What artificial intelligence can and can't do right now

When AI Steps In as Business Co-Pilot

A mid-sized logistics firm finds its analysts buried each morning under spreadsheets tracking shipment delays, fuel costs, and inventory mismatches. Instead of spending hours sorting patterns by hand, the team feeds the raw data into an AI tool that flags anomalies and suggests rerouting options. The analysts shift from data crunching to reviewing those suggestions and making final calls on high-stakes routes. Productivity rises, yet the firm still requires human judgment when weather disruptions create cascading effects no model has seen before.

Andrew Ng's talk centers on exactly this kind of measured deployment. He argues that AI currently works best as a co-pilot that automates repetitive tasks and surfaces insights, thereby lifting productivity and sparking innovation. At the same time, he cautions businesses against expecting the technology to handle complex, context-heavy decisions on its own. The emphasis is on identifying narrow processes where clear inputs and outputs allow AI to perform reliably today, rather than promising broad automation across entire workflows.

Applied to the logistics example, Ng's framing clarifies why the morning anomaly detection succeeds: the task is bounded, data-rich, and repetitive. The same logic points to other strong candidates for adoption, such as invoice processing, basic customer-query routing, or initial code-review checks. These processes share the trait of having well-defined success metrics and abundant historical data, allowing AI to act as a reliable assistant without needing to navigate ambiguous strategic trade-offs.

By contrast, the talk underscores that areas involving novel ethical judgments, shifting regulatory landscapes, or multi-stakeholder negotiations remain outside current AI reach. The logistics firm therefore keeps humans in the loop for route decisions during crises, preserving accountability while still harvesting efficiency gains elsewhere.

The practical takeaway is to audit operations for tasks that are repetitive yet data-intensive; those are the processes best positioned for AI adoption right now. Which routine workflows in your own industry already generate the clean, labeled data that would let an AI co-pilot take the first pass?