the story
Siew Ling runs a building-supplies distributor in Kaki Bukit: 22 staff, four sales people, about 300 trade customers. She and her firm are made up. The problems are ones I hear about often.
Week 1: a price list in a chatbot
A sales executive shows her a quote he wrote in two minutes. He had pasted the customer’s past orders and the firm’s trade price list into a free chatbot on his phone. Two others do the same. Nobody asked, because nobody had said anything.
She doesn’t ban it. She writes one page: which AI tools staff may use, what never goes in (customer details, trade prices, anything under contract), and who to ask. She names herself and her operations manager as the two people who answer for AI use, and puts AI on the monthly management meeting.
UsedSection 2 of ISAGO, IMDA and PDPC’s free self-check guide: who is responsible, and does everyone know the rules.
Week 2: the agent demo
A software vendor demonstrates an AI agent. It reads email enquiries, checks stock, drafts quotes, sends them, and chases late payments. A chatbot only answers. An agent acts, in the firm’s name.
She asks the vendor to walk her through what the agent is made of. IMDA’s guide for agentic AI, updated in May 2026, lists eight parts. She writes the vendor’s answer under each one.
- ModelThe AI that reads, plans and writes.
- InstructionsWhat it is told to do and not do.
- MemoryWhat it keeps between conversations, including customer details.
- Planning and reasoningHow it breaks a job into steps.
- ToolsWhat it can open and change: email, stock system, accounts.
- ProtocolsHow it connects to other software and other agents.
- ControlsAccess limits, guardrails and human approvals.
- Logging and monitoringA record of what it did and why.
Two answers worry her. The agent would get full access to the accounts system, and the vendor can’t show her where its actions are recorded.
UsedSection 1 of IMDA’s Model AI Governance Framework for Agentic AI (version 1.5).
Week 3: where a person says yes
She sorts what the agent would do by two questions: how bad is a mistake, and can it be undone? Checking stock and drafting a quote: low harm, easy to undo, so the agent goes ahead. Sending a quote to a customer, giving a discount, writing off a payment or deleting a record: a person approves first. If the approver is away, the action waits. It doesn’t go ahead by default.
She asks the vendor to give the agent read-only access to the accounts system, and to block the approval-needing actions in the software itself, not just tell the agent in its instructions. Instructions can be ignored or forgotten. A blocked button can’t.
UsedSection 3 of ISAGO on the level of human involvement, and dimension 1 of IMDA’s agentic guide on bounding the risks upfront.
Weeks 4 and 5: two people, one log
The agent starts with two sales staff and one product range. Every action goes into a log nobody can edit. On Fridays the operations manager reads ten drafts at random.
At the end of the month she looks at one number: how often staff changed or rejected a draft before approving it. Early on it is about one in four. If it ever drops to zero, either the agent has become perfect or nobody is reading. IMDA’s May 2026 update suggests tracking this override rate and how fast people approve.
UsedSection 4 of ISAGO on operations, and dimension 3 of IMDA’s agentic guide on testing, gradual rollout and monitoring.
Week 6: telling customers
Quotes now carry one line: drafted with AI, checked by a named staff member, with that person’s number. Customers who reply get a person.
The vendor then asks to use the firm’s customer emails to improve its model. She says no for now. PDPC’s July 2026 guidelines on personal data in generative AI say that where a firm relies on consent to train AI on personal data, it must say so plainly. “To improve our services” is not enough. She would rather not start that conversation with 300 customers.
UsedSection 5 of ISAGO on stakeholder communication, and PDPC’s advisory guidelines on personal data in generative AI.
And the junior staff
Writing quotes was how new sales staff learned the products and the prices. If the agent writes them all, nobody learns. So each junior writes a few quotes a month by hand, and a senior checks them. If the agent stops working one day, someone still knows how.
UsedDimension 4 of IMDA’s agentic guide, on keeping skills and business continuity.