New · See every ThinkChat feature in action. Take a look →
ThinkChat
Start free trial

Knowledge AI

When should your AI customer service ask a human? What new FSB and Singapore guidance says

AI 4 min read

On 15 October 2026, the Financial Stability Board (FSB), an international body that monitors and makes recommendations about the global financial system, published its final guidance on how banks and other financial firms should adopt AI. One of its 12 “sound practices” is human oversight, and it is especially direct about AI agents: set clear points where the agent must stop and ask a person.

Singapore’s IMDA (Infocomm Media Development Authority) says something very similar in its Model AI Governance Framework for Agentic AI, version 1.5, published on 20 May 2026.

You don’t need to be a bank to use this. If an AI answers your customers on WhatsApp, Instagram, Facebook Messenger or your website, the same question applies: when should it handle things itself, and when should it hand over to you?

What the FSB said

The FSB’s guidance is written for financial institutions and isn’t binding, but the ideas are simple:

  • AI agents don’t think like people. The report says firms should recognise that “AI agents do not reason like humans and may interpret pre-defined goals in unexpected ways.”
  • Set clear triggers for human approval. It lists “setting clear, risk-based triggers for when an agent must seek human approval, confirmation or review”, for example before approving “financial transactions above a certain value, or taking actions that may cause irreversible harm.”
  • Have an off switch. It describes “kill-switch” mechanisms to “pause, limit, override or remove an agent’s access.”
  • More risk, more oversight. Higher-risk uses “may warrant layered deterministic controls and more intensive human oversight.”

It even gives a customer service example of lighter oversight, which it calls “human-on-the-loop”: an “AI chatbot autonomously handles routine customer inquiries, with human agents alerted to intervene only when sentiment analysis detects frustration or the request volume exceeds the AI’s capability.”

What Singapore’s framework says

IMDA’s framework says organisations “should define significant checkpoints or action boundaries that require human approval, especially before sensitive actions are executed.” Its examples include:

  • High-stakes actions, such as editing sensitive data or actions that may trigger liability.
  • Irreversible actions, such as “permanently deleting data, sending communications, making payments.”
  • Unusual behaviour, such as an agent reaching into a system outside its job.

It also warns about automation bias, where people approving an agent’s requests stop checking properly. It suggests tracking how often humans reject or change the agent’s actions, because “a low rate may signal rubber-stamping behaviours.”

What this means for a small business

Most small businesses won’t write an AI policy document. But you can make the same decisions in an afternoon. The lists and checks below are our suggestions, drawing on the FSB and IMDA ideas above. They are not quotes from, or requirements of, either document. Write down three lists:

  1. AI handles it alone. Routine questions with approved answers: opening hours, prices you’ve published, what’s included, how to book.
  2. AI hands over to a person. Refunds, complaints, angry or upset customers, anything about money or bank details, medical or legal questions, and anything the AI isn’t sure about.
  3. AI never does it. Promise discounts you didn’t approve, change a booking or payment without a person, or make up an answer.

Then check two things every week or so:

  • Read a few handovers. Was the AI right to pass them on? Did it pass on the ones it should have?
  • Read a few chats it handled alone. If you never find anything to fix, make sure you’re actually reading, not just skimming.

And know where the off switch is. We suggest making sure you can pause the AI and reply yourself whenever you need to.

Where ThinkChat fits

At ThinkChat, we think a good AI customer service employee answers from information you approved, stays honest when it doesn’t know, and can hand over to a person. ThinkChat works across multiple chat channels, with WhatsApp as the main one for most Malaysian businesses. Before you choose any AI customer service tool, ours included, ask the vendor to show you exactly when it hands over to a human and how you switch it off.

Soft next step

If you want to hire a ThinkChat AI CS employee, Talk to Lyia on WhatsApp: +60 11-5515 9317.

Sources

Checked 22 October 2026.

← All articles