A Practical First Agent for Marketing Ops: UTM Generation

I led marketing operations at a cybersecurity software company that marketed to both businesses and consumers. Around 60 people worked across marketing operations, demand generation, social, and paid marketing, and all of those teams needed campaign URLs they could trust for their different audiences and channels.

The process started with a Google Sheet containing dropdowns for UTM values, a field for the base URL, and a formula that assembled the finished link. As the sheet grew, it became unruly: people had to scroll through a large interface, selections were incomplete or incorrect, and the limited logic did little to ensure that links actually worked. A colleague later rebuilt it as a Google Form, which made the interface cleaner but did not solve the underlying problem. I was still seeing malformed links from both the sheet and the form, which made this a good candidate for a more dependable first AI-agent experiment.

The company’s objective was to improve tracking consistency and reporting by making sure campaign links were correctly built and tagged. My objective was narrower and more practical: design and validate a workflow that produced correct URLs, gave people sound guidance on UTM values, and handled several links in one request instead of making them repeat the process.

I owned the workflow and the first agent experiment, but I did not own the broader reporting system or the final rollout. The initial constraint was tool access. ChatGPT could generate and reason through the links, but it could not connect to Google Sheets, Notion, Slack, or Salesforce, so the first version had to work as a standalone agent with manual review. I judged the test against three defined signals: whether malformed inputs produced usable links, whether people could choose UTM values with guidance rather than guesswork, and whether one request could generate the links needed across multiple channels.

I started with a standalone GPT-based agent in ChatGPT because that was the tool I could actually use in December 2025. Dust was not accessible to me until around March, and its external connections came later, so I did not wait for a more sophisticated setup. I wanted to test whether an agent could solve the core workflow before investing in integrations.

First, I defined the agent’s rules for cleaning and building links. If someone supplied a bad URL, it stripped the existing tracking data and applied consistent structure and capitalization rules. It could also return several links in one request, such as versions for LinkedIn, Twitter, and Meta. That directly addressed the sheet and form’s biggest practical failures.

The taxonomy required a second workstream. I worked with the analytics team to define what medium, source, and the other UTM values meant, then documented those rules in Notion and exported them into the GPT. The agent could infer values from plain-language requests such as “I need a URL for a marketing email,” but explained the approved options when the intent was ambiguous.

Because ChatGPT could not write to an external system, I added a prompt asking users to copy their outputs into a dedicated Slack channel. I reviewed those examples and used them as a feedback loop. That surfaced an undefined taxonomy for partner marketing and an edge case where an anchor in the URL caused the hashtag to sit in the wrong place, preventing the UTMs from appending correctly. I added both findings to the documentation and updated the agent’s rules.

The first version was deliberately minimal: a GPT-based agent in ChatGPT with no integrations, including no Google Workspace. ChatGPT proved that smarter logic could improve the process. It cleaned inputs, generated multiple links, and guided people through UTM choices instead of relying on dropdowns they might not understand. In this version, utm_campaign was free text. I used Notion for the standardized UTM definitions built with the analytics team, then exported those rules directly into the agent. A Slack channel served as the manual audit point because the agent could not write anywhere I could review.

I then began testing a move to Dust. Dust could connect the logic to Slack, Google Sheets, Notion, and Salesforce, keeping the instructions grounded in current sources and integrating the workflow more deeply with our technology stack. Its Salesforce connection could read the campaign object’s UTM campaign field and reuse an existing value rather than letting someone invent a new variant. In Dust, that field became part of a standardized naming structure; the agent read it but did not write to Salesforce.

ChatGPT was the important proof of concept, while Dust opened the door to better-maintained logic and future automation. I would use the same progression again: prove the judgment and workflow simply, then add integrations once the rules are reliable. I was the primary tester, with a few colleagues piloting the process, and I cannot claim a full Dust rollout.

Over roughly two months, the ChatGPT agent solved the three problems I set out to test. Bad input no longer broke the output, the agent guided people toward the right medium and source instead of leaving them to guess, and one request could produce every link needed for a campaign. Work that previously took several passes through the form became one request, which made bulk link creation much faster. Roughly 20 people on the marketing team used the agent and created a few hundred links.

The Dust version ran for roughly another two months and moved campaign-value standardization forward by reading existing values from Salesforce instead of relying on free text. I did not complete a full Dust rollout or measure a reporting lift, so I would not claim those outcomes. The remaining adoption gap reinforced a broader lesson: people should not have to remember to visit a separate tool. The next step would be to make the link-generation agent callable from the places where campaign work already happens—by a person, another AI agent, or a broader campaign workflow. That would turn it from a destination people must remember to visit into shared infrastructure for marketing operations.