I Let AI Plan My Entire 3-Day AI EXPO Schedule — A Field Report from a GAI Consultant

This time I attended DIGITIMES AI EXPO 2026 as a GAI Solution Consultant at dentsu Taiwan, while also bringing my own practitioner's lens from building a Multi-Agent Ecosystem architecture. Before the event even opened, I fed the AI EXPO website into Claude, parsed out the full three-day dual-stage schedule, described my background — Agentic AI consultant and architect, running my own multi-agent system, translating frontier AI into enterprise-deployable solutions every day — and asked it to filter which sessions were must-attend, worth-attending, or skip-entirely. It mapped out stage transitions, booth priorities, and a written strategy guide. The result was a three-day battle plan better than anything I would have built myself — while I was still working on Tuesday.

Three days in, I was validating one question I actually cared about: Where is enterprise Agent adoption really stuck? The exhibition floor gave me one answer. The speaker sessions gave me another.

The Exhibition Floor Was Mostly Disappointing — With One Exception

The exhibition floor, honestly, let me down.

The dominant pattern: take an existing product, add an AI label. Or take an AI product, add an Agent label. Done. Slow, not solving new problems — following the discourse, not the era.

Lobsters are everywhere on people's lips, but not every one of them knows why it's there.

One booth stopped me for a long time though — AI SOLLY, an AI social practice companion designed specifically for children with special needs. The design philosophy came from real classrooms: letting children practice conversation, social interaction, and emotional processing in a safe environment, filling gaps in special education resources. In a floor packed with commercial applications, the most genuinely realized use of AI was helping special-needs students learn. I shared this on the spot with my mother — the principal of Hung Jen Catholic High School (宏仁高中) in Chiayi, who has been actively exploring AI in experimental education. Compared to all the commercial software still desperately trying to "add an Agent feature," AI SOLLY showed me something else entirely — it's not AI plastered on top of something. It's a product that only exists because AI made it possible. That's what real change looks like in the AI-Native era.

The Speaker Sessions Were Better Than Expected

The quality of the talks far exceeded the exhibition floor. Three days moving between the Wisdom Hall and Future Stage, I found several genuinely meaningful threads.

The Economics of Inference | NTU Prof. Hung-Min Hsu
NVIDIA AI Lab program lead, approaching the next chapter of LLMs from inference cost. The fundamental nature of autoregression — predicting the next token by looking only at previous ones — is the root cause of LLM instability. That's exactly why context governance and control planes aren't nice-to-haves. They're prerequisites.

Agentic AI Enterprise Deployment | AWS Yang Shu-Wei
"Getting agents to production is still too hard" — that one sentence framed his entire session, and it's the exact wall I hit every time in consulting: after the PoC wraps up, the first question from management isn't "can the tech work?" It's "who can use it, how do they use it, and who's accountable when something breaks?" Bedrock AgentCore's architecture answers exactly that — Runtime, Identity, Policy, Observability all built in. AWS isn't giving you a better tool. They're turning the control plane you'd have to build yourself into a managed service.

GMI Cloud | Fred Jhang
Sharp-eyed friends might notice GMI wasn't in my original battle plan — I had initially ranked hardware infrastructure as lower priority. But after seeing their product I immediately added it to the schedule. GMI's architecture is a full vertical stack: Taiwan's AI Factory with 7,000 NVIDIA GB300 chips, GPU as a Service, up to MaaS — "one API, every model" — GPT, Claude, Gemini, DeepSeek, Llama 4 all on a single bill.

I'm currently running dual 3090s for Qwen, building an API Router to route across inference providers, spending significant time manually handling different LLM output formats, quota management, and fragmented billing. That's exactly the problem GMI MaaS solves. More importantly: the value of a unified API layer isn't just reducing cognitive overhead for humans managing accounts. When an Agent needs to dynamically call different models mid-execution, a single API layer keeps the orchestration logic dramatically cleaner. This isn't a concept — it's something I'm building by hand that they've already turned into a service.

The Agent-Native Era | HCLTech Dr. Paul Yu-Chun Chang
The slide Dr. Chang brought back from Germany had me standing still for a moment: "ARE YOU AGENT-NATIVE YET — OR JUST AI-ENABLED?" He broke enterprise AI evolution into three layers: Cloud-Native changed where software runs, AI-Native changed how systems reason, Agent-Native changes who executes work — humans shift from being executors to intent-setters and governors. The executive takeaway was unambiguous: Copilots help. Agents execute. Winners redesign value chains early. We talked after about the gap between Taiwan and Europe — he's optimistic about Taiwan, I'm more reserved — but he said something I'll carry: someone has to say it first, before anyone else will be willing to do it.

AI Agent Evolution | Google Cloud Neo Chen
Neo presented GCP's complete Agent management framework — Gemini Enterprise positioned as the unified entry point for all enterprise Agents, regardless of who built them or what tools they used, with governance built directly in. The contrast with AWS's session was fascinating: AWS comes in from how to build agents, from infrastructure and architecture decisions; Google Cloud comes in from how to manage agents, from unified enterprise control. Both talking about the same deployment challenge, from completely opposite angles.

From Copilot to Colleague | Microsoft Wu Tzu-Chiang
If Dr. Chang gave us the framework, Wu gave us the enterprise problem list: Can IT have visibility into every Agent operating in the organization? Does the Agent's behavior comply with enterprise policy? Who's accountable when sensitive data is accessed? After running my own 38-agent system and then trying to implement Agent governance inside a corporate project, every single one of these questions showed up in real life. Microsoft Agent 365 bills itself as "the control plane for agents" — not a better Copilot, but the layer that makes those questions have answers. Intelligence + Trust. Drop the Trust half, and your Agent won't survive inside an enterprise.

GEO Survival in the Post-Search Era | Apier Chen Dai-Min
When the search bar disappears, the logic of ads and content has to be rewritten. Chen's three-layer GEO framework was memorable: Technical Layer (AI can find it, AI trusts it), Content Layer (AI understands it, easily extractable), Trust Layer (AI knows who you are, believes you're real). The concept that hit hardest was Extractability — when AI becomes the only gateway, a brand's challenge is no longer ranking or traffic, it's being extractable: being ignored by AI is equivalent to not existing. This reframes budget logic for brand owners, and it's exactly what we've been debating internally.

2026 Tech Salary Trends × Getting Into Foreign Companies | Robert Walters × Tech Craft Podcast
Beyond the technical stack discussions, one session reconfirmed where this era's real opportunity lives.

Xiao's data was more direct than any JD. Taiwan's biggest hiring challenge isn't budget — it's "can't find the right people" (69%). Job movers with in-demand skill sets can expect 10-20% salary increases. But one global figure deserves attention: 52% of employers are deploying AI, and 34% of them are doing it to optimize headcount, not expand it — companies don't want more people, they want the right people. That gap is the real opportunity of this era.

The Most Honest Observation After Three Days

Taiwan doesn't lack people interested in AI. What it lacks are people willing to seriously think through what a governable AI system actually looks like.

The Agent era isn't just about making AI do more things. It's about answering: who's responsible? Where's the boundary? What happens when something goes wrong?

Almost no one on the exhibition floor was answering those questions. In the speaker sessions, some people had already started.

AI SOLLY, that small booth, stayed with me not because of the technology but because of a simple fact: a truly AI-Native product isn't one where AI is added on top of something that already existed — it's a product that for the first time becomes possible because AI exists.

That standard is the measure I use to evaluate this era.

This article was drafted through intensive dialogue with an AI assistant over three days of sessions. All source material, perspectives, and content verification are the responsibility of the author.