Taste, Judgment & Building AI Products: What We Learned at the Snowball Panel

In April 2026, Tanya Chou moderated a practitioner panel at Snowball Asia's AI-in-business event in Hong Kong. Around 150 people showed up. The panelists were Augustin Chan (Digital Rain Technologies, co-host of Hong Kong AI Podcast) and Alexander Payne (Signal Eight, co-organiser of AI Tinkerers HK).

No slides. No pitches. Just two people building with AI every day, comparing notes on what actually works.

Here are the ideas that came out of it.

Payments Took Two Months. Agentic Features Took Two Days.

Alex has been building Signal Eight — an agentic CRM for food and beverage sales teams. His salespeople chat with the CRM through voice instead of wrestling with a complex UI. The surprising part: building the agentic functionality (web search, CRM enrichment, conversational interface) was straightforward. The payment and billing infrastructure? That took two months of head-banging.

Aug had warned him to deprioritise billing early on. Alex didn't listen — and learned the hard way that multi-seat SaaS billing is one of those deceptively complex infrastructure problems that AI doesn't simplify.

The takeaway: the "boring" infrastructure often takes longer than the "hard" AI features. Knowing which to tackle first is an 80/20 decision that can save months.

Roll Out AI in Terms of Capability, Not Functionality

When the audience asked about ROI, Aug reframed the question. Measuring AI ROI through traditional software metrics doesn't work well. Instead, think about rolling out AI in terms of capability — what can this team now do that they couldn't before?

The goal is to enable departments to do self-service development, or at least self-service problem-solving. The tools aren't fully there yet, but they're getting there fast. You want to shoot where things are going.

This connects to what Aug calls "surgical software" — smaller, highly adapted systems that fit the user's actual workflow, rather than forcing users to adapt to standardised platforms. When code is cheap (and it is now), you can have the software adapt to the users rather than the other way around. A Forrester article by Fred Giron captures it well: "When code is free, what's left to sell?"

What's left is judgment.

Code Is Cheap. Taste Is Not.

Both panelists kept circling back to the same idea: writing code with AI is effectively free now. The hard part is knowing what to build and why.

Alex put it plainly: "Knowing what to build and why, and actually what's going to be useful — that really comes down to judgment that only we as humans have, at the moment anyway."

Aug went further, arguing that taste — the ability to look at a direction and know it's wrong, or cringe at something generic — is what separates useful AI products from average ones. AI regresses to the mean because it's trained on everything. Taste is the departure from the mean.

The tricky part: taste is hard to teach. Aug suggested the industry might need to return to apprenticeship-style learning — observation and subtle cues rather than exhaustive documentation. You'd learn more watching Alex do a planning session than reading a how-to guide about it.

Agentic Engineering, Not Vibe Coding

Alex pushed back on the term "vibe coding." When it first appeared, it described writing code with AI. But it carried an implication of going with the flow — tweaking things on instinct without a firm plan.

He prefers "agentic engineering": start from the specification, start from the strategy, understand why you're building something, then go from spec to plan to code. That's very different from vibing.

But even vibe coding requires taste. When you're making rapid judgment calls about what looks right and what doesn't, that's taste operating below the level of conscious articulation. The difference is whether you've built up enough domain knowledge for those instincts to be reliable.

Domain Expertise Is the Real Moat

Alex's CRM works because he spent years as a field salesperson. He knows what it's like to do 20 meetings in a week, fly home, and struggle to recall the details for Monday's pipeline review. That pain is specific and real, and it's what makes Signal Eight's design decisions sharp rather than generic.

Aug's methodology works because he's seen dozens of enterprise software deployments in various states of success. His "surgical software" approach — requirements gathering through WhatsApp audio, extensive transcriptions, detailed planning, data model first, AI last — comes from watching what goes wrong when you skip those steps.

In both cases, the AI is the accelerant. The domain expertise is the direction.

Watch the Full Panel

The full 20-minute conversation is available on our YouTube channel, with English, Traditional Chinese, and Simplified Chinese captions. You can also watch the collaboration detail page for participant bios and links.

For the full-length conversations: Aug is Episode 1 (parts 1 & 2) and Alex is Episode 2 (parts 1 & 2) of the Hong Kong AI Podcast.

Get notified when we publish new articles and episodes. No spam, just signal.