Human Out of the Loop · Part 5 of 6

November 6, 2026

Is the barn door closed?

AI Guardrails and Access Control: What to Put Around AI Before You Let It Act

Q: What is shadow AI?

The use of AI tools for work that an organization hasn't approved or doesn't know about.

Q: What are AI guardrails?

Rules, permissions and checks placed around AI so it only sees and does what it should.

Q: What should an AI acceptable use policy include?

Approved tools, information never to enter, responsibility for checking output, rules for agents, and how to report mistakes.

Q: What is AI access control?

Deciding which information and systems AI can reach and what it can do there, following least privilege.

Q: Can AI agents go rogue?

The real risk is mistakes made with too much access, managed with limited permissions, approvals and logs.

Q: What is an AI governance framework?

Policies, roles, permissions, technical controls and reviews that make sure AI is used safely.

WRITTEN BY

Quinn Merio

Solutions Architect

Artificial Intelligence, Digital Transformation

QUICK SUMMARY

Rogue agents, leaked data and shadow AI are real risks, and they’re manageable. Here’s what AI guardrails and access control mean in plain language, and a simple framework to put in place before AI touches your systems.

If you’ve been following this series, you’ve seen AI go from answering questions to taking actions. That’s where the biggest fears live: an agent doing something nobody approved, customer information ending up somewhere it shouldn’t, or staff quietly using tools nobody signed off on.

Those fears are reasonable. They’re also manageable, with the same kind of controls you already trust on the road: keys, locks, brakes and rules. In AI, people call these guardrails.

What are AI guardrails?

Guardrails are the rules, permissions and checks you put around AI so it can only see what it should, do what it’s allowed to, and get stopped or corrected when it goes wrong.

In our car analogy, the model is the engine and the harness is the car (part 3). Guardrails are the keys, brakes, speed limiter and door locks. Nobody thinks brakes make a car less useful. They’re the reason you can drive it fast.

The three AI risks worth worrying about

Science-fiction scenarios get the headlines. The everyday risks are more ordinary, and much more likely.

1. Shadow AI: tools nobody approved

What is shadow AI?

Shadow AI is employees using AI tools for work that their organization hasn’t approved, reviewed or even heard of, usually with the best intentions.

It’s happening because people are ahead of their employers. Nearly a third of Canadian workers use standalone, general-purpose AI assistants for work, while most report no formal guidance from their employer at all.

2. Data leakage: what goes in the trunk

What is personally identifiable information (PII)?

Any information that can identify a person: names with contact details, employee records, health or banking information, customer accounts. It’s the information privacy laws care about most.

In part 3 we said context is what you put in the trunk, and it travels with the request. When someone pastes a customer list, an HR file or a contract into a tool that isn’t approved for that information, it has left your control. That’s data leakage, and it’s usually an accident.

3. Oversharing: AI can see what you can see

AI assistants built into your workplace tools generally work with the user’s own permissions. That’s good design, and it’s also why permissions suddenly matter. If old project folders were shared with “everyone” years ago, nobody noticed. An AI assistant that can search everything you have access to will find them in seconds.

Agents raise the stakes. An agent acts with whatever access it’s been given. If it has the keys to every room, a mistake can reach every room.

Most workers have no formal AI support

What Canadian workers say their employer has provided when it comes to AI at work.

No formal AI support64%

A list of approved AI tools19%

Guidance on AI risks, privacy, security and limits16%

Written rules on entering company or customer information16%

Informal briefing at a team, toolbox or safety meeting9%

Formal training on using AI in their role7%

A named person or team to ask for help7%

Base: all workers surveyed (1,100). People could choose more than one, so support types don’t add up. Source: Angus Reid Group for Punchcard, September 2026.

AI access control: give it a valet key

What is AI access control?

Access control decides which information and systems an AI tool or agent can reach, and what it’s allowed to do there: read, write, send, delete or pay.

Many cars come with a valet key. It opens the doors and starts the engine, but it won’t open the trunk or the glove box. That’s exactly the idea to apply to AI, and security people call it least privilege: give every tool and agent the smallest set of keys it needs to do its job, and nothing more.

An agent that chases missing timesheets needs to read the timesheet system and send reminders. It doesn’t need payroll, and it certainly doesn’t need to delete anything.

Can AI agents go rogue?

Not the way they do in movies. Agents don’t develop secret plans. What they do is follow instructions imperfectly, misread a situation or act on bad information, using whatever access they’ve been given. The fix is the same as for any new team member: limit what they can touch, set limits on what they can do alone, and watch the work.

The big platforms are building these controls in. Our colleague Dustin Petersen covered how Microsoft is approaching security and governance for its new workplace agents in Microsoft Build 2026: The Digital Workplace Has Changed.

An AI governance framework, without the bureaucracy

“Governance framework” sounds like a binder nobody reads. It doesn’t have to be. For most organizations it’s five layers, and you can map each one to the car.

Rules of the roadPolicy

An acceptable use policyWhich tools are approved, what information never goes in, and who to ask. One page is enough to start.

Driver’s licencePeople

Training and a named ownerShort, role-specific training, and one person or team people can go to with questions or mistakes.

Valet keyPermissions

Least-privilege accessClean up old sharing before switching AI on, and give each agent only the access its job needs.

Brakes and speed limiterPlatform controls

Guardrails built into the toolsApproved tools only, data protection rules that block sensitive information, and limits on what agents can do alone.

DashcamReview

Logs and human checkpointsA record of what AI did and why, and a person in the loop for anything that matters (part 4).

Five layers of AI governance, from the rules people follow to the record of what happened.

What goes in an AI acceptable use policy?

Example

Our AI acceptable use policy, on one page

Approved tools: use the AI tools on our list for work. Ask before trying anything new.
Never put in: customer or employee personal information, passwords, or anything marked confidential, unless the tool is approved for it.
You own the output: check AI work before it’s used, and in proportion to the stakes.
Agents need a sponsor: any AI that takes actions needs a named owner, limited access and a log.
Mistakes get reported, not hidden: tell the AI lead what happened. No blame for honest errors.

What to do Monday

Find out what’s already in use. Ask your team which AI tools they use for work. Make it an amnesty, not an audit, or you’ll only hear half the answer.
Write the one-page policy. Start with the example above and adjust it to your business.
Check who can see what before you switch AI on. Review sharing on your main SharePoint sites, Teams and drives, starting with HR, finance and customer files.
Name an owner. One person or team responsible for AI tools, questions and mistakes.
Give every agent a valet key and a dashcam. Limited access, limits on what it can do alone, and a log you can read.

The five-second version

Give AI a valet key, not the master key. Clear rules, tidy permissions and a record of what happened let you use AI with confidence, instead of hoping nothing goes wrong.

The series

Human Out of the Loop: AI explained for people who don’t work in tech

Frequently asked questions

What is shadow AI?

Shadow AI is the use of AI tools for work that an organization hasn’t approved or doesn’t know about. It’s usually well-intentioned, but it can expose company and customer information.

What are AI guardrails?

AI guardrails are the rules, permissions and checks placed around AI so it only sees what it should, only does what it’s allowed to, and can be stopped or corrected when it gets something wrong.

What should an AI acceptable use policy include?

At minimum: which AI tools are approved, what information must never be entered, who is responsible for checking AI output, rules for AI that takes actions, and how to report mistakes.

What is AI access control?

AI access control decides which information and systems an AI tool or agent can reach and what it can do there. The best practice is least privilege: only the access needed for the job.

Can AI agents go rogue?

Not in the science-fiction sense. The real risk is an agent making a mistake with more access than it needs. Limiting permissions, requiring approval for important steps and keeping logs manage that risk.

What is an AI governance framework?

A set of policies, roles, permissions, technical controls and reviews that together make sure AI is used safely and responsibly. For most organizations, it can start small and grow.

About the data. Survey conducted by Angus Reid Group for Punchcard, September 2–11, 2026, among 1,100 working Canadians. Employer-support questions were asked of all workers. “No formal support” means none of the listed options; it does not exclude informal help. Results are self-reported. Full results are in The State of AI at Work in Canada.

The acceptable use policy is an illustrative starting point, not legal advice.

Is your Microsoft 365 ready for AI?

Before Copilot or an agent goes looking through your files, make sure it only finds what it should. Talk it through with a specialist and leave with a short list of what to fix first.

Get 90 free minutes with a Microsoft 365 specialist punchcard.io/contact/