Direct answer: build a recommendation copilot before an autonomous operator. It reads spend, forecast, constraints, change history, and mature outcomes; produces an evidence-backed proposal; validates the mutation without applying it; requests approval inside a band; executes one change; confirms the effect; and rolls back when a guardrail fails. If the company cannot distinguish scarce demand, restrictive targets, broken signals, and weak commercial quality, pacing automation only accelerates bad decisions.
The investment case emerges when a team reconciles many accounts, campaigns, currencies, and sources every day, rules are stable, delay has material cost, and history supports comparisons among action, inaction, and mature outcomes. A small account with few monthly changes benefits more from discipline. Ten markets and hundreds of campaigns may justify integration—provided value means revenue or margin, not recovered spend.
Pass the PACE gate before funding
- P — Proof: diagnosis is repeatable and the company knows when not to spend.
- A — Addressability: campaign, budget, currency, offer, owner, and change history have stable IDs.
- C — Consequence: accepted lead, opportunity, mature revenue, margin, or another outcome reconciles at a known lag.
- E — Execution control: action and period caps, approval, validate-only, idempotency, expiry, kill switch, and tested rollback exist.
When one item fails, fund that foundation first. The diagnostic in https://makinai.co/insights/en/paid-media-campaign-not-spending-budget-ai-diagnosis defines constraint classes. API write capability is not corporate authority, and native recommendations do not carry company margin, sales capacity, or mature revenue by themselves.
Choose one of five authority levels
- 0. Observe — consolidate data and detect anomalies; make no recommendation.
- 1. Recommend — propose action, evidence, impact, and rollback; a human decides.
- 2. Prepare — build the mutation and run technical validation; a human applies it.
- 3. Execute with approval — apply within an approved band and window.
- 4. Execute by policy — act without case approval only for reversible, low-value cases under versioned policy.
Start at level 1. Advance only after false positives, reversals, and commercial quality remain inside agreed limits for several cycles. Do not confuse a human-reviewed policy with blanket permission to move money.
The minimum production architecture
- Read layer — cost, budget, status, bid, target, forecast, inventory, creative, conversions, and change history.
- Canonical model — normalized IDs, time zone, currency, window, objective, and lag.
- Decision ledger — prior snapshot, hypothesis, evidence, proposal, approver, mutation, outcome, and reversal.
- Policy engine — allowed actions, limits, cooldown, separation of duties, and exceptions.
- Executor — validate-only, idempotency key, post-write confirmation, and partial-failure handling.
- Observability — freshness, volume, drift, calls, errors, external changes, and kill switch.
- Outcome layer — accepted leads, pipeline, revenue, margin, payback, and capacity from CRM or commerce.
Google exposes change history for up to 30 days and supports validation and atomic mutations by default, so persist your own ledger before writing. Meta and LinkedIn expose different surfaces; design connectors by capability rather than pretending every platform shares the same field or latency.
A six-week implementation
- Week 1 — one account, one offer, and 10–20 campaigns; freeze economics, limits, and prohibited actions.
- Week 2 — connect platform reads, change history, and CRM; reconcile IDs, currency, time zone, and lag.
- Week 3 — shadow mode; compare recommendations with actual decisions without writing.
- Week 4 — validate-only and mutation preparation; test idempotency, failures, and rollback.
- Week 5 — execute with approval in one low-value cell under a daily cap.
- Week 6 — compare time, error, reversal, quality, pipeline, and hours saved; keep, expand, or retire.
Cost, accountability, and proof
Cost rises with platforms, accounts, API access, granularity, refresh, history, currencies, identity, offline conversions, CRM depth, security, environments, on-call, and audit. Marketing owns economics and limits; media owns actions; data owns contracts; RevOps validates outcomes; engineering operates connectors; security reviews credentials; finance validates exposure; an independent approver authorizes expansion. Measure freshness, constraint precision, acceptance, decision time, human hours, failed or reversed mutations, pacing, preserved budget, accepted leads, opportunities, mature revenue, margin, payback, and impact against a comparator.
Use https://makinai.co/insights/en/ai-cross-platform-budget-reallocation-marginal-return to evaluate the next budget unit, https://makinai.co/insights/en/ai-crm-revenue-feedback-loop-ad-platforms to close the revenue loop, and https://makinai.co/insights/en/ai-marketing-agent-action-contract-guardrails to specify the action contract. If your team repeats this ritual in spreadsheets, bring MAKINAI at https://makinai.co/services/en/digital-marketing-media-performance-growth-agency one account, 10–20 campaigns, current rules, change history, and an anonymized CRM sample. We can scope authority and the pilot without replacing your stack.
Limitation: APIs, fields, permissions, and versions change. Platform estimates are not causal evidence; human approval does not eliminate risk; spending closer to plan does not prove growth. Preserve credential separation, your own history, and a defensible comparator.