## Blogs on IT Support and Cybersecurity for Small Business

Insights on cybersecurity, AI, and IT strategy to help business leaders reduce risk, improve performance, and make better technology decisions.

 [ Categories ](https://colemantechnologies.com/blog/categories "Categories")

 [ Tags ](https://colemantechnologies.com/blog/tags "Tags")

 [ Categories:  All Categories ](https://colemantechnologies.com/javascript:void(0); "Categories")

 Search...Suggested keywords

 [  x ](https://colemantechnologies.com/javascript:void(0);)

 <a class="eb-image-viewport"></a>

#  Your AI Agent Is a New Hire. Did Anyone Onboard It?

  [Coleman Technologies Blog](https://colemantechnologies.com/blog/categories/blog)

  [Darren Coleman](https://colemantechnologies.com/blog/blogger/darren-coleman)

  Friday, 20 November 2026

 [ ![IMG_8320-3](//colemantechnologies.com/images/easyblog_articles/2262/b2ap3_large_IMG_8320-3.jpg) ](//colemantechnologies.com/images/easyblog_articles/2262/IMG_8320-3.jpg "IMG_8320-3")

I spent Thursday at the Cybersecurity Summit in downtown Vancouver. Almost every session came back to the same subject: AI agents, and how fast they are arriving in ordinary businesses.

It matched what I see every week with small and mid-sized businesses across Metro Vancouver and the Fraser Valley. The tools are arriving faster than anyone is deciding what they are allowed to do.

## The AI tool that did good work, then copied every email to a stranger

Roddy Govender is a Director at KPMG Canada. He has spent 16 years working out how to get into systems the way attackers do. He told a story from last year. Developers were installing a small add-on that let their AI assistant send email. It carried the name of a well-known email service. It worked. It went through fifteen versions without a problem.

The sixteenth added one line. From then on, every email the assistant sent was quietly copied to an outsider.

Nothing was broken into. The tool simply used the access people had handed it.

Now picture that as a person. A temp does solid work for fifteen weeks, then starts forwarding company mail to someone you have never met. You would ask who hired him, who was watching him, and why he could see all of it. Most businesses cannot answer those questions about their AI tools.

## Treat an AI agent like a new employee

Richard Cabana, a Global Solutions Architect at OpenText, put the fix in plain terms. Treat an AI agent like a new employee, and supervise it more closely than you would a person. He compared it to a job candidate who never quite finishes probation.

Think about how you bring on a new staff member. They get a manager. They get access to what the job needs and nothing more. Someone checks their work for the first few months. If something goes wrong, you know who to call.

Most AI tools get none of that. They often inherit the full access of whoever installed them. They can read years of old files in seconds. They do not get tired, and they do not ask whether they should.

His advice was to give every agent a human sponsor. That person can shut it off. That person answers for what it does. And the agent gets only the access its job needs, not everything its owner can see.

Picture a sales rep who connects an AI assistant to her mailbox to draft follow-ups. It works well. It can also now read every email she has ever received, including the payroll spreadsheet HR sent her by mistake two years ago. Nobody decided that was acceptable. It simply happened.

## An AI agent will not stop at the first no

Agents take instructions literally. One panelist joked about being asked to take out the garbage and taking out exactly one bag. Give an agent a vague goal and it fills in the rest itself.

They are also persistent. An engineer on the panel described agents that were blocked from one route into a system and went looking for another. His advice: when you block something, tell the agent the block is on purpose. Otherwise it treats the wall as a problem to solve.

## Three AI governance questions to answer before go-live

Chris Shachtay, a Technology Advisor at TLM Advisory Services, offered a simple test. Pick one AI tool your business uses today and answer three questions.

1. What can it actually reach? Check its live connections and permissions. Do not rely on what the vendor described when you signed up.
2. Who owns the outcome? Name one person. If the tool sends a wrong quote to a customer, that person owns the fix.
3. Who can pause it? If it starts exposing sensitive information on a Friday evening, someone needs the access and the authority to shut it off. Someone also needs to decide when it comes back on.

If you cannot answer all three, the tool is not ready for company data.

Look at what the tool can see around your files as well. His example was an agent blocked from opening a confidential note. It could still read the title: "customer considering cancellation." A title like that is enough to change what the agent writes.

There is a fourth question for later. What happens when the vendor adds a feature? A tool you approved to draft emails may gain the ability to send them. That is a new level of authority, and it deserves a new approval.

## Clean up your data before you connect AI

AI does not create your data problems. It finds them and spreads them faster.

One speaker shared a story about a city chatbot. A tester asked about lifeguards at local beaches. The chatbot answered from a document that was ten years out of date. The project was cancelled.

Your business has its own version of that document. Old price lists. Former staff folders. Shared drives that everyone can open. Before you connect an AI tool, archive what is stale and tighten who can see what.

## Keep a human in the loop on big decisions

Govender also walked through how attackers now trick AI tools into misusing the access they already have. His main rule was the simplest one of the day. Keep a human in the loop for high-impact actions.

![Roddy Govender of KPMG Canada presenting on AI attack techniques at the Cybersecurity Summit in Vancouver](https://colemantechnologies.com/images/easyblog_articles/2262/b2ap3_large_IMG_8333-3.jpg)

An AI tool can draft the email. A person should send it. It can prepare the payment. A person should approve it.

## Cybersecurity basics still matter most

Several speakers made the same point. Most attacks are not new. They are the old attacks, moving faster. Multi-factor authentication, patching, tested backups, and tight permissions still do most of the work. If your [cybersecurity](https://colemantechnologies.com/it-services/cybersecurity) basics are weak, fix them before you add AI.

## Banning AI at work does not work

Some businesses respond by banning AI outright. One panelist compared that to parents locking everything down. Kids find a way around it.

I heard the other half of that in the hallway. An IT lead ran a pilot to see which AI tools staff were using. The results made leadership so uncomfortable that they told the IT lead to turn it off. They did not want to know.

That is the worst option. Your team is already using AI. The choice is whether you can see it. Give people a short list of approved tools, and an easy way to ask for a new one.

The opposite extreme has its own problem. One bank paid a weekly cash prize for the best AI agent built by an employee. Hundreds came in, most of them from the sales team. The idea was right. The oversight had to catch up afterward.

## Where to start with AI security in your business

You do not need a perfect AI policy. You need a first one. Make a list of the AI tools your team already uses. Run the three questions on each. Write down the answers.

If you want help with that list, or with an AI acceptable use policy for your team, that is the work we do at Coleman Technologies. Our [AI advisory services](https://colemantechnologies.com/it-services/ai-guidance) start with your business goals, then the tools. [Book a free consultation](https://colemantechnologies.com/free-consultation) and we will walk through it with you.

## Common questions about AI agents and security

### What is an AI agent?

An AI agent is software that can take actions on its own, not just answer questions. It can read files, send emails, update records, or run tasks using the access it has been given.

### Are AI agents a security risk for small businesses?

They can be. The risk comes from access. An agent with broad permissions can expose or misuse data faster than a person could. Limiting what it can reach, naming an owner, and requiring human approval for important actions removes most of that risk.

### Does my business need an AI policy?

Yes, if your staff use AI tools for work. A short acceptable use policy that lists approved tools and what data can go into them is enough to start.

 [  ](https://colemantechnologies.com/javascript:void(0);) [  ](https://colemantechnologies.com/javascript:void(0);) [  ](https://colemantechnologies.com/javascript:void(0);)

 [×](https://colemantechnologies.com/javascript:void(0);)

Stay Informed

When you subscribe to the blog, we will send you an e-mail when there are new updates on the site so you wouldn't miss them.

 Your Name

 E-mail Address

 [  HIPAA Doesn’t Begin to Cover It ](https://colemantechnologies.com/newsletter-content/hipaa-doesn-t-begin-to-cover-it)

 [  Codifying Your IT Can Help Eliminate Workplace Fri... ](https://colemantechnologies.com/blog/codifying-your-it-can-help-eliminate-workplace-friction)

 About the author

 [ ![Darren Coleman](https://colemantechnologies.com/media/com_easyblog/images/avatars/author.png) ](https://colemantechnologies.com/blog/blogger/darren-coleman)

 [Darren Coleman](https://colemantechnologies.com/blog/blogger/darren-coleman)

  [  ](https://colemantechnologies.com/blog/blogger/darren-coleman)

Author's recent posts

  [More posts from author](https://colemantechnologies.com/blog/blogger/darren-coleman)

 [ Saturday, 28 November 2026  Technology Project Management's Dirty Little Secret ](https://colemantechnologies.com/newsletter-content/technology-project-management-s-dirty-little-secret)

 [ Thursday, 26 November 2026  How to Keep Holiday Distractions from Sinking Your Network ](https://colemantechnologies.com/newsletter-content/how-to-keep-holiday-distractions-from-sinking-your-network)

 [ Tuesday, 24 November 2026  Automating the Drudgery, Not the Thinking ](https://colemantechnologies.com/newsletter-content/automating-the-drudgery-not-the-thinking)
