Nudge users to take actions and make better decisions, so they can focus on what is more valuable. This is what some businesses want artificial intelligence (AI) agents to do, as they deploy agentic capabilities across workflows.
Grab, for instance, is aiming to have AI agents automatically nudge customers to book their ride earlier, if rain is forecast for the day. It will allow them to avoid being unable to secure a ride, amidst an anticipated flood of booking requests whenever it rains.
This means powering the agents with the necessary information, such as real-time weather updates, and predictive capabilities. They then can ingest the data and provide insights that drive actionable decisions, said Jerry Lim, group managing director at Grab.
The Asian tech company, which provides a range of consumer services including rides, food deliveries, and banking, wants to apply AI across the entire customer journey.
This requires optimisation and “balance” in the backend, so its network comprising drivers and riders are in the right place at the right time, said Lim, who was speaking at Salesforce’s Agentforce World Tour conference in Singapore.
Grab is doing what it can to position its fleet in the right areas based on forecasting and AI, and nudging its suppliers to the right areas, he noted.
The aim here, ultimately, is to find the optimal balance between supply and demand, and establish a rider price that also is fair to customers, Lim explained.
The backend infrastructure, too, needs to be built on trust and safety, he said, adding that Grab does this from the ground up, from design.
The company uses AI to manage user identities and payments, so it can detect anomalies and resolve incidents quickly or, when needed, escalate them to human administrators for intervention, he noted.
Tapping AI agents to stand out from the crowd
It creates a robust model, tapping Grab’s “trust signals” and technology, so it can execute more quickly on the ground and in a more localised way, he said.
It also helps ensure its services cannot be easily replicated by generic, competing services in its ecosystem, Lim said.
As adoption of agentic AI grows, however, some companies have expressed concerns about what this might mean for their infrastructure.
In particular, 94% are worried that AI sprawl is driving complexity, technical debt, and security risk, according to an April 2026 report by OutSystems.
In spite of such concerns, few have established a centralised approach to agentic AI governance, the report found.
This suggests most are using agents across fragmented environments, noted OutSystems, which polled almost 1,900 IT leaders worldwide, including 527 in three Asia-Pacific markets — Japan, India, and Australia.
When organisations scale their agentic implementation, the challenge shifts from deployment to coordination, said Gavin Barfield, Salesforce’s vice president and CTO of solutions for Asean.
He urged enterprises to move beyond basic setups to build observability and orchestration capabilities, needed to manage complex workflows.
Unify for clear oversight, governance
“Without this oversight, adding more agents introduces new points of failure instead of building operational leverage,” Barfield said.
He further noted the need for a strong, trusted data foundation and unified platform for organisations to scale effectively, while managing complexity and maintaining costs.
Unifying both structured and unstructured data ensures their AI agents have governed, accurate, and real-time context — instead of disconnected data silos — to operate reliably across channels and systems, Barfield said.
Security also should be built in by design, ensuring that AI agents are managed with the same rigour as human employees, he said.
This includes zero data retention, toxicity detection, role-based access controls, and clearly defined guardrails for what AI agents can and cannot do, he added.
And it seems that, like Grab, other businesses in the region are looking to tap AI agents for operational and productivity gains.
Just 3% of employees in Asean do not expect to use AI agents, while 75% have interacted with or already are using agentic agents, revealed a Salesforce report, which polled 4,062 respondents across the region, namely, Singapore, Indonesia, Thailand, and the Philippines. The study was conducted by YouGov.
Asked how AI will impact their work in future, 35% of the Asean respondents expect to use AI agents to automate some tasks and augment others, the report noted.
Another 48% say AI agents augment work by providing quick access to information, cutting the need for extensive research.
Some 45% say agents assist with writing and communications, saving them time and effort, while 43% credit AI agents for helping them brainstorm.
In addition, 39% expect to use AI agents to enhance their performance at work, while 16% believe the agents power simple automation, the study revealed.
Some 28% anticipate human-to-AI collaboration skills to be crucial in future, with 74% agreeing their job will change at least moderately as tasks are shared with AI agents.
“To succeed in the AI era, companies should see AI not just as a technology investment, but as a people transformation,” said Paul Carvouni, Salesforce’ Asean senior vice president and general manager. “[The] workforce is ready, but it is up to organisations to provide the secure, enterprise-grade frameworks and skills support that turns personal use of AI into a coordinated engine for growth and innovation in the agentic enterprise.”
Grab currently uses its own AI platform Jarvis alongside Salesforce’s AI agent platform, Agentforce, to develop and orchestrate the AI agents it deploys.
The goal is to equip its merchants and sellers with the right tools and the right context, and push the right advice to customers, Lim said.
Merchants that work with Grab can tap the company’s self-service portal to manage their business and access data insights, including business reports and recommendations for next actions to take.
The AI-powered insights can guide sellers through complex negotiations, especially when multiple products are involved, and generate the appropriate pitches, Lim said.
Grab also uses AI to coach its sellers, including helping them manage live calls or customer engagement, he added. For example, it taps AI-powered role-playing and avatars to test the sellers’ capabilities, and identify gaps in their sales path.
Where AI agents can help
AI agents can be triggered to help sellers better manage customer engagements, cutting down on the time they would have spent dealing with mundane queries, said Rene Hefner, Grab’s regional head of sales enablement.
AI agents provide support digesting information and preparing sales representatives for client meetings, Hefner said during a session at the conference.
He added that they can look up past meetings and interactions with a customer, summarise these and pull up the most recent engagement, and recommend actions to prepare for the next meeting with the customer.
Asked what job skills they will need in the agentic era, 38% of respondents in the Salesforce study cite data analysis and interpretation, while 35% point to creative thinking and 34% highlight problem solving.
To feel more confident in their use of AI agents at work, 43% highlight the need to know what actions the agents took and why, the study revealed.
Another 42% point to easy access to approved, quality tools, while 42% cite the need to understand the skills they have to develop.
According to the OutSystems study, 96% of respondents already are using AI agents in some capacity, while 97% are exploring system-wide agentic AI strategies.
Some 49% describe their agentic AI capabilities as advanced or expert, with 31% noting that AI already is integral to their development practices.
As agentic AI deployments grow, though, organisations will need to be careful that their costs do not spike.
Barfield advised companies to track how much work actually gets done by measuring Agentic Work Units (AWUs), or discrete tasks an agent accomplishes.
He noted that AWU growth across Salesforce’s platform now outpaces token consumption, which means more work is done per token.
He attributed this efficiency to tapping hybrid architectures where systems dynamically route tasks.
“By letting an agent decide which steps require the creative, probabilistic reasoning of an LLM (large language model), and which can be handled by cheaper, deterministic workflows, you simultaneously lower the cost per task and increase operational predictability,” Barfield explained.
Unlike deterministic software, agents are probabilistic and require entirely different tools, he added.
“An agent-first architecture must be open by design, where every platform capability is accessible to agents through MCP (Model Complex Protocol) servers, Command Line Interfaces, and APIs (application programming interfaces),” he said. “This allows them to operate autonomously across workflows, without being bottlenecked by a single user interface.”
