I joined Pressable to lead product marketing at the start of 2026, and quickly discovered there was a lot of groundwork to lay. Some parts of my job were eating up a lot of time without adding much value, so I started looking for tasks I could automate with AI. The three workflows I’ll be sharing in this article are my favorites, and I can’t imagine doing them manually again. They are: tracking competitive intelligence, creating product update content, and maintaining our marketing knowledge base.

I’ll be honest, none of these workflows worked perfectly the first time. They all took multiple rounds of edits to get right, and they all work a little differently. The first, (tracking competitive intelligence), runs on Zapier, and the other two, (creating product update content, and maintaining our marketing knowledge base) runs through Claude Cowork.

To create them, I spent hours refining AI prompts, figuring out why certain workflows weren’t running as intended, and staring at errors trying to figure out why a file wouldn’t upload. But they’re all pretty well-oiled machines now, and I’m excited to show you what I built. Hopefully, some of what I’ve learned through the process of building them can save you time if you’re building something similar.

In this article, I’ll walk through what I built, why I built it, what broke, and what I learned along the way. I think what’s probably most important to note is that there is a common thread which runs across all three: I knew how to do the underlying work already. AI just gave me a way to remove some of the repetitive steps.

Workflow 1: Competitive Intelligence, Automated Through Zapier

Why I built it

Pressable competes in a crowded market where positioning, pricing, and feature claims from other players shift constantly. There are a lot of elements to track, things move fast in this space, and frankly, I found it difficult to stay on top of everything. Before I built this workflow, my competitive intelligence process was mostly reactive: after a sales call, during a pricing project, or when something specific prompted me to go looking. The last thing I wanted was to find out about a competitor’s pricing change weeks after someone in sales had already lost a deal over it.

How it works

The workflow collects information from multiple sources, analyses it with AI, and then distributes the information which is either useful or important.

I started by defining which competitors mattered most to us and identifying the pages and channels I wanted to monitor. These included home pages, pricing pages, product feature pages, blogs, social media accounts, email newsletters, and news mentions. I tried to be as exhaustive as possible, assuming that including more sources would mean I’d be able to gather better information.

The process consists of individual Zaps for each source and competitor, so there was a lot to set up. Each Zap records the information gathered in a Google spreadsheet, uses AI to determine what’s relevant, provides a summary, and, where relevant, offers suggestions of actions we might take as a business.

For example, for the webpage scrapes, the previous week's scrape is compared to the current one and any changes are highlighted. In addition to this, the changes themselves are analyzed and further information is recorded such as whether this change creates any competitive gaps, and how Pressable might want to respond.

Finally, it takes all the summaries gathered that week and pushes them into a separate document. That document is then sent to a company Slack channel; while the original spreadsheet keeps all of the previous summaries documented so nothing historic gets lost.

What broke

I think a better question might be what didn’t break?!

First of all, there was a lot of noise. For example, the same information would get picked up repeatedly in the same week, or would keep appearing for multiple weeks as new outlets, channels, and sources picked it up. I had to do a bunch of work on my prompt in order to make sure it reported each piece of new information only once.

The other issue was the AI would pick up on something it deemed a change, but it didn’t really have a major impact competitively. This meant I had to also add rules in order to flag only the meaningful changes.

Lastly, I had to reconfigure the information the AI was pulling in (because I was pulling in too much information, and using way too many credits). To resolve this, I ended up creating new tabs that used formulas to pull just the most relevant information on a weekly basis. I then had the AI analyze that smaller dataset, instead of using the AI to analyze all of the data I had gathered.

The lesson

The quality of the data you’re gathering is more important than the quantity. Initially I wanted to track anything and everything, but once I rationalised what I was gathering the output was actually more useful. In reality, there are really only a handful of types of changes that competitors might make that will impact or affect our sales conversations, and I had to revise my AI prompts a few times before I really nailed it.

Workflow 2: From Ship Notice to Published Changelog

Why I built it

Before AI, a new feature would take weeks or months to ship and the updates coming from our product and engineering teams were pretty manageable for me to read through, summarize, and post about on our changelog, in-product news center, and blog. However, once AI allowed them to start shipping multiple features a week (or sometimes a day!), I had trouble keeping up.

Each time a new feature dropped, I’d get an email notification of the update. I’d click through, read the update, draft a summary, log it in a tracking spreadsheet, and send it to the product team for review. Then I’d copy the summary into WordPress to publish it.

For each new feature I would have to go through this whole process. I knew there had to be a better way.

How it works

I started by training the AI on my exact writing style and format for the changelog, news center, and blog. This training is the reason the output sounds like me and not AI slop.

I then moved to Claude Cowork, to do everything else including reading the email, clicking through to the internal post, pasting that into my trained AI chat, copying the entries back out, and pasting them into a document.

Claude also formats the document, pastes the link into my tracking spreadsheet, and fills the rest of the row with information such as the feature name, its prioritization, and more.

From there, it sends me a notification to review the copy, and I usually make a few changes.

Once I feel good about it, I have it send a notification to the product team in Slack. Finally, Claude opens our content management system and pastes in the copy.

But it doesn’t get published automatically, instead, it sends me the link to the draft. I really prefer to do the final read myself and ensure everything looks correct. Publishing to a live site just wasn’t a step I wanted to hand off to AI at this point.

What broke

Again, a lot. The browser tool that gives the AI control of Chrome wasn't even connected, so the first attempt returned nothing.

On top of that, Claude couldn’t pull information from my inbox properly, and I had to ask it to switch to reading a screenshot instead. The spreadsheet step was a little funky too - it kept pasting everything as one big blob of text into one cell instead of inputting the relevant text into each cell, so that required a bit of prompting to fix. Also, it never quite managed to format my Google Doc just the way I liked it, so I decided I was okay fixing the bolding and spacing during my human in the loop step.

The lesson

Don't assume the tool sees what you see. Just because looking at a page with human eyes seems easy and straightforward, doesn’t mean that’s what it looks like to AI. I had to ask AI to keep taking screenshots of things to ensure it was getting the right information versus just going through text.

Also, it takes time to train the AI on your writing style and voice before you can automate any type of writing. That will help avoid heavy editing every time it produces something. Not surprisingly, the more context and historic details the AI has, the better it will get at giving you exactly what you want and avoid the things you are tired of asking it to stop doing.

Workflow 3: Centralizing a Scattered Knowledge Base

Why I built it

When I first started building out our product marketing workflows at Pressable, there was a lot of groundwork to lay. In real terms, that meant pasting a lot of context into each AI conversation.

Over time, I got tired of having to locate the same set of links and documents before I could even start a prompt. Sometimes I’d forget an important piece of information, or the AI would miss a nuance I had mentioned in another conversation, and the output would be incomplete. Meanwhile, team members were asking questions like “how are we messaging this?” or “how do we position that?” and having to recreate the same context in their own AI conversations.

I knew we needed to build something everyone in our team and organization could use repeatedly. However, I also kept putting this project off because it sounded so tedious to have to go through everything, find all the places these documents might live, and make sure I thought of everything. At the time our knowledge base was scattered across my own personal computer folders (with outdated or duplicate files), an internal resource hub, our own website, an internal search tool, and our team’s internal blog. That’s when I realized, I could probably ask AI to do that too!

How it works

I started with my trusty pal Claude Cowork. I asked it to inventory my local files, flag duplicates and outdated files, and compare them against what was already in Google Drive. Anything missing could then be uploaded or updated.

Next, I had it go through our internal resource hub and pull relevant links directly from the pages. It also went through our website and compiled the information into a reference document, which it uploaded to Google Drive. I then loaded the folder into a dedicated AI workspace that anyone on the team could use as a starting point, without having to recreate all that context themselves. Anyone can also upload relevant documents to it as needed. I could also reuse the workflow periodically to scan those same file locations and keep the knowledge base up to date.

What broke

The AI wasn’t perfect at identifying duplicates because often, whilst two (or more) files weren’t identical, the information inside them was essentially the same. It also didn’t always know which was the source of truth and which was outdated, so it did require me to help with the audit. Some files were too big and got left out, and I had to prompt it to ensure I manually uploaded anything it missed.

It also had trouble going through our internal resource hub because it had multiple pages, and I kept thinking, “Just click next page! It’s not that hard!”. In the end, I had to pause it and explain it was missing documents that weren’t on the first page. Turns out everything was already loaded on the page, it was just visually hidden by page number.

The lesson

Don’t always trust what your automated process is doing. Spend some time watching what’s actually happening in your browser, on your computer, etc. instead of assuming the AI knows exactly what it’s doing.

Final Thoughts

One of the biggest things using AI workflows has unlocked for me is tackling projects I would otherwise procrastinate on because they seemed too complicated, tedious, or time-consuming.

Today I use AI not just to build workflows, but also to brainstorm and think through solutions. A lot of the time, I won’t use the output directly, however, it helps me tackle the task head-on, break it up into smaller, more manageable pieces, and get things done the way I’d like to do them.

If you’re just starting to experiment with building workflows, there are two things I’d recommend:

1. Start with a task you already know how to do really well

I was able to make the three workflows I’ve described in this article work because I had a deep understanding of the underlying tasks that needed to be done before I handed any of them off. I already knew what “good” looked like and which signals actually mattered. I didn’t need AI to teach me the task; I needed it to handle the repetitive mechanics of applying my judgment over and over again.

2. Don’t get discouraged if your workflows don't work the first time around

I’m not embarrassed to say that none of my workflows worked perfectly on the first try. It’s a reminder that we’re all human, and that the AI can only do so much with the information that we give it. It can’t (yet!) read our minds, and having a human in the loop is important. Use AI to troubleshoot by explaining what you expected to happen and what actually happened. A lot of the time, it actually helped me figure out what was wrong with itself.