Can an agent leak customer data?

Yes, if it is given more access than it needs, if it sends data to a tool whose terms allow reuse, or if it can be tricked by instructions hidden in content it reads. Limit what data the agent can see, use business tiers with clear data-use terms, and have a person review anything that leaves the business.

Who is responsible when an agent makes a mistake?

Your business is. An agent acts on your behalf, and Canada's privacy regulators are clear that organizations remain accountable for how they use AI with personal information. That is why the first agents should draft for a person to approve rather than act on their own.

Do I need a policy before using agents?

A short one, yes. One page covering which tools are approved, what data may and may not go into them, who reviews agent outputs, and how to report a problem is enough for a small team. It is also what a client will ask to see.

What are the risks of using AI agents in my business, and how do I reduce them?

The seven main risks of using AI agents in a small business are privacy breaches, wrong or invented outputs, prompt injection, excess access, silent failure, runaway costs, and vendor lock-in. Each is reduced the same way: keep the agent’s job narrow, give it the least access it needs, keep a person approving anything that leaves the business, and monitor it after launch.

None of these is a reason to avoid agents. They are reasons to start with one narrow, reviewed workflow instead of many unsupervised ones.

Updated 2026-09-24 by Esteban Valencia, EV Advisory. Written for founder-led businesses in Canada and the United States. This page is general information, not legal advice.

What are the main risks?

Risk	What can go wrong	How to reduce it
Privacy and data exposure	Personal information is sent to a tool, stored, or shown to someone it should not be	Minimise the data the agent sees, use business tiers with clear data-use terms, document what goes where
Wrong or invented outputs	The agent states something false with confidence, or misreads the input	Keep a person approving outputs, test on real past examples, restrict the agent to your own source material
Prompt injection	Instructions hidden in an email, web page, or document trick the agent into doing something else	Never let an agent that reads outside content also take unreviewed actions; separate reading from acting
Excess access	The agent can change or delete far more than its job needs	Least-privilege access, separate accounts for the agent, no admin rights
Silent failure	A connected tool changes and the agent quietly stops working or produces junk	Monitor volume and error rates, alert when either moves unexpectedly, keep a named owner
Runaway costs	Usage fees grow with volume or loops	Set spending limits with the model provider, review monthly cost against value
Lock-in	The agent runs in a provider’s account and cannot be moved	Run agents in accounts you own, keep configuration and documentation in your hands

How do Canadian privacy rules apply to AI agents?

If your business collects, uses, or discloses personal information in the course of commercial activity, federal privacy law (PIPEDA) generally applies, unless a substantially similar provincial law applies instead, as in Quebec, Alberta, and British Columbia for private-sector activity within those provinces. Using an AI agent does not change those obligations.

The Office of the Privacy Commissioner of Canada’s guidance on AI, privacy, and your business sets out what it expects when businesses use AI with personal information, including:

Legal authority and meaningful consent. Establish legal authority for collecting and using personal information, and when relying on consent, make sure it is valid and meaningful.
Transparency. Be open about how information is used and the privacy risks involved, including when AI generated content or made a decision.
Safeguards. Protect privacy rights and limit the sharing of sensitive or confidential information.
Privacy by design. Build privacy into tools from the start.

Canada’s federal, provincial, and territorial privacy regulators have also published principles for responsible, trustworthy and privacy-protective generative AI, which apply to organizations that use generative AI tools as well as those that build them.

Serving clients in the United States adds state privacy laws that vary by state. Ask a qualified advisor about your specific situation.

How do you keep a person in the loop?

Decide, for each agent, which of three levels it runs at:

Draft only. The agent prepares; a person reviews and sends. The right level for every first agent.
Act on routine, escalate exceptions. The agent handles clearly routine cases and hands anything unusual to a named person. Earned after weeks of accurate drafts.
Act, with sampling. The agent acts on its own and a person reviews a sample. Only for low-risk, internal, reversible tasks.

Write the level into the agent’s documentation, and make the escalation path concrete: who gets the exception, how, and how fast.

What should you monitor after launch?

Correction rate. How often a person changes the agent’s output. Rising means something changed.
Volume. A sudden drop often means a silent failure upstream.
Exceptions and escalations. What the agent hands to people, and why.
Cost per run. Against the value of the work it does.
Access. A quarterly check that the agent still has only the access it needs, and that people who left no longer control it.

Sources

Office of the Privacy Commissioner of Canada, AI, privacy, and your business.
Office of the Privacy Commissioner of Canada, Principles for responsible, trustworthy and privacy-protective generative AI technologies.
Innovation, Science and Economic Development Canada, Toolkit for SMEs deploying artificial intelligence.

Next step

EV Advisory builds first agents at the draft-only level, with least-privilege access in accounts you own and a runbook your team can follow. See how on the AI agent implementation page, or book a free 20-minute Live Audit to talk through the risks for your own workflow.