AI Agents are too good at marketing now

Summary

The episode presents a marketing-engineering approach: use AI agents to perform much of the research, creative production, editing, and campaign “middle work,” while humans build the systems, supply context, set quality standards, and supervise decisions. The core idea is not simply to prompt a model, but to equip it with business data, approved platform access, specialized tools, and a growing library of marketing knowledge and creative examples.

The speakers also position advertising as a way for early-stage companies to test whether a product resonates with customers—not just as a channel for scaling an established business.

Marketing-Engineering Playbook

Build the Agent’s Operating Environment

• Business data: Collect data from sources such as advertising platforms, analytics, CRM, and payment systems.
• Data pipeline and warehouse: The proposed setup uses Airbyte to move data into ClickHouse, so an agent can query consolidated, frequently synced data instead of repeatedly pulling from source APIs.
• Tool access: Provide APIs for tasks such as image generation, video, transcription, lead research, and language models.
• Platform access: Give the agent the required permissions to create or manage ads and publish content.
• Human oversight: Add approval steps where a person should review work before it goes live.

Separate Reading Data from Taking Action

The agent should read performance data from the warehouse and use advertising-platform APIs primarily for actions such as uploading creative or pausing and moving ads. This reduces repeated, high-volume requests to platform APIs.

Schneider warns that agents can make excessive requests and violate API terms or trigger account problems if access is configured carelessly.

Create a Repeatable Static-Ad Workflow

1. Provide a brand URL and a static-ad skill.
2. Have the agent extract brand details and search for existing ads.
3. Research customer pain points and desired outcomes using sources such as Reddit and social media.
4. Generate multiple ad concepts and headlines.
5. Create the visual assets, then review and refine them.
6. Upload approved ads to the advertising account.

The demonstration generated concepts for products including Idea Browser, Late Checkout, Cursor, and other brands. Schneider recommends supplying a brand-style guide—such as fonts, colors, and other constraints—rather than relying only on what the agent infers from a website.

Create a Lower-Cost UGC-Style Video Workflow

1. Write a script and prepare a reference image for the intended avatar.
2. Generate the spoken audio, then send the audio, script, and image together to the video model. This constrains the output and reduces the chance that the model invents dialogue or an unsuitable avatar.
3. Generate a low-resolution first pass to control costs.
4. Transcribe the output and compare it word for word with the script; locate and correct pronunciation errors.
5. Upscale or otherwise process the video, then add subtitles and animated graphics.
6. Use a review step to check the final video before publishing.

The workflow can use AI-generated avatars or human UGC creators. Schneider describes using creator platforms such as Tribe or Sideshift to coordinate scripts and briefs through APIs, with the agent helping manage creator relationships and campaigns.

Add Creative Quality Control

• Treat creative review as a defined job that can be encoded into a workflow.
• A “creative director” agent could inspect video frames, identify problems, and send assets back for revision.
• Add a human approval gate or require a minimum quality score before publication.
• For regulated industries such as financial services, an agent can check creative against permitted claims and revise it through a few review cycles. Schneider says these workflows typically take about two or three cycles.

Build a Swipe and Knowledge Repository

Store examples of effective creative, brand guidance, and reusable workflows so agents can draw on them. The discussion recommends making this repository a continuing team resource, not a one-off collection. Examples of tools mentioned include Sublime, Cosmos, and Graph.

Isenberg suggests giving a creative-director agent reference material from David Ogilvy’s books and leading advertising discussions, as well as a large collection of creative examples.

Operating Principles and Recommendations

• Focus on building the system around the model—data, tools, context, evaluation, and permissions—rather than expecting a generic agent to perform well on its own.
• Treat agents as powerful but potentially unreliable operators: constrain their access, define what they may do, and include review steps for consequential actions.
• Let humans act as coordinators and editors of the system rather than manually performing every production task.
• Use cheap first passes and review them before spending more on higher-quality renders or large production runs.
• Scale creative volume according to budget and campaign economics. The speakers caution that producing more ads is not automatically valuable if the budget cannot meaningfully test them.
• Connect each ad to downstream business outcomes. The goal is to learn which creative brings in customers with stronger lifetime value—not merely which ad gets clicks or initial conversions.

Examples and Metrics Cited

• Search-query qualification: Schneider says an agent checked 5,000 paid-search queries in 20 seconds, assessing whether the search intent matched the landing page before spend was committed.
• AI UGC production cost: The demonstrated video was estimated at about $4.75, described as roughly $5 after adding animated graphics. A run of 10 would therefore cost about $50, and 100 about $500, before any other costs.
• Video resolution: The workflow creates a 480p first pass. Schneider says a 1080p version costs about $14; he describes using a lower-resolution pass and subsequent processing to manage expense.
• Human creator cost: A creator example was priced at about $40 per video. Schneider contrasts low-cost AI production with human production that he estimates could reach roughly $300 for a creator-plus-motion-graphics workflow, and elsewhere refers to a $300–$500 range for comparable work. These are illustrative figures from the discussion, not a controlled cost comparison.
• Creative volume: Schneider describes companies moving from about 10 ads a month toward testing batches of 30 or, in some cases, 300. He emphasizes that appropriate volume depends on ad budget and cost per lead.
• Illustrative campaign example: A batch of 30 ads produced an example in which one ad generated about two demo bookings for $50, while the company’s average price was said to be around $150. The speakers say they might stop spending on an ad after limited results and test other creative; they do not provide enough detail to assess profitability or establish this as a general benchmark.
• Quality and iteration: Schneider claims some generated creative can be roughly 95% good, while acknowledging variable model quality and the need for iteration. These are his assessments, not independently verified results.
• No stated business outcome: The episode does not provide measured campaign revenue, return on ad spend, customer acquisition cost, or lift from the demonstrated ads. Its quantified examples mainly concern production cost, workload, and testing volume.

Presenters and Sources Mentioned

Presenters: Greg Isenberg (host) and Cody Schneider (guest).

Named tools, platforms, and reference sources: Airbyte, ClickHouse, Facebook Marketing API, Google Ads API, Claude Code, Grokbot, Graph, Kai AI, Fal, OpenRouter, Apollo, GitLeads, Seedance, FFmpeg, Mirage Tesseract, Tribe, Sideshift, Sublime, Cosmos, Reddit, David Ogilvy’s books, and advertising podcasts and creative examples.