Testing generative AI for customer-facing work
Trial AI-assisted marketing and service content with checked source facts, human review and measures that include correction time and customer impact.

Generative AI can draft marketing copy, propose ideas or help staff prepare a customer response. A useful pilot needs to show that the result is accurate, fits your customers and doesn’t take too long to check.
Choose one task with a limited audience and an established human owner. Don’t connect it to live customer interactions until you’ve tested the workflow.
Start with your own reference material
Prepare a small set of current service facts, prices and policies the team can use. Leave out anything your business hasn’t agreed to put into that AI service.
The Canadian privacy regulators’ generative AI principles address legal authority, data minimization and accountability. Check privacy and supplier terms before submitting personal information.
Publicly visible information can still identify a person. Don’t assume a customer review, social profile or conversation is unrestricted training material.
Try a draft with a clear reviewer
For marketing, ask for a short draft based on verified facts. Have the content owner check each product claim, price, link and call to action.
Review tone, accessibility and Canadian spelling. Check generated images and other assets against licensing terms and brand requirements before using them.
Keep the source facts beside the draft. A polished sentence can introduce a capability your business doesn’t offer.
Bound customer-service use
Begin with staff reviewing proposed responses. Define which questions need escalation, such as account disputes, privacy requests or advice outside the business’s scope.
If you later consider a customer-facing assistant, test what happens when it doesn’t know, when a record is outdated and when a customer asks for a person. Set access limits and logging before rollout.
An answer about policy must match your current written policy. The business can be held to what its assistant tells a customer: in Moffatt v. Air Canada, British Columbia’s Civil Resolution Tribunal made the airline pay a refund its website chatbot had described wrongly.
Verify analysis separately
For a proposed trend or calculation, compare the output with the source data and an independently checked result. Check missing records, definitions and time periods.
Do the arithmetic in a spreadsheet or analysis tool and keep the calculation reproducible. Don’t make pricing, hiring or customer-eligibility decisions from an unverified generated explanation.
Measure the complete workflow
Record drafting time, review time, corrections and any complaints or rework. Compare with the current process using similar tasks.
Include licences, integration and maintenance costs. A faster first draft can still take longer to review.
At the review date, decide whether to continue, narrow the task or stop. The first-automation guide helps separate this trial from wider process changes.


