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

Giving employees access to an AI tool is a visible step. Changing how an organisation works is harder to observe and harder to complete. The difference may explain why businesses with access to similar products can obtain very different results. Their capabilities in data preparation, process design and ongoing evaluation can matter as much as the tool itself.

Singapore’s Digital Economy Report 2025 reported AI adoption among SMEs rising from 4.2% in 2023 to 14.5% in 2024, compared with 62.5% among non-SMEs in 2024. These are Singapore-specific historical adoption measures, not an assessment of every deployment’s success or a current regional adoption rate. IMDA’s report highlights.

Adoption and implementation measure different things

A company can adopt a tool for occasional writing while leaving its core processes unchanged. Another can integrate AI into a recurring workflow with data access, quality checks and an accountable owner. Both may appear in an adoption count. Their operating effects, costs and risks can be quite different.

DBS reported more than 2,000 models and over 430 use cases in 2025. These company-reported figures describe deployment breadth, not the value of every model. DBS’s 2025 annual report letter.

The useful lesson is organisational rather than a prescription to match a large bank’s scale. A small company may need only one well-chosen workflow. It still needs reliable inputs, a clear definition of an acceptable result and a way to handle exceptions. Without those elements, a promising trial can become an additional layer of checking rather than a reduction in work.

Evaluate the process that surrounds the model

NIST’s Generative AI Profile provides a voluntary framework for examining risks throughout development and use. Its discussion includes confabulation, data privacy and dependencies across components. These are practical implementation concerns, not merely properties of a model in isolation. NIST’s Generative AI Profile.

For an internal knowledge assistant, evaluation should therefore include whether the correct documents are available, whether permissions are respected and whether answers remain traceable to information employees can inspect. The organisation must also decide what happens when no adequate answer exists. Improving the model alone cannot resolve a missing or contradictory company policy.

Build a repeatable operating capability

A sensible rollout starts with a task that has an identifiable owner and measurable results. The team can compare performance with the existing process, record failures and decide whether the remaining work justifies expansion. This is an analytical recommendation, not a claim that there is one implementation sequence suitable for every organisation.

The divide to watch is the ability to repeat that process. Businesses that can evaluate, integrate and maintain systems may benefit from each new generation of tools more quickly. Wider access remains valuable, but implementation capability determines how much of that access becomes dependable everyday work.

Sources

See something we should correct? Contact the newsroom.