How to choose an AI agent vendor in Quebec

Let's get this out of the way: we sell custom AI agents. This article about choosing a vendor is written by a vendor, with the conflict of interest that implies. We would rather name it than dress it up, because the most useful thing anyone can hand an SMB shopping around is the list of questions that make vendors squirm. Us included.

The timing is right for a buying guide. According to the Institut de la statistique du Québec, 12.7% of Quebec businesses were using AI applications for production in the second quarter of 2025, up from 9.4% a year earlier. Supply has followed demand: large consulting firms, Microsoft integrators, SaaS platforms, independent studios. Everyone sells "AI agents". Nobody sells the same thing.

So this article maps the market by vendor type rather than by name, because the business model predicts the behaviour, not the logo. Then it gives you the five questions to ask before you sign anything.

The short answer, for the busy

Quebec's AI agent market has four types of vendors: large consulting firms, Microsoft Copilot integrators, SaaS platforms, and custom studios. None of them is bad by design; each has a business model that creates different incentives. Before signing, ask five questions: who owns the code, the prompts, and the data if you leave; what maintenance actually covers (model drift, API changes, monitoring); how the vendor tests for hallucinations before production; where your data is physically hosted, because Quebec's Law 25 requires an assessment before any transfer outside the province; and what cancellation looks like in practice. A serious vendor answers all five in writing. Evasive answers on code ownership and contract exit are the most reliable warning sign. And be wary of any project sold without a specific use case and a defined total cost: the vendor landscape is easier to read once you see each type's incentives.

Four vendor types, four business models

Large consulting firms sell time. Their structural incentive is mandate length: more phases, more committees, more deliverables. They are built for complex projects that touch many systems inside organizations with hundreds of employees, and they are genuinely good at that. For an SMB with 10 to 50 employees, the problem is arithmetic: the entry cost of a mandate often exceeds a reasonable total budget for the project.

Microsoft integrators deploy agents in Copilot Studio and the Power Platform ecosystem. If your business already lives in Microsoft 365, it is a defensible path: authentication, data, and tools are already in one place. Two things to watch, though. First, consumption billing: Copilot Studio is paid in credits, US$200 per month for a pack of 25,000, or pay-as-you-go, and each agent action burns a variable number of credits depending on its complexity, which makes the bill hard to predict before you have real volume. Second, the underlying incentive: the integrator wins by anchoring you deeper into the ecosystem, never by preparing your exit.

SaaS platforms sell an agent in a box: monthly subscription, live in days, unbeatable entry price. The incentive is volume, so standardization. It is the right choice when your need is generic. The trade-off: your "agent" is a configuration inside their system. The day you leave, nothing follows you. And the data often flows through US servers, which is not a footnote in Quebec (more on that below).

Custom studios, our category, build an agent around your specific process, often on subscription. The incentive is retention through delivered value: if the agent stops being useful, you leave. The honest weakness of the type is dependence on a small team. Which is exactly why the ownership and reversibility questions matter most with a studio. For real numbers by model, we broke down what an AI agent actually costs for a Quebec SMB.

Question 1: who owns what when I leave?

The most uncomfortable question, so ask it first. An AI agent is code, prompts, test sets, conversation logs, API keys, and hosting accounts. Each of those pieces has an owner, and the contract should name it.

On a SaaS platform, the honest answer is "almost nothing": you can usually export your data, rarely the logic. With an integrator or a studio, insist that intellectual property in the code and prompts is assigned to you, and check whose name is on the accounts: model provider, hosting, domain. A vendor who keeps the API keys in their own name has a hold on you, even with a perfect contract. The right answer fits in one sentence: "everything built for you belongs to you, and here is the list of accounts with their holders."

Question 2: what does "maintenance" actually cover?

An AI agent does not age like a website. Model providers regularly retire versions from the market, the APIs of connected tools change without notice, and an agent's behaviour can drift without a single line of code changing.

So there are two definitions of maintenance, and they do not resemble each other. The minimal version: we fix what breaks when you report it. The serious version: we actively monitor the agent's answers, measure quality over time, migrate model versions before retirement, and alerts exist for silent failures. Three questions separate the two: who pays for the migration when a model version is retired? Is answer monitoring included or billed extra? What warns you that the agent has started answering badly, before a customer does?

Question 3: how do you test for hallucinations before production?

Virtual agents and chatbots accounted for 24.8% of reported AI uses among Canadian businesses in the second quarter of 2025, according to Statistics Canada. It is also the application type where an invented answer costs the most: a wrong price, a fictional commitment, bad advice given to a customer under your brand.

A serious vendor has a method, not an opinion. Concretely: a test set built from your real business questions, an error rate measured before go-live, an explicit refusal behaviour (the agent must know how to say "I don't know" and hand off to a human), a supervised break-in period, and logged answers you can audit afterwards. If the answer to this question is "the models have become very reliable", keep shopping. And remember that everything starts with scope: an agent that does one well-defined thing can be tested, an agent that does "everything" cannot. That is exactly why the use cases that work in SMBs are narrow and measurable.

Question 4: where does my data physically live?

In Quebec, this is not an academic question. Law 25 requires a business to conduct a privacy impact assessment before communicating personal information outside the province. Most large AI models are hosted in the United States. Nothing forbids using them: the law requires documenting the transfer and framing it contractually, which is doable, but not improvisable.

The test is simple. A vendor deploying agents in Quebec should be able to answer in three sentences: where the data sits at rest, where it travels while the model works, and what the contract says about all of it. If they discover the question in front of you, they will discover compliance on your invoice.

Question 5: what does the exit look like?

Ask for the cancellation procedure in writing before you sign, not after. Four things belong in black and white: the notice period, the export of your data and logs, the transfer of accounts and access, and the cost of the operation. A vendor confident in the value they deliver does not need contractual handcuffs; a 36-month term with penalties tells the opposite story.

The fastest test on the market is one question: "if I leave in twelve months, what do I take with me, and what does it cost?" The vendor who answers in one page stands behind their model. The one who answers "we'll see when we get there" just answered.

Funding tests your vendor too

One last filter, less known: knowledge of public programs. The Regional Artificial Intelligence Initiative (RAII) offers SMBs an interest-free repayable contribution covering up to 50% of authorized costs for an AI adoption project, in effect until March 31, 2031. The decisive detail: only costs incurred after the application is filed are eligible.

A vendor who knows the program will structure the project into fundable phases (diagnostics, design, implementation, knowledge transfer) and will tell you themselves to wait for the filing before signing the first invoice. A pushy seller who wants to start "this week" may have just cost you half the project. We covered how the RAII works and its filing mechanics if you want to dig in.

Ask us the five questions

This guide is not disinterested, as we said up front. But it is falsifiable: every one of these five questions gets a written answer in our proposals, code ownership included. If you are shopping for an AI agent, put them to every vendor on your list, us included. That is exactly what a first conversation is for: defining the task, the real cost, and the exit path before committing a dollar. 30 minutes, no commitment.

→ First conversation, no commitment

This article helps an SMB structure its vendor search; it is not legal advice. The exact obligations under Law 25 (assessments, contractual framing of transfers) depend on your situation: validate with legal counsel where needed.