The New Solution Architect: Why the Bottleneck Isn't Code Anymore

​In Part 1, I argued that AI commoditized building. Value shifted to knowing what to build. In Part 2, I showed how that discovery works. Finding the knot beneath the stated problem.

But that's an oversimplification. I made it sound like discovery happens first, then building happens second.

Reality is messier.

‍

The Knot Beneath the Knot

You find the real problem. You start building. Then you hit another knot.

This happens constantly. Organizations are layered. The stated problem sits on top of unstated processes, which sit on top of inconsistent data, which sit on top of tribal knowledge that never got documented.

Every layer reveals something new.

Story 1: The Construction Data That Kept Unraveling

We were building a demand forecast for a heavy equipment dealership. Partway through, the client asked if we could incorporate their construction data. Upcoming projects in their region. Should help predict demand, right?

"If we add it directly," I said, "we might double-count. Some of these projects are probably already reflected in your baseline forecast, just implicitly."

We landed on incrementality: only flag projects that represent demand above the baseline.

But then: "Even if a project is incremental, does that mean you'll capture it?"

They knew some opportunities were theirs to lose. Others were long shots. The data couldn't tell us which.

The solution: human-in-the-loop. The system surfaces net-new opportunities. They decide whether to "add to demand" based on what they know about the account.

What started as "add construction data" became a layered system with incrementality logic and human judgment built in.
‍

Story 2: The Protein Bars That Were Really About Ingredients

A client came to us drowning in spreadsheets. "Help us predict demand for our protein bars."

We started building a demand forecasting model. As we got deeper into the work, connecting to their data, understanding their workflows, something became clear. They didn't actually care about protein bar forecasts in isolation.

What kept them up at night was one step downstream. Knowing when and how many ingredients and packaging materials to order.

The demand forecast was a means to an end. The real pain was supply chain optimization for components. That's what we ended up building.

We also discovered that different teams maintained different versions of their data. Sales had one set of numbers. Operations had another. We had to figure out which source to trust, which required understanding their organization, not just their data.
‍

Discovery Recurses

These stories share a structure. Client states a problem. You start building. The build reveals another layer. That layer requires another judgment call. Sometimes several.

Discovery doesn't hand off to building. It continues through building.

AI is powerful here. It processes data fast enough to expose inconsistencies. It builds quickly enough that you hit these walls in weeks instead of months.

But AI can't decide what to do when the wall appears. It can't recognize that construction data might double-count demand. It can't navigate the politics of which data version is authoritative. It can't design a human-in-the-loop solution because it understood that some decisions require local knowledge.
‍

Accountability

Judgment calls are only as real as the person willing to own them.

AI can generate options. Surface inconsistencies. Recommend paths forward. But it can't take responsibility when the choice turns out wrong. It can't sit across from a client and say "your current process has a gap, and here's what we should do about it."

Accountability is what makes judgment calls stick. Without it, you're generating possibilities. With it, you're solving problems.

When things get messy, clients need a person to call. Someone whose reputation rides on the outcome. Someone who says "I've got this" and means it.

That's not a limitation of current AI that better models will fix. It's how trust works.
‍

The Real Model

It's not "humans discover, then AI builds." It's continuous collaboration at every layer, with humans owning the judgment calls that require organizational context.

AI makes you faster. It makes custom solutions viable. It surfaces problems you might have missed.

But the knots keep coming. Every layer you touch can reveal something that requires human sense-making. The job isn't to find the first knot. It's to keep finding them, keep making calls, and keep owning outcomes as the build progresses.

That's the work. And it's more valuable now than ever.