendations, reversals, gate violations, forecast error, cohort stability, incremental lift when available, margin, payback, operating cost, and the value of the experiment reserve. Benchmark against a simple policy such as holding spend or rebalancing by a fixed rule. If AI does not beat the rule after cost and risk, keep the rule.
Read the platform/company boundary at https://makinai.co/insights/en/paid-media-ai-beyond-google-meta-automation, align maturation at https://makinai.co/insights/en/attribution-window-revenue-maturation-ai-marketing-budget, connect margin and payback at https://makinai.co/insights/en/ai-customer-acquisition-margin-payback-value, and preserve experiment capacity at https://makinai.co/insights/en/ai-marketing-capacity-valid-experiment-operating-system. If budget is still decided by spreadsheet and dashboard conflict, bring eight weeks of cost, outcome, and constraints for two channels to MAKINAI at https://makinai.co/services/en/digital-marketing-media-performance-growth-agency. We can scope the shadow layer before discussing automated execution.
The sources document capabilities and methods; they do not prove that a specific move will create value. Curves can confuse correlation, carryover, and selection, while experiments may be infeasible or inconclusive. The contract and pilot are editorial guidance to validate in the real workflow.