A good AI pilot answers a business question: can this workflow become more reliable at an acceptable cost? A fluent demo does not answer it. For a Canadian or North American service business, start with one process you can observe, such as after-hours enquiries or following up on unsent estimates.
The checklist below is our suggested working method, not a claim that every business needs a month-long AI project. If the process is low-volume or your records are incomplete, allow more time to establish a useful baseline.
Choose a workflow with an owner
Write one sentence: “When an after-hours enquiry arrives, collect the service needed and contact details, then place it in the team's queue for review.” Name the person who owns that queue. List what the system may do and what it must hand to a person. Do not start with “automate customer service.”
If your scheduling rules change every day and only one person knows them, document those rules first. A shared checklist or a better booking form may solve the problem without AI.
Record the starting point
For a representative period, record the number of eligible enquiries, how many receive a response, how many become completed jobs, and how much staff time is spent fixing mistakes. Separate spam and requests you could never serve. Keep your measurement definition the same during the pilot.
| Measure | Write down before launch |
|---|---|
| Eligible volume | Which requests count and which do not? |
| Completion | What is a successful outcome: a qualified enquiry, confirmed booking, or completed job? |
| Quality | What counts as an incorrect answer, duplicate record, or failed handoff? |
| Human effort | Minutes spent reviewing, correcting, and maintaining the workflow. |
| Cost | Setup, subscriptions, usage, and staff time, in the same currency. |
Test the awkward cases before real customers
Use invented or appropriately de-identified examples first. Include a caller changing their mind, an unavailable appointment, a duplicate enquiry, an unclear address, a request outside your service area, an outage in the booking system, and a customer asking for a person. Confirm that an unsuccessful write to the calendar cannot produce a “booking confirmed” message.
For sensitive or urgent requests, define the escalation with the responsible professional. A receptionist should not improvise clinical, legal, or emergency advice. Your team needs to know what happens when the system cannot safely complete the request.
Start with a limited live scope
Use one channel, a restricted service, or a review-before-send queue. Keep a clear fallback that staff can activate. Review the first live cases closely and record the reason for every correction. Change one part at a time so you can tell whether a fix worked.
A suggested sequence is baseline, offline testing, limited live operation, and a decision review. Those stages matter more than an arbitrary deadline. Do not widen access just because the pilot calendar says week three.
Count value without counting it twice
Use contribution from additional completed work, not the full value of every enquiry. Deduct the direct cost of delivering that work. Count saved staff time separately, and only treat it as a cash saving if it actually changes your spending. Subtract ongoing system and review costs. Show setup costs separately so a promising monthly result does not hide a long payback period.
Compare like-for-like periods and note seasonality, promotions, staffing changes, and small sample sizes. A busy week following a marketing campaign is not proof that the assistant caused the increase.
Make the decision in advance
- Expand: the agreed quality and economic targets are met, and the team can operate the fallback.
- Revise: the problem is specific and fixable, with a bounded additional budget.
- Stop: the process does not need AI, the volume cannot justify the cost, or unresolved errors make it unsuitable.
Stopping a pilot is a valid result. Keep the documented process, baseline, and lessons; those remain useful whether you buy software, hire a specialist, or improve the manual workflow.
Keep a record of the data decision
List what customer information enters each service, who can access it, and when it is deleted. Canadian privacy authorities explain that existing privacy obligations still apply to generative AI and vary by organization and activity. Check the requirements for your business before using real customer data. Their generative AI principles are a useful starting point.
For a broader framework, NIST's voluntary AI Risk Management Framework offers a structured way to think about evaluation and ongoing risk. This checklist is our operational application, not a certification.
Next: use our vendor handover checklist to decide what must remain under your control, or review the cost breakdown before setting your pilot budget.