The “Walking Fairway” Test

Designing a Voice AI That Works Where Software Usually Fails

Golf isn’t played in front of a screen—it unfolds in motion, under sunlight, in conversation. That makes it a surprisingly hostile environment for traditional software. When we partnered with Wolfhammer, the requirement wasn’t simply to “add voice.” It was to capture structured, accurate game data without interrupting the player’s physical rhythm or social flow. In other words, this was a real-world interface problem, not a feature request—and most systems fail at exactly this boundary.

The Problem: The “App Gap” Traditional apps assume user attention. Real life doesn’t. In golf, even a simple action—recording a score—introduces friction:

• Pull out the phone

• Unlock it

• Navigate UI layers

• Input structured data manually

  
  

Now add complex side games like Wolf, Vegas, Nassau. These aren’t just scores—they involve dynamic rules, rotating roles, and contextual scoring. The result:

• Cognitive load increases

• Game flow breaks

• Adoption drops (usually by mid-round)

  
  

We call this breakdown the App Gap: the space between how software expects users to behave, and how humans actually behave in motion. Voice seems like the obvious solution—until it meets reality. Standard voice systems break down in three critical ways. 

1. The Correction Loop

Speech recognition is probabilistic. Even a 95% accurate system fails in practice—because that 5% requires interruption. In a live setting:

• “No, not Mark—MARC.”

• “That was a 5, not 9.”

Every correction collapses user trust.

2. The Context Collapse Problem

Humans don’t speak in structured commands. A single sentence might include:

• Game data

• Social banter

• Interruptions

Example: “Yeah, yeah nice shot… okay Mark got a 4 there… wait, whose turn is Wolf?” Systems can fail if they treat voice as linear input, not multi-threaded intent.

3. The Reliability Gap

Backends don’t accept “almost correct.”

In sports systems:

• One incorrect score corrupts the leaderboard

• One missed event breaks game logic

Voice AI must produce: Deterministic, structured outputs—not raw transcripts

To solve this, we built an AI system embedded into a smartwatch experience that continuously understands human intent as a real-time stream and transforms it into structured, validated, and context-aware actions with precision and reliability.

One of the more important design decisions was recognizing that voice alone isn’t enough—experience matters. Instead of positioning voice as a utility, we embedded it into a game layer. Rather than navigating menus to configure complex games, players simply speak naturally: “Start a Wolf game with the same group.” The system responds, guides, and participates, turning setup and interaction into part of the experience rather than a task. This shift—from tool to interaction—dramatically improves adoption because users aren’t learning a system; they’re engaging with it.

The result a better voice software that disappears into the background. Players capture scores mid-round without breaking stride. There’s no learning curve because the system understands natural speech. Most importantly, the system aligns with how people already behave, rather than forcing behavioral change.

At its core, this work reflects a broader belief: meaningful AI systems are built by deeply understanding the problem space—its constraints, edge cases, and real-world unpredictability—and then engineering for precision, reliability, and trust at every layer. The real challenges are prediction, generation, and designing systems that consistently translate messy, human intent into accurate, structured outcomes without losing context or fidelity. When that level of robustness is achieved, AI stops being an experimental layer and becomes a dependable infrastructure that quietly powers critical decisions.

That is the standard we build toward—where clarity of problem understanding and rigor of engineering come together to produce systems that simply work, in the environments that matter.