21 checks to run before that becomes your problem — written for Ontario businesses of 5 to 100 people, under PIPEDA and PHIPA rather than someone else's jurisdiction. Twenty minutes with whoever runs your IT.
Straight to your inbox, and shown on this page immediately. No sales call attached.
List every AI tool in use — including the ones nobody approved. ChatGPT, Copilot, Gemini, an AI notetaker in Teams, a transcription plugin, a resume screener. Shadow AI is the norm, not the exception.
For each one, write down what data staff actually paste into it. Client names, invoices, contracts, medical details, payroll. This is the list that matters, not the vendor's feature list.
Check whether the free tier trains on your input. Consumer tiers usually do; business tiers usually do not. The same brand can be both, so the tier is the control, not the brand.
Find out which country the data is processed in. If the answer is "United States", that is a cross-border transfer you are responsible for disclosing.
Obtain a data processing agreement, not a marketing page. If the vendor will not sign one, you cannot use them for client data.
Confirm the retention period for prompts and outputs. "Retained for 30 days for abuse monitoring" means a copy of your client's data exists on someone else's server for 30 days.
Confirm whether sub-processors are used, and who they are. Most AI vendors are a thin layer over a larger model provider.
Confirm the breach notification commitment and its timeline. If the vendor is breached, you still owe notification to your clients and possibly to a regulator.
PIPEDA applies to almost every Ontario business handling personal information. Consent must cover the purpose you are actually using the data for — "we use AI" is not a purpose.
Health information is PHIPA territory. If you are a clinic, a therapist, or anyone acting as an agent for one, health data goes only into a system that has signed on as your agent under PHIPA.
Legal practices: the Law Society's confidentiality duty does not have a convenience exception. A public model is a disclosure to a third party.
Anything sold to a US client or handling EU residents adds GDPR or state privacy law on top. Do not assume Canadian rules are the ceiling.
Write down who at your business owns the decision. Unassigned accountability is where every one of these programmes stalls.
Turn on SSO and MFA for every AI tool, on the same identity provider as the rest of your stack. An AI tool with its own separate password list is an offboarding failure waiting to happen.
Remove AI tool access in your offboarding checklist. Departed staff keeping a Copilot licence is common and invisible.
Restrict which data sources an AI assistant can reach. Copilot indexing a wide-open SharePoint site will confidently surface salary spreadsheets to whoever asks.
Keep a log of AI-assisted decisions that affect people — hiring, credit, pricing. If a decision is challenged, "the model said so" is not a defence.
Classify your data into three buckets: fine for a public model, acceptable in a contracted business tier, and never leaves our infrastructure. Most businesses discover the third bucket is bigger than they assumed.
For that third bucket, evaluate a self-hosted model. Open-weight models now run on a single workstation-class GPU and are good enough for summarisation, drafting, classification, and document search — which is most of what SMBs actually use AI for.
Price it honestly: self-hosting trades a per-seat subscription for hardware and someone competent to run it. It wins on privacy and on cost at scale, not on day one.
Write a one-page AI use policy and have staff read it. It should name the approved tools, the banned data, and who to ask. One page gets followed; twelve pages do not.
That is the normal outcome. We run this exact review for Ontario SMBs, and where the answer is "this data should never leave the building", we deploy self-hosted models on your own infrastructure instead of shipping it to a US provider.
See transparent pricing