Answers, not hype

Straight answers to the questions we hear most

Short, plain answers with a link to the official source for each. Checked against Microsoft and Canadian regulators, with the date we last reviewed them.

Building agents

Should we build in Agent Builder or Copilot Studio?

Start with Agent Builder when one person or a small team needs a quick helper that answers from content they already have, like SharePoint files or email. It lives inside Microsoft 365 Copilot and needs no code.

Move to Copilot Studio when the agent needs workflows, connections to other systems, publishing beyond Microsoft 365, or proper development, test and production stages. You don’t have to start over: Microsoft supports copying an Agent Builder agent into Copilot Studio.

Our rule of thumb: if the agent will be used by more than one team, touches business systems, or has to pass an audit, build it in Copilot Studio from day one.

Source: Microsoft Learn — Choose between Agent Builder and Copilot Studio

Last reviewed November 21, 2026

Cost

How is Copilot Studio billed?

Agents use Copilot Credits. You either buy credits in advance as prepaid capacity, or pay as you go through an Azure subscription (about US$0.01 per credit as of November 2026). Each answer, action or workflow step uses a set number of credits, so cost depends on what the agent does, not how many people have access.

Two things keep costs predictable: send fixed, rule-based steps to workflows instead of the AI model, and set a monthly limit on every agent (see the next answer). We price every step before an agent goes live, so finance sees the number first.

Source: Microsoft Learn — Manage Copilot Studio capacity

Last reviewed November 21, 2026

Cost

Can we stop an agent from overspending?

Yes. In the Power Platform admin center you can give each agent a monthly credit limit, get an email as it approaches that limit, and have the agent switch itself off when it reaches it. Turning it back on, or raising the limit, then becomes a deliberate decision.

Two caveats: limits aren’t applied instantly, so leave some margin, and an agent’s limit doesn’t cover the agent flows it calls. Pay-as-you-go Azure subscriptions also have no hard spending cap of their own, so per-agent limits plus an Azure budget alert are the safety net.

Source: Microsoft Learn — Per-agent monthly consumption limits

Last reviewed November 21, 2026

Regulation

Does OSFI Guideline E-23 apply to AI agents?

Very likely, if you’re a federally regulated financial institution and the agent informs decisions. E-23 takes effect on May 1, 2027 for all federally regulated banks, insurers and trust and loan companies. It covers models regardless of the technology used, and OSFI added specific guidance on AI and machine learning models.

In practice that means an inventory of your agents, a risk rating for each, testing before release, monitoring after, and a clear owner. Third-party models also need validation and monitoring in line with their risk. We build agents with that evidence produced as part of the work, so your model risk team isn’t chasing it later.

This is general information, not legal advice. Your compliance team decides how E-23 applies to you.

Source: OSFI — Guideline E-23 Model Risk Management (2027)

Last reviewed November 21, 2026

Regulation

We're a federal department. What does the Directive on Automated Decision-Making require?

The directive applies to any system, tool or model used to make, or help make, an administrative decision about a client. If your agent does that, you need an Algorithmic Impact Assessment before it goes into production. The assessment sets an impact level, and that level decides which extra requirements apply, such as human review, notice to clients and peer review.

The results must be published and updated when the system changes. Agents that only answer general questions, without deciding anything about a person, usually fall outside it, but confirm that with your department.

This is general information, not legal advice.

Sources: Treasury Board — Directive on Automated Decision-Making · Government of Canada — Algorithmic Impact Assessment tool

Last reviewed November 21, 2026

Regulation

We're an Ontario health organization. What do we need before an agent touches patient information?

Under Ontario’s Personal Health Information Protection Act (PHIPA), you stay accountable for personal health information whichever tool handles it. Before an agent reads or produces patient information, plan a privacy impact assessment: what information it can reach, who can use it, where the data is stored and processed, and how access is logged. The Information and Privacy Commissioner of Ontario publishes guidance on doing one.

Many health organizations start with agents that don’t touch patient information at all, like policy look-ups or staff scheduling, while the assessment is done.

This is general information, not legal advice.

Source: Information and Privacy Commissioner of Ontario

Last reviewed November 21, 2026

Security

Does our data stay in Canada?

Mostly, if you set it up that way. Create the environment in the Canada region, and Copilot Studio keeps your stored data in that geography. The exceptions to plan for:

  • Generative AI processing: if a feature isn’t hosted in Canada, prompts and answers may be processed elsewhere, but only if an admin allows “Move data across regions”. Data already moved can’t be pulled back by switching it off later.
  • Web search: Bing search for agents needs that same setting turned on.
  • Services you connect: connectors to outside systems can send data out of Canada.

We document each of these per agent, so your privacy team sees exactly where data goes before anything is switched on.

Sources: Microsoft Learn — Data locations in Copilot Studio · Microsoft Learn — Move data across regions for generative AI features

Last reviewed November 21, 2026

Working with MindLab

How long does it take to get an agent into production?

For one well-defined process, about six weeks with our Pilot to Production offer: agree the measure of success, build and test the agent, set up development, test and production environments, and hand over with the evidence your risk team needs.

What usually slows things down isn’t the build. It’s access to systems, sign-off on data, and deciding who owns the agent afterwards. We sort those out in week one.

See the six-week plan or book a free 30-minute call.

Last reviewed November 21, 2026

Don't see your question?

Book a free 30-minute call, or ask the Ask MindLab agent in the corner of this page.