Most of what's been written about AI in marketing is either uselessly abstract ("AI will revolutionize the funnel!") or uselessly tactical ("here's the ChatGPT prompt that wrote our hero copy"). Neither helps you operate.
We've shipped 40+ production AI workflows across our portfolio in the last 12 months. The ones that survived past month two share a common architecture pattern. The ones that died died for the same reasons. This is what we learned.
The pattern that works: bounded agent, narrow workflow, human checkpoint. An AI workflow earns its keep when three conditions are true. First, the scope is narrow — one job, well-defined inputs, well-defined outputs. "Write our content" fails this. "Summarize the inbound lead's company from their domain and tag intent signals" passes. Second, the agent is bounded by tools and policies — it can call Clearbit, it can call our CRM, it cannot send external email without human approval. Third, every output has a checkpoint — a human reviews before the artifact ships externally. The agents that get switched off in month two are the ones that hallucinated a customer's name in an outbound email and the founder found out from a screenshot in Slack.
The architecture we actually deploy. Three layers. Layer one is a workflow orchestrator — n8n, Temporal, or in larger orgs, a custom service. Layer two is the agent layer, typically Claude (Anthropic) for reasoning-heavy tasks and GPT-4 for tool-use heavy ones. Layer three is the tool surface — CRM API, ESP API, data warehouse, vector store. The agent doesn't talk to the world directly; it talks through the tool layer. This sounds bureaucratic until you've watched an agent confidently send 400 cold emails to the wrong segment.
Three workflows that have paid for themselves 10x.
Lead enrichment + scoring on form fill. When a lead submits a form, an n8n workflow fires. Clearbit enrichment runs in parallel with a Claude call that scores the lead on 23 signals — company stage, recent funding, public hiring activity, web tech stack. Total runtime: 6 seconds. The lead lands in HubSpot scored 0-100 with reasoning attached. Sales no longer wastes a discovery call on a $50K-revenue prospect who said they had a $5M budget on the form.
Content research agent. Before our writers draft a piece, an agent assembles a research brief — top 20 ranking competitor pages, the entities they cover, the questions they don't answer, three customer-side anecdotes pulled from our Gong transcripts. The writer starts at draft 1.5 instead of draft 0. Cycle time on long-form articles dropped 38%, quality went up.
Reply triage on the founder's inbox. Every reply to an outbound sequence runs through a triage agent. Three buckets — buying signal, polite no, autoreply. Buying signals get drafted-but-not-sent reply variants. The founder reviews and ships. Inbox triage time went from 90 minutes a day to 12.
The workflows that died. Anything that tried to be 'the AI marketer.' Anything where the agent was the final decision-maker. Anything where the output was directly customer-facing without a checkpoint. The pattern: as scope expanded, error rates compounded. AI in marketing in 2026 isn't a replacement for marketers — it's a 3-5x productivity multiplier for marketers who already know what good looks like.
What to build first. Pick the most repetitive 30-minute task your senior marketer does every day. Build the agent for that. Ship it in two weeks. Measure the time saved. Then build the second one. The teams that try to ship an "AI marketing platform" die. The teams that ship 12 small, boring, useful agents end up with operating leverage their competitors can't match.
