Can You Truly Automate Marketing Workflows? Part 1: Automating Asset Creation to Drive Traffic
AI agents make marketing automation sound simple.
Connect the right tools, give the agent a task, and let it create the content, design the assets, prepare the upload, and publish.
For this first experiment, I focused on a relatively straightforward content workflow: creating Pinterest pins in Canva and preparing them for Pinterest upload.
The task felt simple enough. I already had content and images. Canva could be used to combine image and the title, Pinterest supports bulk upload. Claude Cowork could theoretically coordinate the workflow through MCP and connected tools.
In reality, the experience was much more complicated.
1. What I tried to do
The goal was to automate the process of turning existing content and images into Pinterest pins.
The workflow I wanted looked something like this:
Take existing content, pull the relevant images, create Pinterest pin designs in Canva, generate the supporting Pinterest post copy, and prepare everything for bulk upload.
To start, I used GPT and Claude chat to help me create a prompt for Claude Cowork. The prompt was meant to reflect the tools and capabilities involved: Claude Cowork, MCP, Canva, and Pinterest.
Then I gave that prompt to Claude Cowork and asked it to execute the workflow.
On paper, this felt like a good test case for AI marketing automation. The task was repeatable, structured, and not strategically complex. It was not “build a full campaign from scratch.” It was more like, “Take these ingredients and turn them into formatted social assets.”
2. What actually happened
The initial experience was not great.
The first prompt missed a lot of important restrictions and requirements. Some of those gaps were obvious only after the workflow started breaking.
For example, Canva needed direct image uploads, not just image URLs. Pinterest, on the other hand, needed public image URLs for bulk upload.
That created an immediate mismatch. The same image asset had to move through different systems in different formats.
The initial prompt also suggested using Canva bulk upload in places where that was not actually the right approach. What I needed was for Claude Cowork to use the Canva MCP connection to create new pin/image assets directly from a template. Canva bulk upload is useful for certain templated workflows, but it was not the same as having Claude create and export final image assets.
I also still had to manually create the Canva template for Claude to use. I am sure AI could help with that step too, but I did not include that in this test. I wanted to keep the experiment focused on execution: using an existing template, existing images, and existing content to generate Pinterest-ready assets.
Even with that constraint, the workflow produced issues: missing image URLs, weak content, formatting errors, tool confusion, and a lot of troubleshooting.
What I thought might be a simple automation task turned into a full day of refining the prompt and making the instructions much more specific.
3. What AI agent lacked
The biggest learning was that the AI did not lack general capability. It lacked operational context.
It could understand the goal at a high level. It could help create copy. It could interact with tools. It could follow instructions.
But it did not understand the requirements of the workflow.
It did not reliably understand that Canva and Pinterest needed different image inputs. It did not know which step should happen inside Canva versus outside Canva. It did not consistently account for the need to host finished images somewhere public before generating the Pinterest upload file.
It also struggled with sequencing and asset handling
A human marketer would naturally think: first create the image, then export the final image, then host it, then use that hosted image URL in the Pinterest CSV.
The agent needed that sequence spelled out.
The AI also lacked judgment around when to use a tool’s feature versus when to work around it. Just because Canva has bulk upload does not mean Canva bulk upload is the right mechanism for this workflow. Just because Pinterest needs a URL does not mean the source image URL is the final image URL.
Those distinctions are obvious when you understand the workflow end to end. They are not obvious when an agent is operating from a broad prompt.
The biggest gap was not creativity. It was workflow reasoning.
4. What I had to do
To make the workflow work, I had to become much more specific.
I had to define the inputs, the outputs, the order of operations, and the handoffs between tools.
I had to clarify that Canva needed direct image files, while Pinterest needed public URLs of the final exported pin images.
I had to specify that Claude should use the Canva MCP connection to create new pin assets, not rely on Canva bulk upload for the creative generation step.
I had to tell it where images should be saved, how they should be hosted, and how the final Pinterest CSV should be structured.
I had to stop thinking like a prompt writer and be a workflow architect.
AI can automate parts of marketing execution, but it needs a clean operating system around the task. The more tools involved, the more important this becomes.
This does not mean AI automation is useless. It means the value comes after the workflow is clearly mapped and repeatable.
For a one-off task, the time spent troubleshooting may not be worth it. For a recurring workflow, the upfront investment could pay off quickly.
5. What the final automated flow looked like
The successful version would look like this:
Claude pulls the source content and selected the relevant images for each Pinterest pin.
The source images saved as usable image files.
Claude uses the Canva MCP connection and an existing Canva template to create new pin assets.
The final pin images exported from Canva.
The exported images uploaded to Google Drive or another hosting location.
Public image URLs generated for each final pin image.
Claude creates the Pinterest bulk upload CSV using the final public image URLs, titles, descriptions, links, and board information.
I manually upload the CSV to Pinterest (Pinterest <> Claude MCP limitation).
6. What I learned about marketing automation using AI agents and what this means for marketers
The good news is that it is possible to create an automated workflow, but the workflow architecture needs to be very specific and foolproof.
“Create Pinterest pins from this content using this Canva template” is not enough. They need instructions like “Use these source files. Create these assets in this template. Export them in this format. Host them here. Generate public URLs. Use those URLs in this CSV structure. Validate that every required field is complete before handing it back.”
For marketers, this is the shift AI is creating. The work does not disappear. It moves up a layer.
Instead of manually creating every image or every post, marketers need to design the repeatable systems that let AI do those tasks reliably.
Tips and best practices
Start by manually mapping the workflow before asking AI to automate it.
List every input, output, tool, file type, and handoff point.
Be explicit about platform-specific requirements. Do not assume the AI knows that one tool needs image files while another needs public URLs.
Separate the workflow into smaller steps instead of asking the agent to complete everything in one broad prompt.
Use templates wherever possible. AI performs better when the creative structure already exists.
Build in validation. Before upload or publishing, check that image URLs work, metadata fields are complete, formatting is correct, and the final asset matches the intended design.
AI can automate marketing tasks, but it cannot yet reliably automate messy marketing workflows without a human first turning the workflow into a clear, repeatable system.
I’m still working on finetuning the actual workflow/prompt to use for this specific workflow, so stay tuned if you are interested in seeing the details.