5 Levels of AI Marketing (And How to Master Each One)

Originally published by Shann³ (@shannholmberg) — read the original thread

I came across this thread a few weeks ago and it's been living in my head ever since. Shann breaks down something I've been thinking about for a while — the gap between people who use AI as a shortcut and people who use it as infrastructure. I wanted to share it here with a bit more context and some examples to make each level really click.

Andrej Karpathy scored 342 US jobs on a 0–10 AI exposure scale. Marketing scored a 9.

That means almost everyone in marketing is at risk — but only if they stay passive. The real danger isn't AI replacing marketers. It's marketers at level 1 being replaced by marketers at level 5.

Shann³, an AI marketing practitioner, spent the past several months building a multi-agent content system — not as a side project, but as the actual production workflow for his agency. Along the way, he mapped out five distinct levels of how people use AI for marketing.

Almost everyone is stuck at level 1 or 2. Here's the full picture.

Level 1: Custom prompts

This is where 90% of people start — and where most of them stay forever.

You open ChatGPT, type "write me an email for my product launch," and get something that sounds vaguely professional but could have been written for literally any product on earth. Every session starts from zero. No memory of your brand, no context about your audience, no knowledge of what worked last time.

Generic input produces generic output. One line of instruction produces one dimension of result.

What this looks like in practice:

A SaaS marketer needs a LinkedIn post about their new feature. They open ChatGPT, type "write a LinkedIn post about our new dashboard analytics feature," paste the output with minimal edits, and publish. It performs okay. Next week they do the same thing for a different post, starting from scratch with no connection to the last one.

Why it matters:

Level 1 is a starting point, not a strategy. It's better than nothing, and it's genuinely how most people discover that AI can help them. But the compounding cost of staying here is invisible until you see what level 3 looks like. Every piece of content is a one-off. Nothing builds. Nothing learns.

Level 2: Manual skills

Level 2 is where most "AI-savvy" marketers land. Instead of one-line prompts, you build detailed templates packed with frameworks — Schwartz awareness levels baked into email sequences, proof stacking patterns for landing pages, specific CTA structures tested across hundreds of campaigns.

A skill at this level can be 1,200+ lines of structured framework. The output quality difference vs. level 1 is night and day.

But you're still the engine. Every piece still requires your hands on the keyboard. The ceiling here is your time.

What this looks like in practice:

That same SaaS marketer now has a 60-line prompt template for LinkedIn posts. It includes the audience persona, the awareness level to write for, a hook formula, post structure, and five examples of what "good" looks like. The output is dramatically better — on-brand, structured, and ready to publish with a light edit. But they still have to pull up the template, fill in the inputs, and run it manually every time.

Why it matters:

This is where AI starts paying real dividends. You've encoded expertise into the system — your frameworks, your instincts, your standards. The problem is you're still the bottleneck. This level scales your quality but not your capacity. You still need to show up for every single piece.

Level 3: Skills + brand foundation

This is where most people should be right now. Almost nobody is.

You create a brand foundation file: your voice, your tone, words you always use, words you never use, your audience's pain points, your positioning, and the phrases that sound like you versus the ones that sound like AI slop. Then every skill references this foundation automatically. Every output passes through your brand filter before it reaches you.

One foundation feeds all your skills — email, social, landing page, ad copy — all pulling from the same source of truth.

At level 2, you might produce a great email that doesn't sound like you. At level 3, everything sounds like you, because the system knows who "you" is. Your AI content starts scaring competitors because it doesn't read as AI anymore.

What this looks like in practice:

The marketer now has a brand foundation document that defines their company voice ("direct and a little irreverent, never corporate"), banned phrases ("leverage synergies," "best-in-class"), their ICP's core fears and motivations, and five real examples of content that nailed the tone. Every prompt template references this doc. A LinkedIn post, a cold email, and a product landing page all come out sounding unmistakably like the same brand — without manually re-explaining the brand each time.

Why it matters:

This is the consistency layer that most teams spend thousands of dollars and hours trying to enforce manually. Brand voice reviews, editorial guidelines no one follows, revision cycles that balloon because "it doesn't sound like us." Level 3 bakes that judgment into the system once and applies it automatically to everything after. It's also the level where AI content stops being detectable as AI.

Level 4: Agents with skills

Most marketers have never touched this level.

An agent reads your goal, plans the steps, selects the right skills, and executes. One command triggers a full workflow. Want to research the top 20 tweets about a topic this week, generate 5 content ideas based on what performed, and draft the best one? The research agent pulls the data. The idea agent structures the concepts. The writer agent drafts it. A critic agent scores it before you ever see it.

The key concept here is subagents — isolated specialists that run in parallel. The coordination cost drops to zero. No briefing, no waiting for deliverables, no managing handoffs. One person doing the work of a small team.

What this looks like in practice:

A content team wants a weekly newsletter. At level 4, one command kicks off a workflow: a research agent scans the top industry news and competitor content from the last 7 days, an analysis agent identifies the 3 most relevant angles for their audience, a writer agent drafts the newsletter using the brand foundation and a newsletter-specific skill, and a critic agent flags anything off-brand before it hits the human editor's inbox. The editor reviews and approves. Total human time: 20 minutes instead of 3 hours.

Why it matters:

This is where the leverage becomes real. You stop doing the work and start reviewing it. The agent does the research, the drafting, the quality check. You bring the judgment call at the end. You're not the writer anymore — you're the editor.

Level 5: Autonomous agent teams

Very few people are here yet.

The difference between level 4 and 5 is coordination. At level 4, subagents work in isolation. At level 5, agents share context, build on each other's outputs, and compound their knowledge over time. The research agent's findings feed directly into the writer's context. The critic's feedback loops back into the next draft cycle.

And then there's memory.

Run 1 produces a brand voice document and a first draft. Run 5 adds performance data and tighter headlines based on what actually got clicks. Run 20 has audience feedback baked in, personalized patterns for different segments, and a track record of what converted versus what bounced. The system gets better every time it runs.

Shann knows of an 8-person team currently running 38 agents across 8 departments — content, SEO, paid ads, email, social, analytics, CRM, and reporting — each with their own agent cluster sharing context through a central knowledge base.

At this level, AI marketing isn't a tool you open. It's an OS you run.

What this looks like in practice:

That same content team, six months into their level 5 system, now has agents that know which subject lines got the highest open rates, which content formats drove the most signups, which audience segments respond to humor vs. data-heavy posts. The newsletter agent doesn't just draft from a template — it drafts from a living performance record. Each edition is shaped by everything that came before it. The system compounds.

Why it matters:

This is the moat. Speed is easy to copy. A competitor can hire more people or buy more tools. But six months of accumulated performance data, audience memory, and baked-in judgment? That takes six months to build, minimum. The teams that start now will have a structural advantage that latecomers simply can't shortcut.

Why most people won't do this

AI-generated ads show a 19% improvement in clickthrough rates. But purchase likelihood drops 33% when consumers know AI created the content. "AI slop" mentions went from 461K to 2.4M in a single year — a 5x increase in people calling out bad AI content.

The gap between levels 1–2 and levels 3–5 is taste, judgment, and technical skill. People at the lower levels are producing volume. People at the higher levels are producing volume that doesn't read as AI, because they baked their judgment into the system.

Think of Rick Rubin — he doesn't play instruments on most albums he produces. But every album sounds intentional because his taste shapes every decision. That's the model. You're not the one writing anymore. You're the one deciding what stays, what gets cut, and what gets refined.

The moat is taste. Anyone can set up agents. The hard part is having judgment worth adding to them.

How to move from one level to the next

1 → 2 (an afternoon): Write 3 detailed prompt templates for your most common tasks. Include frameworks, constraints, and examples of good output. Go from a 1-line prompt to a 50-line skill document.

2 → 3 (a weekend): Write your brand foundation. Voice, tone, audience, positioning, words you never say. Then make every skill reference it.

3 → 4 (1–2 weeks): Pick one workflow you repeat and automate it end-to-end. Identify the steps, build a skill for each one, then connect them through an orchestrating agent.

4 → 5 (ongoing): Add memory. Log what performs and what doesn't. Feed that data back. Build a shared context layer all your agents can read from.

The window

Right now, operating at level 3 or above puts you in a small minority. Most agencies are still at level 1, maybe level 2. The ones building their systems today will have compounding advantages that late adopters simply can't catch up to.

Because the advantage isn't just speed — it's accumulated taste, judgment, and months of memory that a competitor starting from scratch doesn't have.

Every week you wait is a week of compounding you don't get back.

This post is based on a thread by Shann³ (@shannholmberg). Read the original thread here.