# I was the bottleneck Amplitude's marketing ops team automated request triage using Asana AI Teammates and Claude, cutting a three-person workflow to a single QA review. Source: https://amplitude.com/en-us/blog/automating-request-triage --- [Chelsie Cauthon](/blog/author/chelsie-cauthon) [Senior Marketing Transformation Manager](/blog/author/chelsie-cauthon) [](https://www.facebook.com/sharer/sharer.php?u=https%3A%2F%2Famplitude.com%2Fblog%2F%2Fblog%2Fautomating-request-triage)[](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Famplitude.com%2Fblog%2F%2Fblog%2Fautomating-request-triage)[](https://twitter.com/intent/tweet?url=https%3A%2F%2Famplitude.com%2Fblog%2F%2Fblog%2Fautomating-request-triage\&text=I%20was%20the%20bottleneck)[](mailto:?subject=Checkout%20this%20Amplitude%20Article\&body=Check%20this%20out%3A%20https%3A%2F%2Famplitude.com%2Fblog%2F%2Fblog%2Fautomating-request-triage) A few weeks ago, I messaged a coworker about a bizarre comment in Asana. It looked like it was yelling, and it looked like it was from me, but it wasn’t. **Me:** So sorry about that comment. That was Claude moonlighting as me. **Her:** OMG, creepy. I thought you just decided to randomly type in title case. She wasn’t wrong to feel unsettled. Every marketing ops request that landed in our queue used to wait on me. Not to build anything, just to look at it. To check if the campaign brief was filled in, if the event had a Salesforce campaign, and if the requester provided copy for their emails. That doesn’t happen anymore, and most people haven’t noticed. Well, unless they’ve noticed the random title case. For years, I assumed the slow part of marketing ops was building the assets. I was wrong. Requests hit our queue around the clock, but they could only get scoped between 8 a.m. and 5 p.m. Eastern, Monday through Friday, because that’s when I was at my desk. The slow part was never the work. It was waiting for one person (me) to decide what the request was, if it was complete, and if we should move forward. . ## The build was automated but the queue still crawled Marketing requests come to our team through Asana: a webinar, a field event, or an email send. Before this year, each one passed through three people. - One person scoped it, meaning they read the brief, fixed the naming convention, deleted the subtasks for assets nobody had asked for, and chased down the missing SFDC campaign name. - A second person built the Marketo program. - A third person QA'd it and shipped it. Last year, we implemented a solve for the middle. When a task moves to "QA Requested” in Asana, Workato picks it up, clones the right Marketo program, populates every token from the campaign brief, and builds the emails. This piece saves the team about 1,100 hours a year, or roughly $68,750. The queue still crawled. Workato is deterministic. Deterministic automation can’t run on a request that isn’t clean yet. If the task name is off-convention, the brief is missing the event location, or the subtasks don’t match what the requester checked, the automation doesn’t fire or fires incorrectly. Somebody has to make the request legible first. That somebody was me. A request submitted at 6 a.m. in EMEA sat until I woke up in the States. Not because it was hard, but because I was asleep. ## First attempt: Asana AI Teammates So I tried to automate the scoping step, starting with the [AI Teammates](https://asana.com/product/ai/ai-teammates) beta in Asana. I already lived in Asana, so not needing to introduce a new tool was a real advantage. Two problems surfaced early, but this was a beta, so kinks were expected. The first was approvals. I hit "Always Allow," and we were off. But then the Teammate began asking me to sign off on every action. This undermined the entire purpose of handing over the work. The second was sequencing. The Teammate would act before the campaign brief link existed on the task. This meant it was scoping a request with nothing to read. I solved that with smart rules, making sure the Teammate doesn’t act until the campaign brief is present. After some iterations and plenty of support from Asana, I got the tool to work. This was a custom Teammate I built rather than one of Asana’s pre-built ones. So, it took real tinkering to understand the behavior and learn which instructions and rules to adjust to get the outcome I wanted. It was a good education in prompting and micro-edits. ## Second attempt: the same workflow, written as a Claude skill I wrote the same triage rules as a Claude skill and ran the two side by side. The difference between what the two could do was smaller than I expected, but the maintenance with Claude was more straightforward. When the skill made a mistake, I had a dialogue with Claude, pointed out the misstep, and prompted it to update the skill accordingly. In one instance, it asked a requester to supply the copy for email assets they hadn’t requested in the brief, essentially chasing them for work they never wanted. This is exactly the friction the whole project was supposed to remove. I was able to fix this in a simple conversation. This is now written into the skill as explicit rules: check the request type before deciding which fields are required. Skip the email copy check entirely if no email assets are checked. These rules are in the file. They’re straightforward. I can read them, and so can anyone else on my team. My favorite fix is the least impressive one. Our campaign brief template doesn’t reliably export checkbox state. If you pull the doc programmatically, every asset looks the same whether it’s ticked or not. The Claude skill fixed that. It now opens the brief in a browser and looks at it, checking with its eyes, the way I used to, instead of guessing from context. ## The shorter correction loop wins Triage rules are never right on the first pass. Every queue has edge cases that only surface in production. Each one becomes a rule you didn’t know you needed. So when you’re choosing a tool for this kind of work, you’re really choosing how long your correction loop is going to be. Both of our tools got there in the end. The difference lies in what each cost me to correct. Making a fix with the Asana Teammate meant adjusting instructions and smart rules, rerunning, and watching to see what changed. With the Claude skill, a fix was a conversation. I’d explain what went wrong, and the new skill was in the file a minute later. The question isn’t "which tool is smarter?" but "which one can I fix on a Tuesday afternoon, by myself, without filing a ticket?" Because that’s the big win for a workflow you personally own. Asana AI Teammates was in beta, and betas get better. For now, we’re going with the Claude skill because the loop back to a fix is shorter. ## The human moved to the end of the line With AI, three people became one person at QA. Requests are now scoped in the timezone they arrive in. A request submitted at 6 a.m. in EMEA gets read at 6 a.m. in EMEA. The brief gets checked, and the name gets normalized. If anything is missing, the requester gets a comment listing what’s needed. The whole thing is queued for the build before I’ve had coffee. Better still, if the request arrives complete, the entire program is built by the time I’m at my desk. Once I log in, there’s really only one job left. Look at the finished product and decide if it’s right. The human is still in the loop. I’ve just moved to the end, which is where my judgment actually matters. ## On being replaced by something that types in title case I keep coming back to my coworker’s reaction. There *was* something creepy about the comment, but I couldn’t put my finger on why or what. What Claude replicated wasn’t my judgment. That part I could write down, and writing it down made it sharper than it had been in my head, where it lived as a set of habits I’d never had to state out loud. What it replaced was my availability, which is what my stakeholders around the world were waiting for. Ultimately, the question isn’t whether AI can do your job. It’s which part of your job is actually judgment, and which part is just you being awake. Those are very different things, and only one of them should be a bottleneck. ##### Ready to use AI to transform your product? Amplitude Agents help you understand your users more easily than ever. [Get started now](/signup?source=blog-ai\&topic=ai\&siteLocation=blog-inline-cta) About the author Chelsie Cauthon Senior Marketing Transformation Manager [More from ](/blog/author/chelsie-cauthon) [Chelsie](/blog/author/chelsie-cauthon) Chelsie Cauthon is a Project Manager who thrives at the intersection of strategy and execution. Drawing from a background in Marketing Operations, she builds the systems, automations, and processes that keep teams moving while turning complex operational challenges into clean, scalable solutions. Topics [AI](/blog/tag/artificial-intelligence) #### Recommended Reading [Read ](/blog/ai-agents-need-good-context) [Insights](/blog/ai-agents-need-good-context) ###### [Your agents are only as good as your data context](/blog/ai-agents-need-good-context) [Sep 4, 2026](/blog/ai-agents-need-good-context) [9 min read](/blog/ai-agents-need-good-context) [Read ](/blog/trust-economy-financial-services) [Insights](/blog/trust-economy-financial-services) ###### [The New Trust Economy in Financial Services](/blog/trust-economy-financial-services) [Sep 2, 2026](/blog/trust-economy-financial-services) [11 min read](/blog/trust-economy-financial-services) [Read ](/blog/agent-data-security) [Product](/blog/agent-data-security) ###### [How to Secure AI Agent Traces Without Losing the Signal](/blog/agent-data-security) [Aug 28, 2026](/blog/agent-data-security) [7 min read](/blog/agent-data-security) [Read ](/blog/connecting-agent-performance-to-product-outcomes) [Product](/blog/connecting-agent-performance-to-product-outcomes) ###### [Connecting agent performance to product outcomes](/blog/connecting-agent-performance-to-product-outcomes) [Aug 20, 2026](/blog/connecting-agent-performance-to-product-outcomes) [10 min read](/blog/connecting-agent-performance-to-product-outcomes)