Start with the work, not the model. A small business does not need an “AI strategy” before it can improve one painful workflow. It needs to see where time, errors, delay and customer friction accumulate - and whether a bounded intervention can remove enough of that cost to justify itself.
1. Build a friction inventory
Ask each person to list work that is repetitive, rule-heavy, delayed by handoffs or regularly re-entered. Look at inboxes, calendars, shared drives, paper forms and spreadsheets. For one week, record five facts for each candidate:
- How often it occurs.
- Minutes of human effort per occurrence.
- Loaded hourly cost of the people involved.
- Rework, error or delay rate.
- Business consequence: lost cash, slow response, poor experience or compliance exposure.
Do not count only keystrokes. A ten-minute task that interrupts a manager twenty times per week can be more expensive than the visible labour suggests.
2. Score value and feasibility separately
Use a one-to-five score for six dimensions. Value consists of volume, time/cost and consequence. Feasibility consists of input consistency, rule clarity and reversibility. Multiply the two averages to get an opportunity score out of 25.
A high-value, low-feasibility process should be redesigned before it is automated. A highly feasible but trivial process is a useful practice exercise, not a priority.
3. Calculate the conservative return
A distributor receives 160 quote-request emails per month. An inside-sales employee spends 12 minutes extracting product, quantity, delivery date and contact information, then re-enters it into a tracker. Loaded labour cost is estimated at $42 per hour.
| Current monthly labour | 32.0 hours |
| Current monthly labour cost | $1,344 |
| Expected time removed at 70% | 22.4 hours |
| Monthly labour capacity recovered | $941 |
| Estimated monthly software/monitoring | ($180) |
| Conservative monthly benefit | $761 |
If discovery, setup, testing and training cost $6,000, simple payback is about 7.9 months. That does not mean laying anyone off; it means deciding whether 22 hours of recovered capacity can be redirected to quoting faster, following up or handling more volume.
Add error avoidance only when you can defend it. If four incorrect quotes per month create an average $125 correction cost and the pilot can prevent half, that adds $250 per month. Do not assign imaginary revenue to every saved minute.
4. Decide whether AI is even needed
- Use a form or validation rule when inputs can be standardized.
- Use conventional workflow automation when the logic is deterministic: if this happens, create that record and notify this person.
- Use AI when the useful input is unstructured language, images or documents and the output can be checked.
- Keep a human decision when the consequence is material, the evidence is ambiguous or the action affects employment, credit, safety, health or legal rights.
Statistics Canada found text analytics, data analytics and virtual agents among the most common AI applications used by adopting Canadian businesses in 2025. That supports a practical starting point: reading, classifying and drafting around existing work, not autonomous control of the company.
5. Design a reversible pilot
- Choose one input source and one output.
- Run the new method beside the old process for two to four weeks.
- Require human approval before any external message or record-changing action.
- Measure cycle time, accuracy, exception rate, adoption and hours actually recovered.
- Define a stop condition before launch.
Map the implementation in Pathway
Track the baseline, acceptance tests, exceptions, owner and pilot results in the free project tracker.
Sources and methodology
- Statistics Canada: AI use by Canadian businesses, second quarter 2025.
- NIST AI RMF Core, including defining business context and specific supported tasks.
- Canadian privacy commissioners' principles for generative AI.
- Walker Deibel, Buy Then Build - pages 12, 23 and 63; Codie Sanchez, Main Street Millionaire - pages 53, 89 and 199. Book references informed the operating lens; Canadian claims rely on the linked primary sources.
Bring one stubborn workflow.
We will help you identify the smallest useful intervention, the real baseline and the evidence that would make a pilot worth continuing.
Book a free call →This article is general operational information, not legal, privacy, employment or financial advice. Validate requirements for your industry and jurisdiction.