We Built an AI Workflow-Only to Throw It Away
It's 11 PM. I'm staring at a broken workflow.
Not broken in a dramatic way. Broken in the quiet, annoying way where it almost works. Ninety percent of the contacts coming through are clean. Ten percent crash the whole pipeline and I still can't figure out why.
This has been going on for six weeks.
Hey friend, Rich here.
I own our Go-To-Market strategy at The Change Fluency Co., which means I live in the messy middle between "we should be talking to more people" and "someone has to actually do that." And for a while, that someone was doing it manually.
For every decision-maker we wanted to reach, every leader, every executive with a consulting budget, someone on our team was digging through LinkedIn, pulling context, copying it into HubSpot, our contact database. Hours. Every week. The kind of work that isn't hard, just relentless.
Not Jay. He flies too much.
So I decided to automate it. I built a workflow in n8n, an AI and automation tool that connects your apps and automates the steps between them. The idea was simple: take a lead, look them up, push the relevant context into HubSpot automatically.
I was fascinated by the problem. I genuinely love this stuff. Give me a workflow to untangle and I will happily disappear for a weekend.
So I disappeared for a weekend.
The first version kind of worked. The second version mostly worked. As the team continued to use the workflow more edge cases showed up, breaking the automation.
What if there's no LinkedIn URL?
What if someone enters a conference name instead of a contact?
What if HubSpot already has the record and I'm about to overwrite clean data with garbage?
I fixed those. Then I hit a wall I couldn't fix. The whole workflow depended on third-party scrapers to pull LinkedIn and Instagram data. Fragile services I didn't control. It worked until it didn't, and when it didn't, everything stopped.
That's where the obsession curdled into frustration. Six weeks in, 11 PM, staring at a workflow that was so close and still broken.
Then I gave Claude Cowork the task.
Not to fix the workflow. I was testing something else entirely.
I described the research we needed on a lead and told it to go find them. And then I just watched. My browser opened. Claude started navigating, pulling up LinkedIn, cross-referencing, finding contact details and background, moving through the internet the way a person would.
I sat there and ran test after test. What about this format? Can you pull this field too? What if the LinkedIn page has no company listed?
Every problem I threw at it, it handled.
It was beautiful.
At some point I opened Loom, hit record, and captured the magic. Sent the video to Jay on WhatsApp without saying anything.
Jay was wrapping up dinner. He glanced at it, put his fork down, and texted back: "No way. This will 3x our productivity. You can run this on a daily basis without you."
He was right. The blocker that had consumed six weeks of iteration wasn't a blocker anymore. Claude in Chrome browses LinkedIn and Instagram the way a person would. No scraper. No API. No fragile dependency sitting between us and the information.
I had spent six weeks building an elaborate solution to a problem that had just been quietly solved for us.
Here's the part I want to be honest about.
Before we ever touched n8n, we'd done something that turned out to matter more than any tool decision. We wrote out exactly what our process actually was. Every step.
Open the post. Find the name. Search LinkedIn. Copy the title. Research the event. Open HubSpot. Find or create the contact. Paste everything in.
Written out, it was seventeen steps.
And immediately, four of them obviously shouldn't exist. Habits from old tools. Artifacts of a process that had grown sideways over time. We cut them before we automated a single thing.
You cannot automate waste. You have to see it first.
That same mapping discipline forced us to define what "done" actually looked like. What does the output need to contain? What format does HubSpot need it in? Most teams skip this. They go from "we do this manually" straight to "let's get AI to do it," and then they're surprised when nothing works, because no one ever defined what working looked like.
When Claude Cowork arrived and we realized it could replace the scraper dependency entirely, it wasn't confusing. It was a clean swap. We knew exactly what it needed to replace because we'd already drawn the map.
The six weeks weren't lost. They were the map.
Maybe you're in business development and all of that made sense. Maybe not.
If you're in sales, you've felt this. The research pile that never shrinks. The CRM that's only as good as what someone remembered to update. The hours between "I have a lead" and "I have enough context to actually reach out."
If you're not in sales, swap the terminology. The same sequence plays out in operations, HR, finance, any function where a process has been running long enough to grow sideways. The tools in this story are specific to what we do. The sequence isn't.
Map Your Process: Start with one process. Write out every step, more than you think there are.
Streamline: Look for the steps that shouldn't exist and cut them.
Define Done: Define what done looks like before you touch any technology.
Tinker with Tech: Then, and only then, bring in the tool, whether it’s AI, Power Automate, or something else. It will be clear what you can use to accelerate your effectiveness.
The tools will keep changing. That part I can promise you.
The process clarity you build by doing this work is the only thing that doesn't expire. If you’re curious about turning a messy process into an AI workflow, let’s chat.
Happy tinkering.
Richard Wong is co-founder of The Change Fluency Co. and works at the intersection of AI strategy, workflow design, and organizational change. Alongside Jay Kiew, he helps teams move from AI curiosity to AI capability. Change Fluency is available now.
