Most of what gets written about AI in marketing is theater. Someone pastes a prompt into a chat window, gets a passable blog post, and calls it transformation. That isn't a system. It's a parlor trick that doesn't survive contact with a real backlog.
When I talk about building marketing systems with AI tools, I mean infrastructure: a repeatable pipeline where each stage has a defined input, a defined output, a tool doing the work, and a human at the gates that actually matter. I'm Sebastian Behar, and over the years I helped scale a company through a period of major revenue growth and its IPO. Almost none of that came from one-off prompts. It came from systems, and what follows is one I actually built and still run.
The system: a content and lifecycle production pipeline
The problem was boring and expensive. We had more content and lifecycle work queued than the team could ship, and most of the bottleneck wasn't the writing itself. It was the connective work around writing: pulling the brief, matching it to data, drafting variants, getting it into the CMS, wiring it to the right audience. That's where the days went.
So I built a pipeline that treats AI as the labor between the decisions rather than the decision-maker. Five stages, each one a clean handoff to the next:
- Intelligence — pull the performance signals and audience context that tell us what to build and for whom.
- Brief — turn that signal into a structured brief a writer or a model can act on without guessing.
- Draft — generate first-pass copy and variants against the brief, not against a blank page.
- Production — push approved copy into the CMS and lifecycle tooling as structured, deployable assets.
- Activation & feedback — ship to the audience, measure, and feed the results back into stage one.
Stage 1 and 2: intelligence and the brief
This is where the system earns its keep, and where most people skip straight to drafting. Garbage in stays garbage no matter how good the model is, so I'd rather spend the time here than anywhere else.
I pull behavioral data from Amplitude and session context from FullStory to see where intent is concentrated and where pages leak. On the enterprise side, Demandbase and 6Sense tell me which accounts are in-market. That signal feeds a brief, and I use Claude to compile it: the audience, the job-to-be-done, the proof points, the constraints, the angle. The model is good at synthesis here precisely because I've handed it structured inputs instead of a vague ask.
The human gate is the brief itself, and I read every one before it moves. A wrong brief multiplies downstream in a way a wrong sentence never does, so that's the place to be careful.
Stage 3: drafting against a brief, not a blank page
Drafting is where AI looks most impressive and, honestly, matters least. A model generating copy from a tight brief is genuinely fast, but what it gives you is a strong first draft and a set of variants, not a finished asset.
I run drafts through Claude, and for anything that touches code or templating I use Codex. The variants matter more to me than the polish, because the next thing I'm going to do is test them. The pipeline doesn't replace judgment about what's good; it produces enough credible options that a human can pick and sharpen instead of starting cold.
Stage 4: production into the real stack
This is the stage that separates a demo from a system. Copy that lives in a chat window is worthless; copy that's a structured entry in your CMS, mapped to a campaign and an audience, is an asset you can actually deploy.
Approved content moves into Contentful as structured entries, then into the lifecycle layer: Braze for consumer messaging, Marketo and Salesforce for the enterprise motion. Personalization runs through Mutiny and experiment variants through Optimizely. The AI step here is mechanical and high-leverage, shaping copy into the exact fields and formats each system expects, which is precisely the tedious work that used to eat a marketer's afternoon.
None of this is exotic. It's the same tooling in the stack I run every day, and the pipeline just connects the pieces that used to require a human to ferry data between them.
Stage 5: activation, feedback, and the loop that makes it a system
A pipeline that only runs forward is a conveyor belt. What makes this a system is the loop. This is the growth loops, not funnels idea that Reforge has done so much to popularize: durable growth comes from compounding loops where each cycle's output becomes the next cycle's input, not from a linear funnel you keep refilling at the top. Once content ships, performance flows back into Looker and Amplitude, and those results become the intelligence for the next cycle. Winning variants inform the next brief, and the losing ones tell the model what to stop generating.
This is also where the discipline of measurement keeps the whole thing honest, because the variant testing baked into stage three is what turns directional gut calls into evidence. That same habit sits behind the wins I've shipped over the years: meaningful conversion-rate lifts, real incremental revenue, and a 7× ABM conversion improvement from a build earlier in my career. Systems compound in a way that one-off campaigns simply don't.
I go deeper on the tooling choices in my full marketing tech stack, and on the testing discipline in how I run experiments with AI.
What to take from this
The takeaway isn't the specific tools. It's the shape of the thing: define your stages, assign a tool to each, and put humans at the gates where being wrong is expensive, which for me is the brief and the facts. AI is the labor between your decisions, not the decisions themselves.
Build the pipeline once and it pays you back every cycle, and that, more than any single tool, is what separates using AI from building with it.
