Part 1: The Big Picture
Part 1 of 7 · Reviewed 31 Aug 2026

The Big Picture

Where models can assist media buying, where deterministic controls must take over, and what evidence is required before any production use.

⏱ Self-paced
🎓 Beginner Level
Educational and Safety Boundary

This framework does not promise revenue, ROAS, productivity, career outcomes, or safe autonomous operation. Platform behavior, API versions, prices, permissions, and policies change. Begin with synthetic fixtures or authorized, minimized, redacted data. Production use requires qualified privacy, security, legal, finance, analytics, and advertising-platform review appropriate to the account.

Section 01

AI Assistance Is a Workflow, Not a Job Title

A model can draft and challenge. It does not own the objective, evidence, account, or decision.

Generative tools can help structure a brief, normalize approved data, surface anomalies, create alternative hypotheses, and draft a change proposal. Fluency is not evidence, and a model output is untrusted input until people and deterministic controls verify it.

The safe default in this course is read-only analysis and proposal-only output. The model has no platform write credential, never sees a client secret or access token, and cannot publish, send, pause, launch, change a budget, or alter a customer or product record.

Three Labels on Every Output

FACT
A dated claim tied to an authorized source
Include tenant, account, entity, metric definition, attribution setting, time zone, date range, currency, source ID, capture time, and known lag where relevant.
INFERENCE
A ranked explanation, not a discovered cause
Show assumptions, plausible alternatives, uncertainty, evidence against the explanation, and a validation step. An anomaly or correlation does not establish causality.
PROPOSAL
A reversible option awaiting a named reviewer
State the exact entity, before/after values, rationale, dependency, cap, expiry, measurement plan, stop condition, and rollback. It remains inert until separately approved.
A More Useful Mindset

Do not aim to “replace clicks.” Aim to improve the evidence presented to an accountable operator while reducing unauthorized data exposure and preventing accidental side effects.

Section 02

The Controlled Architecture

Collection, reasoning, approval, and execution are separate trust boundaries.

read → reason → approve → execute
1. READ SERVICE tenant-scoped identity · allowlisted fields · rate limits
2. MODEL redacted evidence · no credentials · typed proposal only
3. REVIEW QUEUE full immutable diff · sources · risks · exact approval
4. EXECUTOR separate write identity · server binding · schemas · policy · caps
5. CONTROL stable idempotency · audit · monitoring · stop · rollback
// The model cannot cross these boundaries by asking itself for permission.

What Each Boundary Must Enforce

  • Read service: a distinct read-only service identity, tenant and account allowlists, minimum authorized fields, explicit date ranges, pagination, rate-limit handling, source snapshots, and retention limits
  • Model boundary: redaction before inference, no secrets or credentials, no direct network or write tools, schema-constrained proposals, and “insufficient evidence” when the approved pack cannot answer
  • Approval service: server-authenticated reviewer and tenant binding, authorization checks, complete immutable before/after diff, sources, policy result, expiry, and an approval bound to the exact proposal hash
  • Deterministic executor: a separate narrow write identity; platform, tenant, account, entity, action, and destination allowlists; type and range validation; spend and count caps; PAUSED-only creation; and no model-generated retry logic
  • Operations: a stable idempotency key derived from approved intent, tamper-aware audit events, reconciliation, monitoring, rate-aware backoff, kill switch, credential revocation, incident ownership, and tested rollback
Approval Is Not a Chat Reply

“Looks good” in a conversation is not authorization. Approval must be verified server-side and bound to the authenticated person, tenant, account, entities, exact immutable diff, caps, and expiry. Any changed value invalidates it.

Section 03

Use the Minimum Evidence Needed

More data is not automatically better, safer, more lawful, or more useful.

Start from the decision and request only the authorized fields needed to evaluate it. Spend, impressions, clicks, platform-reported conversions, configured attribution, inventory state, and margin inputs each have different owners and definitions. Join them only when the purpose, keys, rights, quality, and retention have been approved.

📊
Structured API Data
Validate schema, units, currency, time zone, attribution window, freshness, pagination, account binding, and missing rows. Structured does not mean complete or correct.
📄
CSV Exports
Use an authorized export, remove unnecessary identifiers and free text, scan for formulas, record provenance, encrypt storage, limit access, and delete on schedule.
📷
Screenshots
Crop and redact names, emails, IDs, audiences, URLs, notifications, and account details. Record consent or authority, purpose, destination, access, retention, and deletion.
🧪
Synthetic Fixtures
Use invented tenant, account, entity, metric, and failure data for learning, schema tests, prompt-injection tests, duplicate handling, partial failure, and rollback drills.

Data Collection Gate

  • Document purpose, controller or owner, lawful basis or contractual authority, fields, source, recipients, retention, deletion, and prohibited reuse
  • Keep each client’s credentials, storage, logs, prompts, evals, screenshots, CSVs, reports, and indexes isolated by tenant and purpose
  • Do not pool client data, infer sensitive traits, upload customer-level records by default, or use one tenant to benchmark another without a specific approved arrangement
  • Stop collection on authorization failure, unexpected fields, tenant mismatch, pagination error, excessive response, stale definitions, or rate-limit exhaustion
Section 04

Diagnose With Hypotheses, Not Stories

Performance movement rarely has one observable cause.

A change in CPA, ROAS, conversion count, or delivery can reflect auction conditions, attribution lag, budget edits, creative mix, inventory, price, tracking, consent, site availability, product mix, or ordinary variation. A model should rank plausible explanations and propose validation—not announce a root cause.

FACTS
- Metric definitions, source IDs, time zone, currency, date range, lag
- Account and entity IDs; before/after values; data-quality warnings

RANKED HYPOTHESES
- Supporting evidence, conflicting evidence, alternatives, uncertainty
- Validation query or controlled test for each hypothesis

PROPOSAL
- Reversible option only; exact diff, cap, owner, approval, expiry
- Primary measure, guardrails, stop condition, incident route, rollback
Causal Language Gate

Use “associated with,” “consistent with,” or “hypothesis” unless a valid experiment or causal design supports a stronger claim. Platform-reported attribution is not, by itself, proof of incremental impact.

Section 05

Earn Readiness With Dated Evaluations

Choose models and workflows for a defined task, account context, language mix, and risk level.

Do not select a model from a leaderboard, brand preference, or one impressive prompt. Create a versioned, task-specific evaluation set with synthetic or authorized redacted cases. Record the model, configuration, prompt, retrieval pack, policy, schema, date, reviewers, thresholds, disagreements, and limitations.

  • Representative: relevant account structures, currencies, time zones, Arabic and English content, missing values, attribution lag, and low-volume cases
  • Adversarial: prompt injection in names and documents, cross-tenant requests, malicious URLs, secret requests, approval spoofing, and unsafe write instructions
  • Regression: rerun after any model, prompt, retrieval, schema, policy, API version, permission, or executor change
  • Operational: pagination, rate limits, duplicate delivery, partial failure, stale data, reconciliation, stop, credential revocation, and rollback
Release Ladder

Move from synthetic fixtures to an isolated test account, then shadow analysis, then a smallest capped canary only after the relevant gates pass. Preserve human approval. Expand, pause, revise, or retire based on evidence—not elapsed time.

Section 06

Minimum Readiness Checklist

No single prompt, skill, connector, or agent satisfies this checklist by itself.

  • Approved purpose, users, data classes, rights, retention, provider terms, tenant boundary, and named owners
  • Minimum-field read identity separated from the narrow write executor; credentials stay in a managed secret store and never reach the model
  • Typed input and output schemas, policy enforcement, full immutable diff, exact server-side approval, caps, expiry, and stable idempotency
  • Representative, bilingual, adversarial, privacy, safety, and regression evaluations meet task-specific thresholds
  • Audit, reconciliation, rate-limit behavior, monitoring, kill switch, incident response, credential revocation, and rollback are tested
Current References

Verify implementation details in the current Meta Marketing API documentation, Shopify authentication guidance, NIST AI Risk Management Framework, and OWASP guidance for LLM applications. Official documentation is a source for reviewed code, not permission for a model to rewrite a live integration.

Next: See the Bounded Use Cases

Part 2 turns this architecture into scenario designs with evidence, permissions, review, stops, and rollback made explicit.

Built by @itsmazinzaki — AVAMARTECH