Slow down to speed up: Guardrails are key to scaling agentic AI

You hand an AI agent a task and let it run. It pulls the data, works through the steps and completes the job in minutes.

But which source did the AI agent trust? What steps did it take? If it did something wrong, could you pinpoint it?

Unlike a chatbot that simply answers questions, an AI agent can autonomously plan, execute, and verify multi-step tasks, with human oversight during critical decisions.

That is why anyone who deploys an AI agent needs to know how it reached its answer, and why, said Priscilla Chong, managing director for Singapore at Amazon Web Services (AWS). AWS is a subsidiary of US-based e-commerce giant Amazon and provides cloud computing, AI, and generative AI services.

Auditability – a record of every action an agent takes and the data it accessed – is one of four key guardrails organisations must put in place to run AI agents at enterprise scale, she added.

The other three, said Chong, are: 

THE BUSINESS TIMES INSIGHTS

Powering Asean's Future 2026

Find out how enterprises are navigating power constraints and sustainability targets

ACCESS THE INSIGHTS

• Breaking work into small tasks that can each be checked; 
• Setting clear limits on what each agent can do, with humans signing off on consequential decisions; and 
• Giving agents secure access to well-governed data.

Far from slowing firms down, these guardrails are what let them move faster, said Chong.

For example, global payments and financial services firm Stripe built what Chong said was “a handful” of AI agents on AWS to handle financial compliance tasks – and scaled the number to more than 100 in less than a year.

This allowed Stripe’s compliance team to reduce handling time for compliance reviews by 26 per cent, said Chong.

How did they do that? Chong explained that instead of handing AI agents the entire compliance review, Stripe broke it into small sub-tasks. The artificial intelligence agents gather and prepare the information; a human reviewer makes the final call.

Once that pattern worked, Stripe could repeat it with confidence, she said.

Regulated and ready 

Most would assume that a tightly regulated industry like financial services would face more barriers to adopt agentic AI. Chong argued the sector is one of the most ready for it.

Financial institutions like banks run on high-volume, well-defined work, she explained, comprising tasks where an agent can do the groundwork while a human makes the call.

These include verifying documents to onboard customers, assembling case files for financial crime reviews, investigating payments and triaging disputes.

Regulation actually gives the sector an advantage, Chong said. Financial institutions already have mature frameworks for risk, governance and internal audit – the very guardrails agents need.

In Singapore, regulators have laid the groundwork for the adoption of agentic AI. The Monetary Authority of Singapore (MAS) published Safeguards for Agentic Finance at Runtime (SAFR) in July 2026.

Developed in collaboration with financial institutions and fintech firms, this framework sets out four safeguards: 

• Identity: Verifying that each agent is who it claims to be.
• Permissions: A clear rulebook of what each agent may do, and within what limits.
• Checks before action: Assessing every action an agent proposes against its permissions.
• Records: A tamper-proof log of every decision.

Even with these safeguards in place, a human must stay in the loop, said Chong.

This is especially important in financial services, she added, noting the sensitive nature of the transactions and information involved.

“An agent helps you expand the surface of thinking; then a human (comes) in to make the decision,” Chong said.

This was produced in partnership with the Monetary Authority of Singapore and the Global Finance & Technology Network.

For more stories, go to bt.sg/sff26