Amelia Lerutte built an AI marketing agent in 15 minutes on stage at SaaStr AI 2026. Five months later, SaaStr runs nearly 30 agents used almost a million times. The playbook she shared is worth dissecting — not because it's magic, but because it exposes exactly where human judgment still matters and where it doesn't.
The agent, called 10K, reportedly handles about 60% of a traditional VP of Marketing's functions: campaign management, email drafting, dashboard synthesis, win-back workflows. It doesn't manage people. That 60/40 split tells you everything about where AI sits in marketing leadership right now.
Why This Matters for Your Budget Conversation
More than two-thirds of marketing leaders feel pressure from CEOs and CFOs to deliver cost savings and efficiency gains. But here's the trade-off nobody's spelling out: if you position AI purely as a cost play, you invite budget shrinkage. Research suggests teams that use AI for growth (not just speed) see roughly 2× higher marketing-driven profitability. The framing you choose in your next board deck has real consequences.
Meanwhile, 82% of SaaS companies have already invested in AI for their products. The question isn't whether to adopt. The question is whether marketing leads the implementation or gets dragged along by the CTO's roadmap. Right now, the CTO/CIO leads AI strategy in 44% of organizations. The CMO leads it in only 32%.
That gap is the real problem this playbook solves.
The 10-Step Build, Stripped to What Matters
Lerutte's process started from a mundane pain: she was manually copying marketing dashboards into Notion every week. The agent grew from there. Here's the condensed sequence, with the operational details that actually determine success or failure.
Step 1: Pick one number. Not three. One. Paid attendees, pipeline dollars, qualified signups — whatever your team's single north-star metric is. The spec flows from this choice. Lerutte's point: detailed specs produce better agent performance. Vague briefs produce vague outputs. No surprise there.
Step 2: Collect your data. Every spreadsheet, every historical report. This is where most teams stall. Key barriers to AI efforts cited by practitioners include data quality and technology integration with legacy systems. If your CRM data is a mess, your agent will confidently report garbage. Fix the foundation first.
Step 3: Build v1 fast. Lerutte used Replit (a vibe coding platform) to get a working version quickly. The goal isn't perfection. The goal is a functional prototype you can pressure-test against real data within days, not quarters.
Step 4: Connect Salesforce first. Read/write access to pipeline and revenue data. This is the integration that makes the agent useful rather than decorative. Everything else is secondary until the agent can see your actual numbers.
Steps 5–6: Add APIs and workflows incrementally. Marketing automation, Slack, Google Calendar — one at a time. Build workflows for specific tasks: daily idea generation, win-back campaigns, newsletter drafts. Trying to wire everything at once is the fastest way to kill the project.
Step 7: Define autonomy levels. This is the step most teams skip, and it's the one that creates the most risk. Which tasks can the agent execute without approval? Which require a human sign-off? Lerutte's framework uses two layers: an autonomous layer (dashboards, scheduled jobs, drafted emails) and an operator layer (manual analysis, outbound tasks that need judgment).
Steps 8–10: Guard against hallucination, build memory, verify everything. Implement data accuracy checks before any communication goes out. Maintain a single memory file with institutional rules and corrections. And manually verify initial outputs before trusting the system with anything customer-facing.
The Operator Layer Is the Actual Moat
The autonomous layer gets the attention. The operator layer is where the value compounds. Each interaction with the agent builds a library of reusable scripts and knowledge. Over time, the agent gets better because the human operating it leaves a trail of corrections, preferences, and context that no off-the-shelf tool can replicate.
This matches what experts have been saying: AI adoption in marketing is a workflow shift, not a business model shift. Human judgment remains critical for managing brand risk and converting AI outputs into actions that actually land with buyers. The 61% of B2B SaaS companies already using AI for lead scoring and qualification are seeing results (teams report 2.8× higher demo-to-paying conversion rates), but those results come from humans tuning the system, not from the system running unsupervised.
What to Measure (and What Not to Over-Interpret)
Success = reduction in manual hours per workflow, plus improvement in the north-star metric you picked in Step 1. Guardrails = hallucination rate on outbound communications, data accuracy vs. source-of-truth CRM. Stop-loss = if the agent's outputs require more correction time than the manual process it replaced, pause and retrain.
Don't treat early wins as proof of full autonomy. The 60% coverage claim is directional, not definitive — it depends heavily on your org's complexity, data quality, and how many systems the agent can access.
Lerutte's original problem was a weekly copy-paste job. Five months later, the agent drafts campaigns, scores pipeline, and sends reminders. The distance between those two states wasn't covered by a single breakthrough. It was covered by a thousand small corrections logged in a memory file, each one making the next interaction slightly less wrong. That's the actual playbook: not the 15-minute demo, but the five months of teaching that followed it.
