Your Staff's AI Agent Is Asking for Permission: Singapore's New Advisory Makes It Your Call
On 6 November the Cyber Security Agency of Singapore, GovTech and IMDA published a joint advisory on using personal AI agents safely. These are not chatbots. They are assistants that connect to your email, calendar and apps, and then act: sending messages, making bookings, filling in forms, sometimes carrying on in the background while you do something else. The advisory names several, including Dots, Hermes, Grok Bot, OpenClaw and Meta's Muse, which launched in Singapore on 27 October with free core features and the ability to connect to Gmail and Google Calendar.
The advice is sensible and written for consumers. But one line in it is aimed squarely at every employer in the country. On using an agent with a work account, staff are told to "check if your employer allows it". That is the only workplace guidance in the document. In most SMEs, nobody has decided what the answer is.
The question has already been asked, just not of you
The way these agents get into a business is not dramatic. An employee installs one on their phone, it asks to connect to their email and calendar, and they tap allow. If that account is their work account, the agent now has standing access to client correspondence, meeting invites, shared documents and whatever else the account can reach. No procurement, no IT ticket, no vendor review.
That is the gap. Most SMEs have some form of AI policy by now, and it is usually about what staff may paste into a chatbot. An agent is a different risk. It doesn't just receive data; it holds a key and uses it. The advisory itself spells out why that matters. Agents can make mistakes or take actions the user didn't intend. Some keep running when nobody is watching. Many keep persistent memory. And they can be manipulated by instructions hidden in web pages, emails, calendar invites, documents or images. For a business, that last one is the important one. Your staff's inboxes receive content from strangers all day.
"The user is responsible" doesn't help the employer much
The advisory is clear that people remain responsible for what their agents do on their behalf, and may be held liable depending on the circumstances. That is a reasonable message to consumers. It is cold comfort to an employer. If an agent connected to a sales manager's mailbox forwards a client list after reading a poisoned email, the fact that the employee is "responsible" doesn't undo the disclosure, and it won't be the employee's name in the incident report the client reads.
It also sits uneasily with Singapore's own position on agents deployed by organisations. IMDA's Model AI Governance Framework for Agentic AI, launched in January and updated in May, rests on the principle that humans are ultimately accountable, and the natural reading is that the organisation letting an agent act carries that accountability. An employee's personal agent working inside a company account falls between the two documents: too personal for the framework, too corporate for the consumer advisory.
Banning it is the wrong instinct
The reflex will be a one-line rule: no AI agents on work accounts. I'd push back on that. Bans on AI tools have a poor track record. Deloitte's October research on personal agents at work reports that 49% of respondents had used unapproved AI tools, and that senior leaders were more likely to do so than their staff. A ban without enforcement just means the agent is still connected and nobody asks about it.
Deloitte's suggestion is more useful: keep personal agents away from sensitive data and customer-facing work, and pilot them under oversight, which may reduce the pull towards unapproved tools. In other words, decide where they are allowed, make that path easy, and close the doors that matter.
Closing those doors is mostly configuration you already own. Google Workspace lets an administrator see which third-party apps users have already let into company data, and mark each one as trusted, limited or blocked, or block unconfigured apps altogether. Microsoft 365 lets you restrict which apps users can consent to on their own, and Microsoft's own guidance recommends allowing self-consent only for apps from verified publishers. Note the catch in Microsoft's documentation: changing the setting affects future consents, not the grants that already exist. You still have to go and look.
What an SME should do before the next allow button gets tapped
First, find out what is already connected. Pull the list of third-party apps with access to your email and file accounts and read it properly. Anything that describes itself as an assistant or an agent, or that has mail send or full drive access, deserves a conversation with the person who connected it.
Second, answer the question the advisory told your staff to ask. Write it down in plain language: which agents, if any, may connect to which work accounts, with what permissions. The advisory's own consumer advice makes a good employer baseline. Start read-only. Require approval for sending, deleting, sharing and posting, and prefer one-time approval over "always allow". Set those limits in the app's settings, not in a chat instruction the agent may not follow. Keep agents away from finance, HR and anything holding customer identity documents.
Third, protect the authentication path. The advisory tells consumers to keep one-time passwords in a mailbox the agent cannot reach. For a business, that means the accounts used for password resets and MFA codes should never be the ones an agent can read. An agent that can read the inbox where reset links arrive, or whoever manages to steer it, is one step away from resetting passwords.
Fourth, add agents to offboarding and incident response. When someone leaves, revoke the app grants on their account, not just the password. And if something goes wrong, the advisory's own checklist works for a company account too: stop the agent, disconnect it from both sides, change passwords, sign out all sessions, and look for new forwarding rules, unexpected messages and changed recovery details.
Finally, check where the data goes. The advisory notes that Muse may use interactions for model training by default, and that Meta does not review custom connectors. If a client's correspondence passes through a personal agent, you should know whose terms it is governed by.
None of this needs a new budget line. It needs someone to decide, before the next employee taps allow, what the answer to "does my employer allow it?" actually is.
Belian Advisory helps SMEs write an AI policy that covers agents rather than just chatbots, and audit which apps already hold the keys to their accounts; if your staff are asking whether they can connect one, get in touch.