AI Tools at Work: A Few Guardrails Go a Long Way

From Answering Questions to Taking Actions

The first wave of AI tools answered questions. You typed something in, it wrote something back, and a person decided what to do with it.

The next wave takes actions. These tools, often called AI agents, can send email, edit files, schedule meetings and work across connected apps on your behalf. They can save real time. To do that, they need real access to your systems.

Why Agents Change the Picture

An AI tool that drafts a paragraph can get something wrong, and a person catches it. An AI agent with access to your mailbox and files can act on that mistake before anyone notices.

Agents can also be manipulated. If an agent reads an email or document that contains hidden instructions, it may follow them, for example by forwarding information it shouldn't. The agent isn't malicious; it simply can't always tell a legitimate request from a planted one.

Each agent also adds a new kind of account to your environment, with its own permissions. Many firms don't yet track which AI tools have access to what.

Shadow AI: The More Common Risk

For most firms, the bigger issue is less dramatic. Staff find AI tools on their own, sign up with a work email, connect them to their mailbox or files, and paste in client information to get faster results.

This is often called shadow AI: tools in use that IT doesn't know about and can't manage. It usually comes from good intentions. People want to work faster.

The result, though, can be confidential client or personal information sitting with a vendor no one has reviewed, under terms no one has read, possibly stored outside Canada or used to train the tool.

Attackers Use AI Too

Attackers are using AI to write more convincing phishing emails, imitate voices, and find weaknesses faster. Popular AI software has also been tampered with to steal passwords from the systems that installed it. AI tools deserve the same scrutiny as any other software with access to your data.

Putting Guardrails in Place

• 1.Write a short AI policy. Which tools are approved, what information can go into them, and who reviews AI-generated work. It doesn't need to be long to be useful.
• 2.Offer an approved option. Staff turn to unapproved tools when there's no approved one that does the job. Giving them a good option reduces the temptation.
• 3.Control app connections. Microsoft 365 can require admin approval before staff connect new apps to their email and files.
• 4.Give agents the least access they need. An agent that summarizes documents doesn't need permission to send email.
• 5.Keep a person in the loop. Require human approval before an AI tool sends, deletes or shares anything outside the firm.
• 6.Review what's connected. Check regularly which AI tools have access to your systems, and remove the ones no longer in use.
• 7.Train your people. Help staff understand what's safe to put into AI tools, and why.

The Goal Isn't to Say No

AI tools can make firms more efficient, and staff who use them well are an asset. The aim is simply to make sure the tools are chosen deliberately, with the right access and the right oversight.