GAIC Hangzhou 2026: AI Was Becoming an Industry of Workflows

In March 2026, while spending time in Hangzhou, I attended a GAIC artificial-intelligence industry event. I went as part of HAAM's wider field research into how AI products were moving from model demonstrations into real organisations, creative industries, and operational systems.

The useful part of an event like this is rarely the claim that AI will transform everything. That claim has become ambient. The useful part is seeing what companies choose to place between a model and a paying customer: which workflow they claim, which knowledge they structure, which risks they absorb, and which result they promise to deliver.

Across the exhibition floor, the strongest pattern was clear. The commercial unit was no longer the prompt or even the model. It was the workflow.

The products appeared unrelated at first. One company worked with engineering drawings. Another managed electronic contracts. Red Bear AI focused on persistent memory. Dreamweaver Studio offered an automated system for producing and distributing comic dramas.

Yet each booth was making the same structural argument. A useful AI product must remember specialised context, guide a sequence of decisions, connect several tools, produce an auditable output, and remain present after the first generation step.

This is a major change in product language. The empty chat box asks the user to invent the process. These systems tried to own the process. Their value proposition lived in the chain between intention and outcome.

Jinkou Liangce presented AI-enabled engineering design as an intelligent design engine rather than a general assistant. Its display referred to engineering-industry models, drawing models, tens of thousands of technical standards, enterprise knowledge, calculation, documentation, and drawing generation.

That combination matters more than the presence of a large language model. Engineering work is constrained by standards, geometry, calculation, accountability, local practice, and accumulated organisational memory. A model becomes valuable only when those constraints are represented in the system around it.

The booth made a durable product principle visible: domain AI wins by converting institutional knowledge into a repeatable operating environment. The interface must reveal where information came from, what rule was applied, what remains uncertain, and who is responsible for approval.

Asign's display framed the electronic contract as a data and intelligence hub. Drafting, comparison, review, workflow, reporting, signature, and evidence were presented as connected functions rather than separate documents or utilities.

This is the shape of serious enterprise AI. The generation step may be impressive, but the surrounding controls create trust: identity, permissions, version history, legal validity, review, storage, retrieval, and proof. The product has to survive disagreement, audit, and time.

For product designers, this means governance cannot remain a compliance page hidden after the interface is finished. It has to become visible interaction design. Users need to understand what the system did, what it recommends, what it cannot decide, and how a human can intervene.

At the Red Bear AI booth, the screen introduced Memory Bear 2.0 and a self-reflection mechanism. The product language suggested that memory was not merely chat history. It was an active layer that could organise context, connect prior events, and influence future behaviour.

Persistent memory is one of the boundaries between an AI demo and an AI system. Without memory, every interaction begins from zero. With poorly designed memory, the system becomes invasive, inaccurate, and difficult to correct. The opportunity therefore creates a parallel design obligation: users need to see what is remembered, why it matters, how long it remains, and how to amend or remove it.

The next generation of AI interfaces will need memory controls with the clarity that file systems, calendars, and account permissions gradually developed over decades. Memory cannot remain invisible plumbing when it shapes the system's judgement.

Dreamweaver Studio presented the most complete chain on the floor. Its automated comic-drama workflow moved from an initial idea through story refinement and production to finished-video export, platform distribution, paid promotion, and revenue sharing.

The inclusion of distribution and monetisation was the important detail. Many generative tools stop when an asset appears. Creators still have to adapt files, maintain continuity, find channels, manage promotion, understand platform requirements, and recover revenue. Dreamweaver Studio treated those later stages as part of the product.

That model points toward a broader future for creative software. The winning platform may not generate the most beautiful isolated image. It may be the one that preserves characters and story logic, manages rights and versions, packages the work correctly, reaches an audience, and returns useful economic information to the creator.

The exhibition also showed where the industry remained immature. Many displays were dense with diagrams, arrows, product layers, claims, and technical vocabulary. The systems promised simplification while their own explanations often demanded expert interpretation.

This is not a cosmetic problem. When AI acts inside engineering, contracts, memory, and media production, comprehension becomes part of safety. A user must be able to distinguish source data from generated output, a suggestion from an automated action, confidence from certainty, and a reversible step from an irreversible one.

The design opportunity is therefore larger than making AI feel friendly. It is to make complex agency legible. The interface should show state, provenance, constraints, handoffs, cost, permissions, and consequences without turning the product into another control room that only specialists can operate.

The event strengthened a conviction that now runs through HAAM's work: AI products become durable when they are designed as institutions in miniature. They need memory, standards, roles, evidence, continuity, and a way to improve without quietly rewriting their obligations.

It also clarified where HAAM can contribute. The market already contains ambitious models and increasingly complete automation chains. The missing layer is often trustable interaction design: accessibility, performance, localisation, evaluation, recovery, and a visual language that lets people understand the system before they surrender control to it.

Participation should leave more than a badge photograph. This field note records the change I saw on the floor in Hangzhou. AI was leaving the blank chat window and entering the machinery of industries. The next design question is whether people will be able to see, question, and govern what that machinery is doing.