Buying an AI agent platform in Canada

Evaluating an AI agent platform means looking past whatever the product demo shows and asking what happens when the agent gets something wrong, because it eventually will. A platform’s ability to take autonomous action is exactly what makes it valuable and exactly what makes it risky, and a buyer who focuses only on the capability without pressure-testing the controls around it is buying half the picture.

What a good one actually looks like

Look for verifiable task-completion data measured against real customer workflows, not a vendor’s own marketing benchmark, and for guardrails and permissioning documentation detailed enough that an enterprise compliance team would actually approve it in production. Deep integration with the systems the agent genuinely acts on — a CRM, an ERP, a finance tool — is a stronger signal than a broad list of surface-level connectors that were never hardened past a demo.

Judge the revenue model, not just the demo

A strong demo tells you what the agent can do; it tells you almost nothing about what the business actually earns. Revenue on an agent platform is usually tied to completed tasks or actions rather than a flat subscription, so ask for the trend by customer, not just the topline, and check whether retention is holding as usage scales rather than being propped up by new-customer growth. Then weigh that against the cost side: inference cost per completed task tends to rise with task complexity, which means revenue can grow while margin quietly compresses if the mix of tasks shifts toward harder ones. A platform that can show both sides of that relationship clearly, revenue and the cost that shadows it, is a materially different opportunity than one that can only point to top-line growth. If the seller can only produce a single blended margin figure and not the underlying trend, treat that as a gap to close before you settle on a price, not a detail to sort out later.

Ask the single-vendor question directly

Find out whether the orchestration depends entirely on one foundation model’s native tool-calling feature, and ask what the fallback plan is if that vendor changes pricing, changes terms, or discontinues the feature. A seller who has clearly thought this through and built some resilience around it is running a different business than one who has never had to answer the question.

Confirm what actually transfers with the business

Ownership of an agent platform is only as good as what actually moves to you at closing. Confirm that the orchestration engine and the guardrail and permissioning infrastructure are independently owned by the seller rather than licensed or borrowed from another vendor, because a platform built on someone else’s proprietary orchestration layer is a very different asset than one you’ll fully control after the sale. Check that the API credentials and integrations into each customer’s CRM, ERP or finance system are actually assignable, and ask what re-authorization each one requires, since a system that has to be manually reconnected by the customer isn’t something you can simply guarantee will happen on schedule. Get clear as well on whether the audit logs and incident history transfer with the business or stay with the seller, because you will need that history to run the platform responsibly from day one.

What a seller may not volunteer

Ask directly, more than once, about any incident where the agent took an incorrect or costly action on a customer’s behalf. A platform that reports zero incidents ever is either genuinely careful or simply isn’t tracking its own failures, and it’s worth finding out which before you rely on that clean record. Push specifically for how any past incident was resolved and where the resulting liability landed, since that answer tells you more about how the business is actually run than the incident itself does.

Qualifying yourself to take this on

No AI-specific licence governs owning an agent platform in Canada today, but if the agent acts inside a regulated domain — initiating payments, or producing what reads as financial or legal guidance — the same rules that would apply to a person doing that task can apply to you once you own the business delivering it. Know what regulatory exposure comes with the specific use cases the platform serves before you close, not after.

Reading the regulatory exposure you’re buying into

PIPEDA applies wherever the agent processes personal information, and for Quebec customers, Law 25’s automated-decision disclosure rules are particularly relevant whenever the agent’s action materially affects a person. Check what the target has actually built to comply — documented processes, not just a privacy-policy statement — rather than taking the seller’s word that it’s handled.

Red flags worth slowing down for

• No meaningful guardrails around what actions the agent is permitted to take.
• An orchestration layer that is essentially a thin wrapper over one foundation model’s tool-calling feature.
• Task-completion claims that cannot be verified against actual customer data.
• Any incident involving the agent acting incorrectly that was never disclosed until you asked directly.
• Contractor-built orchestration or guardrail code with no confirmed IP assignment.