Simbel AI — Your AI Marketing Team

Part 1 of a 4-part series on building Simbel AI.

I've spent 20 years working in data systems and AI infrastructure — the kind of work that rarely shows up in a product demo but makes everything else possible. Pipelines, data models, systems that move information reliably from one place to another. Not glamorous, but I understood it deeply.

Then I started paying attention to how marketing teams actually operate.

It was humbling. Not because the people weren't smart — they were. But the tooling they were working with was a disaster. Not a subtle, nuanced disaster. An obvious, daily, everyone-knows-it's-broken disaster.

The Stack of Duct Tape

A marketing manager at a mid-size company described her morning to me once. She checks analytics in three separate platforms. Logs into Google Docs to review the strategy someone wrote two weeks ago. Pastes content prompts into a general-purpose AI chatbot one at a time. Manually schedules posts across four platforms. Downloads reports from each platform, then consolidates them into a spreadsheet. Repeats tomorrow.

I asked how many tools were involved. She counted on her fingers and got to nine before stopping.

The "integration" between these tools wasn't a technical integration. It was her — sitting in front of too many browser tabs, doing the connective tissue work manually. She was the pipeline. A human being was serving as the ETL layer for her own marketing operation.

This isn't unusual. This is how most marketing teams work.

The fragmentation isn't accidental — it's the cumulative result of a decade of specialized tools being built in isolation. Each one solves its narrow piece well. But nobody designed the system. Nobody asked what it costs for a human to be the glue holding it all together.

When I looked at this through an engineering lens, the answer seemed obvious: this is an automation problem. A data pipeline problem. A multi-agent orchestration problem. It was the kind of problem I'd spent 20 years building solutions to — just applied to a domain I hadn't worked in before.

So I started building.

The First Attempt Was Terrible

Honesty is important here, because the tech industry has a habit of skipping the embarrassing middle section.

My first version was just an LLM wrapper. You gave it a topic and a brand name, and it generated social media captions. It technically worked. The output was coherent. You could copy it and paste it somewhere.

I showed it to some marketing friends. Their feedback was diplomatically delivered but clear: it saved them maybe 20 minutes. The other eight hours of their day were unchanged. The strategy had to come from somewhere. The research had to come from somewhere. Platform-specific thinking — what works on LinkedIn is very different from what works on Instagram — wasn't there. The content had no memory of what had been posted before, no awareness of the competitive landscape, no consideration of timing.

I had built a caption machine, not a marketing system.

The lesson was obvious once I saw it: a single AI model, no matter how capable, cannot replace a marketing function. It can replace one task within that function. That's a completely different thing.

You Don't Need One AI — You Need 37

The breakthrough came when I stopped thinking about AI tools and started thinking about how marketing agencies actually work.

A good agency doesn't have one generalist who does everything. They have a strategist who researches the market and builds the plan. A copywriter who understands the brand voice and knows what resonates on each platform. A designer who thinks visually. A media buyer who manages paid campaigns. An analyst who tracks what's working and feeds that back into the next cycle.

Each specialist has a defined role, a clear scope, and they hand off their output to the next person in the chain. The quality of the final product depends on all of them doing their specific job well — not on one person trying to do all of them at once.

This is the correct mental model for AI in marketing. Not one model doing everything. A team of specialized agents, each with a focused role, working together in a coordinated pipeline.

That's what I built. 37 agents organized across 7 functions.

The Seven Functions

Let me walk through them at a high level — I'll go much deeper on each in future articles.

F1: Research & Strategy — 7 agents

This is the homework before any content gets written. These agents analyze your competitive landscape, map your target audience, identify trends, benchmark best practices for each platform, and synthesize it all into a marketing strategy document. This document becomes the foundation that every downstream agent uses as context.

Nobody should be writing content before this work is done. Most tools skip it entirely. We don't.

F2: Content Creation — up to 10 agents

These agents generate platform-specific content based on the strategy. Notice "platform-specific" — not one piece of content reformatted for multiple channels. Instagram captions, X threads, LinkedIn articles, email campaigns — each format requires different thinking, different length, different tone, different structure. An agent for each.

The content creation layer is also where our Arabic dialect support lives, but I'll cover that in part 3 of this series.

F3: Auto-Publishing — 6 agents

Scheduling isn't trivial. These agents determine optimal posting times based on audience data, manage the actual API connections to each platform, handle retry logic when things fail (and they do fail), and confirm delivery. Somebody has to own the end-to-end delivery problem. These agents do.

F4: Analytics & Optimization — 6 agents

Performance tracking, A/B test analysis, ROI calculation, audience psychographic profiling, trend analysis. But more importantly: the learnings from every campaign get stored in a knowledge base that future campaigns can retrieve. The system gets smarter over time. This is the part that makes the compounding return possible.

F5: Paid Ads — 3 agents

Google Ads, Meta Ads, LinkedIn Ads. Strategy, creative generation, performance optimization. Most marketing automation tools treat paid and organic as separate worlds. We don't. The same research that drives organic content informs the ad strategy.

F6: SEO Audit — 3 agents

Technical SEO, content SEO, and AI search visibility scoring. If you're spending money generating content and no one can find it, that's a problem. These agents audit your site, identify prioritized fixes, and score how visible you are in AI-driven search results — an increasingly important surface that most SEO tools haven't caught up to yet.

F7: Cold Outreach — 2 agents

Email finding and verification, personalized multi-step sequences, subject line variants. Not glamorous, but it closes the loop on acquisition. The outreach context is informed by the same research that drives everything else — so the personalization is genuine, not a mail-merge with someone's first name.

The Hard Problem: Making 37 Agents Work Together

Building the individual agents was the easy part. The hard part was making them work together reliably.

A few things I learned the painful way:

Not every agent needs the same brain. Early on I used the most capable available model for everything. This was slow and expensive. A sophisticated language model reasoning through your competitive strategy is well worth premium compute time. The same model generating hashtag suggestions is massive overkill. We now tier agent intelligence based on task complexity, and the system runs much more efficiently as a result.

Sequential and parallel execution both matter. Some agents need to run in order — you can't write content before research is complete. Others can run in parallel — if you're creating content for six platforms, there's no reason to create them one at a time. Getting this right required thinking carefully about the dependency graph. Getting it wrong is expensive.

Things fail. Always. API calls time out. Rate limits get hit. A third-party data source returns garbage. The naive assumption when building any automated system is that each step succeeds. The production reality is that failures happen constantly, and the question is whether your system degrades gracefully or collapses. We built an explicit state machine for every campaign that can recover from failures at any stage without restarting from zero.

Context is everything. Each agent needs to know what happened before it. Within a function, each agent receives the output of all previous agents as context. Across functions, the strategy document persists as a shared reference. Across campaigns, the RAG knowledge base accumulates learnings that inform future runs. Without this context architecture, you don't have a team — you have 37 individuals who don't know what the others are doing.

The Arabic Question

I want to mention this briefly here and return to it in depth in part 3.

The reason I built Simbel specifically — not just another generic marketing automation tool — is Arabic.

400 million Arabic speakers are actively underserved by marketing technology. The problem isn't just that existing tools don't support Arabic. It's that they don't understand the fundamental reality that Arabic isn't one thing. Modern Standard Arabic — the formal written language — is not how people talk. It's not how people post on Instagram or write captions for TikTok. Every region has its own dialect, its own idioms, its own cultural references.

Saudi content needs to sound Saudi. Egyptian content needs to sound Egyptian. Marketing that doesn't account for this sounds like a news anchor trying to sell you something at a restaurant. It works technically; it doesn't work culturally.

I grew up switching between dialects depending on who I was talking to. This isn't an academic problem for me — it's something I've lived. Building the dialect system was personal in a way that the agent architecture wasn't.

Why I'm Writing This

I'm writing this series because I think transparency about what we're building is useful — both for people who might use Simbel and for people building similar systems who might learn something from the choices we made and the mistakes we made along the way.

The marketing automation space is full of demos that make things look simpler than they are. I'd rather tell you what this actually took to build, including the parts that didn't work the first time.

Part 2 goes into the technical architecture in detail — what happens, step by step, from the moment you click "Launch Campaign" to the moment your posts are live.

Bassem Zohdy is the founder of Simbel AI, an AI-powered marketing automation platform built for Arabic-speaking markets.

Why I Built an AI Marketing Team of 37 Agents