Moving from AI assistance to autonomous operations: Lessons gleaned from PSA BDP at the Data & AI Summit Singapore 2026

Flow of digital information. Global connection concept. Technology futuristic background. Big data visualization.

In global supply chains, the value of identifying a disruption depends on how quickly that information reaches the people who need to respond. For PSA BDP, this made disruption monitoring a practical starting point for moving AI beyond assistance and into day-to-day operations.

At the Data & AI Summit Singapore on 23 September 2026, Gaurav Shukul, Director of Digital Trade at PSA BDP, a member of PSA International, shared that experience during a panel entitled “When Agents Become the Operating Model: How Leaders are Re-Architecting the Enterprise.” Drawing on disruption monitoring and customer shipment enquiries, Gaurav explored what it takes to give AI more responsibility while simulataneously maintaining reliability and control.

The turning point is a change in the workflow

That journey is reflected in Risk Monitor, PSA BDP’s risk monitoring digital product, which has evolved from traditional machine learning through foundation models to reasoning models. Each stage has expanded the role AI can play in identifying and publishing disruption events.

With reasoning models, accuracy levels are now very high. As Gaurav shared, around 98% of disruption events in Risk Monitor are created by AI without human involvement.

The turning point came when events could be published directly to the platform without waiting for individual approval. In an operation spanning time zones, this removes the need for an event identified in one region to wait for a reviewer elsewhere to come online, thereby delaying the visibility altogether. AI becomes part of the operational workflow, with responsibility for completing a defined task.

Better model performance therefore creates an opportunity to rethink how work gets done. The question becomes where human involvement adds value and where a task can proceed independently.

Building trust through relevant and consistent answers

When customers ask an AI agent for information, they need an answer they can use to make a decision. Gaurav’s example of shipment enquiries highlights what this requires: understanding the question, drawing on the appropriate data, and presenting the answer clearly.

Consider a customer asking when a shipment will arrive. The carrier’s latest estimate, the original booking estimate and a predictive arrival time each offer a different perspective. A useful answer needs to establish which one the customer is asking for and make clear what the date represents.

The same applies to follow-up questions. A request such as “what about last week?” depends on the earlier conversation. Maintaining that context helps the AI agent provide consistent, relevant answers. It also needs to recognize requests outside of its capabilities and make those limits clear.

Use the right approach for each task

Building that reliability takes more than detailed instructions. As Gaurav explained, adding rules can eventually create contradictions. A more effective approach is to decide which tasks require exact results and place them outside the model’s discretion.

PSA BDP uses defined software rules and data functions for tasks that require precise results, such as calculations, identifying ports and interpreting dates. These tasks follow established logic, while the AI model handles language and communicates the results. This gives each part of the system a clear role in producing a reliable answer.

The AI agent accesses data through approved functions and uses standard responses for requests it cannot support. Customer feedback helps identify scenarios to include in ongoing testing, so the team can check that improvements hold as the system evolves. Reliability needs to be demonstrated consistently across repeated interactions.

Expand authority in deliberate steps

The same care applies to what an agent is authorized to do. Access to information and authority to change it are separate decisions. The AI agent can retrieve information that the customer is entitled to access but is largely read-only. Its actions are limited to reversible steps, such as setting a favorite or subscribing to an alert, with confirmation provided to the customer.

Looking ahead, Gaurav sees scope to move from identifying disruptions towards recommending recovery options and opening rebooking conversations with carriers and forwarders. These remain as future steps, with additional confirmation as the agent moves towards actions carrying greater operational consequences.

Across both examples, human judgement is still central to deciding which responsibilities can be delegated and what controls they require. Direct publication may be appropriate for disruption events, while changes to a shipment demand a different level of oversight.

Moving from AI assistance to autonomous operations: Lessons gleaned from PSA BDP at the Data & AI Summit Singapore 2026

The turning point is a change in the workflow

Building trust through relevant and consistent answers

More from the BDP Blog