Prepared with AI assistance from the linked primary sources. Company and government claims are attributed to their issuers. How we use AI.

A chat window is an interface, not a description of an entire AI system. Developments across Asian businesses show a wider set of tasks: checking engineering documents, helping employees retrieve information and moving materials inside an operating facility. These examples do not establish a universal adoption rate. They show why the work being changed deserves more attention than the appearance of the software.

Engineering work has its own constraints

On 19 February 2026, Panasonic Connect announced internal deployment of a Manufacturing AI agent that extracts information from PDF drawings and compares specifications. The company describes a semi-automated checking process supported by Snowflake’s Cortex AI. This is a defined engineering task, rather than a general conversation service. Panasonic’s deployment announcement.

The distinction matters because the output has to fit a decision someone already makes. A plausible explanation of a component is less useful than a correctly identified mismatch, tied to the relevant document version and presented for review. Our assessment is that these narrower tasks offer a more concrete basis for evaluation: teams can define the input, the required result and the cost of getting it wrong.

Enterprise adoption has several layers

DBS’s October 2025 account of responsible AI describes customer-service assistance, internal information retrieval and fraud monitoring. These are different tasks within the bank’s wider AI programme; its descriptions remain the company’s account of its operations. DBS’s operational examples.

That breadth makes aggregate AI claims difficult to interpret. A fraud model, an employee writing assistant and a document classifier have different objectives. Putting them under one adoption total can conceal whether each service is accurate, used regularly or economical to maintain. Readers should look for the task behind the headline.

Some systems act in physical environments

DHL reported a December 2025 deployment at its Singapore Asia Pacific Advanced Regional Centre: an autonomous vehicle moving materials within Infineon’s operations. Its release distinguishes that in-plant use from a proposed public-road expansion subject to approval. DHL’s deployment description.

Physical actions introduce a different set of questions from text generation. A failure may interrupt a process rather than merely produce an unhelpful answer. The relevant evidence includes the permitted operating area, intervention procedures and the handoff to people when conditions change.

Singapore’s January 2026 agentic AI guidance similarly focuses on limits to agents’ authority, meaningful human accountability and lifecycle controls. It is guidance for responsible deployment, not proof that every agent already meets those conditions. IMDA’s framework announcement.

The next meaningful milestone

Watch whether these systems remain useful after their initial launch: are exceptions handled, outputs checked and responsibilities clear? The takeaway is not that chatbots have disappeared. It is that AI becomes operationally significant when an organisation can explain the work performed, the boundaries imposed and the evidence used to keep it running.

Sources

See something we should correct? Contact the newsroom.