A framework for evaluating bounded, supervised workflows across Salla, Zid, Shopify, Meta, TikTok and WhatsApp—without giving a model unchecked authority over customers, spend, prices or orders.
This is educational content under source review—not a production runbook. Platform APIs, policies and product behavior change. Re-check current Salla, Zid, Shopify, Meta, TikTok and WhatsApp documentation; pin supported versions; and involve qualified engineering, privacy, security, accessibility and legal reviewers.
What completion means: you can draft a proposal-only architecture and pilot contract. Use synthetic fixtures, OAuth with least privilege, signed and deduplicated webhooks, strict tenant isolation, schema validation, representative and adversarial evaluations, human approval, logs, spend and rate caps, stop conditions, a kill switch and tested rollback. It does not establish growth, savings, compliance or production readiness.
Illustrative scenario, not a customer case: a Gulf retailer has fragmented store, media, fulfillment and support data. The first useful experiment is not seven autonomous agents. It is one read-only brief that reconciles definitions, cites sources, flags uncertainty and helps a named operator make one decision. Compare it with the current process before adding any tool or write permission.
These are proposed service boundaries, not autonomous employees. Pilot one role at a time and preserve human ownership of every material decision.
You have a defined operational problem, an accountable owner, permission to use the relevant data, and capacity to run a controlled pilot.
You need local research across language, payment, fulfillment, accessibility, customer expectations and applicable law—not a borrowed regional average.
You want better evidence and review workflows while keeping media decisions and spend accountable to people.
You want to ship an agentic system but need the GCC-specific context — payment norms, language mix, cultural calendar. All included.
You can enforce client-by-client isolation, authorization, approvals, contracts, incident response and deletion.
You're building internal AI capability. This is the architecture blueprint — vendor-neutral, swappable, owned by you.
This is technical. API schemas, prompt engineering, evaluation matrices. If you want introductory AI literacy, start with THE AI SHIFT instead.
Document the customer problem, current process and data rights first. Architecture cannot substitute for product-market fit or operational readiness.
Before writing a single prompt, you need to understand the Gulf market at operating depth and design the data spine that all seven agents will read from. Skip this phase and everything downstream breaks.
Research payment behavior, order value, active periods, language, Ramadan and WhatsApp by country, category and store. Regional averages are hypotheses.
How agents coordinate. Event-driven vs polling. LLM orchestration. Shared state. Where to use Claude vs GPT vs Gemini. The whole system on one diagram.
Salla, Zid, Shopify APIs. Meta & TikTok Ads APIs. WhatsApp Cloud API. The canonical data schema every agent reads from. Supabase/Postgres setup.
Four parts, seven agents. Each with full technical specs: inputs, outputs, prompt templates, evaluation metrics, failure modes, and the exact orchestration pattern that keeps them from stepping on each other.
Use official public tools and permitted pages, respect terms and rate limits, retain provenance, and never treat ad visibility or longevity as performance evidence.
Arabic-first prompting. Brand voice encoding. Cultural calendar. Static + video + carousel formats. Variant generation. Review queue. Shipping workflow.
Data-quality and anomaly checks, alternative explanations, draft budget changes, explicit approval, idempotency, spend caps and a decision log.
Consent and suppression-aware eligibility, destination-specific data rules, minimized audiences, and cited Arabic or English briefs.
The final part turns the support role into an approval-led pilot and replaces a fixed deployment timeline with evidence gates, shadow mode, canary release, monitoring and rollback.
"Indeed, Allah will not change the condition of a people until they change what is in themselves." — Qur'an 13:11
Begin with one decision and its accountable owner. Establish a baseline, write the data and permission boundary, assemble representative normal and adversarial cases, and pre-commit the stop rule. The first successful outcome is trustworthy evidence—not automation or growth. Start with Part 1.