What AI can and can't do right now

A logistics company rolls out generative AI across its entire operation, expecting the tool to rewrite route planning, driver scheduling, and customer communications in one stroke. Within weeks, plausible-sounding outputs clash with real traffic patterns and regulatory limits, forcing manual overrides that erode the promised efficiency gains.

Ng's Central Claim and Argument

Andrew Ng's talk examines what AI can and cannot accomplish at present. He shows AI already reshaping routines in transportation, education, and personal assistants while stressing that these successes remain narrow. Ng develops the point by contrasting concrete examples of functional AI with domains where current systems still fall short, thereby separating demonstrated capability from broader expectations.

Applying the Insights to the Logistics Scenario

Ng's framing clarifies why the company's sweeping rollout encountered friction. By confining generative AI to discrete tasks—such as generating personalized driver training modules drawn from education examples or refining personal-assistant-style alerts for route changes—managers could have matched the technology to proven boundaries rather than demanding enterprise-wide reinvention. This targeted approach would have preserved existing workflows while capturing incremental value.

Guidance for Generative AI Business Strategy

Companies navigating generative AI can therefore prioritize small, well-defined pilots that echo the domains Ng highlights. In transportation, an AI assistant might handle routine schedule queries without attempting full optimization. In education-oriented internal programs, it could produce tailored learning snippets. Such restraint prevents the overextension that turns promising tools into sources of inconsistency.

Lasting Question

How might leaders map their own operations onto AI's current, limited strengths before scaling further?