Turning expert knowledge into automated workflows

Rachel (name changed) managed compliance for a real estate brokerage. Her job was to make sure every transaction met strict legal requirements. Signatures needed to be verified, documents needed to be checked, and discrepancies needed to be caught before they became liabilities.

It was repetitive, high-stakes work. And there was a lot of it.

To keep up with the volume, her team relied on third-party BPO firms. On paper, outsourcing made sense. In practice, it created new problems.

• It was expensive. High labor costs ate into margins.
• It was slow. Offshore teams in different time zones meant constant delays.
• It was error-prone. Language barriers and manual handoffs led to frequent mistakes.
• It was fragile. The legacy software Rachel’s team used had no APIs, which made traditional automation tools useless.

I looked at what was already out there. Zapier, Retool, UIPath, and of course the BPO firms Rachel was already using. I evaluated each one on budget, accessibility, and whether it could actually handle the complexity of Rachel’s workflow.

Traditional RPA tools like UIPath were powerful, but they required near-developer expertise. Rachel’s team couldn’t use them without engineering support.

Simpler tools like Zapier and Retool were easy to pick up, but they could only handle linear logic. “If this, then that.” Rachel’s work didn’t move in straight lines.

And the BPO firms were the status quo. Expensive, slow, and hard to scale.

The gap was clear. Nothing on the market could handle genuinely complex workflows while still being accessible to someone like Rachel.

I spent time with Rachel and her team mapping out what the compliance process actually looked like in practice. What I expected to be a fairly standard checklist turned out to be far more complex.

A single transaction could involve 15 to 20 different documents. Each document had its own set of validation rules. And the process was full of conditional logic. If a signature was missing on one form, the next step was completely different than if it was present.

The thing that stood out most was that none of this was written down. Most of the knowledge lived in Rachel’s head. She wasn’t just following rules. She was making judgment calls at every step.

That was the core insight. Any tool we built for Rachel needed to support the way she actually thinks: flexible, branching, and full of decisions.

My first instinct was the simplest approach: a linear, step-by-step execution model. You define a sequence of actions, and the system runs them in order.

For basic workflows, this worked fine. But the moment I tried to map Rachel’s actual compliance process into it, the model fell apart. Her work branches. One missing signature sends the entire flow down a different path. Linear execution just couldn’t represent that.

So I started looking at how other tools handle branching. Flow-based programming, node editors in creative software, visual scripting in game engines. The same pattern kept showing up: an infinite canvas where you connect things together visually.

That became the core design decision. Visual blocks on a canvas, connected with conditional logic. Each block is one step. The connections between blocks represent the decisions.

For Rachel, this meant she could actually see her workflow laid out the way she thought about it. Not as a list, but as a map with paths and branches. It matched her mental model.

The final design let users create blocks for each step in their workflow, connect them with branching logic, and write instructions in plain language.

We layered in large language models so Rachel didn’t need to learn any technical syntax. She could type something like “Check if the signature is on line 35 of the buyer agreement” and the AI would know exactly what to do. No code, no technical vocabulary. Just the language she already used every day.

That was the moment it clicked for her. The barrier between what she knew and what the system could do disappeared.

Rachel’s firm ran legacy software that had no API connections. There was no way to plug into it with traditional automation tools.

Instead of relying on APIs, Momentum could interact directly with the computer screen. Clicking buttons, scrolling through documents, reading text, verifying information. It worked the same way Rachel did, just faster and without getting tired.

As the sole designer, I built a component library where each block type shared a consistent visual language. The system scaled to support over 40 different node types while keeping everything visually consistent. For teams stuck on older systems with no modern integrations, this was the only way to automate.

My role went well beyond designing interfaces. I led the definition, proposal, and scoping of most of the product’s features. I created templates for how we wrote feature proposals and organized our product thinking. Day to day, I worked in Notion alongside the founders and engineers, helping decide what got built in the first place.

I also built the brand from scratch. Logo, color system, typography, brand guidelines. The marketing site tied it all together into one story, serving as our front door for demos, investor conversations, and recruitment.

We designed Momentum to replace Rachel’s costly, BPO-led compliance workflow with something she could own and run herself. Using her existing outsourced process as the baseline, here’s the impact we projected:

These projections were modeled against Rachel’s existing outsourced workflow, based on the reduction in repetitive manual handling and review time.

This project taught me how to design AI tools for real people. Not for engineers or power users, but for someone like Rachel who just wants to get her work done faster and more reliably.

The hardest part wasn’t making the AI capable. It was making Rachel believe she could actually use it. I learned that trust isn’t built by explaining how the AI works. It’s built by giving people control. Letting them set the rules, see what the system is doing, and correct it when it’s wrong.

If I could go back, I’d invest more in structured user testing earlier and focus more on onboarding. The canvas was powerful once you understood it, but the leap from an empty screen to a working workflow was steep. A library of starter templates would have made that first experience much smoother.

Owning product, brand, and marketing on the same project changed how I think about design. I stopped seeing it as one phase in a process and started seeing it as a lens that touches everything.

We started Momentum with a simple question: where is AI actually going?

Our belief was that AI would eventually run complex, multi-step workflows on its own. Not just single tasks, but entire business processes with branching logic, document handling, and real decision-making. In 2023, that future wasn’t here yet. So we worked backwards from that vision and built the scaffolding.

Looking back, the bet was right. The workflows people build today with tools like Claude, Cursor, and multi-agent systems are exactly what we were designing toward. We just got there before the technology fully caught up.

The future we designed for is becoming the present.