AI and automation pay back on a narrow set of tasks: frequent, repetitive work where a person can check the output in seconds. The way to find them is to score candidates on volume, minutes per item and the cost of errors, then do payback math that includes the human review time most vendor calculators leave out. Test on a sample of 50 real items before rolling anything out, and agree in writing what data the tools may and may not touch.
The worked example below uses supplier invoice processing, a common first candidate. All figures are illustrative assumptions in Canadian dollars.
Step 1: list candidates and score them
Start with a list of 10 to 15 tasks your team repeats, then score each one. Strong candidates are high volume, take several minutes, follow a pattern and are easy to verify.
| Task | Volume per month | Minutes each | Pattern | Easy to check? | Data sensitivity |
|---|---|---|---|---|---|
| Keying supplier invoices | 600 | 6 | Strong | Yes | Low |
| First reply to common questions | 400 | 4 | Medium | Yes | Medium |
| Sorting incoming requests | 900 | 1 | Strong | Yes | Medium |
| Writing custom proposals | 20 | 90 | Weak | No | Medium |
| Reviewing employee files | 15 | 30 | Weak | No | High |
Monthly minutes is volume × minutes: invoices 3,600, first replies 1,600, sorting 900, proposals 1,800, employee files 450. Proposals look large, but they follow no pattern and are hard to check, so they are poor first candidates. Employee files carry high sensitivity and low volume. Invoices come out on top.
Step 2: payback math that includes review
The honest calculation subtracts the cost of the tool, the time people still spend checking and the time spent fixing errors. Here is the invoice example at a loaded cost of $40 per hour.
Today: 600 invoices × 6 minutes = 3,600 minutes, or 60 hours a month. 60 × $40 = $2,400 a month.
With automation:
- Tool fee: $300 a month (assumed).
- Human review: 600 × 1.5 minutes = 900 minutes, or 15 hours × $40 = $600.
- Error fixing: if 4% need correction, that is 24 invoices × 20 minutes = 480 minutes, or 8 hours × $40 = $320.
- Monthly cost: $300 + $600 + $320 = $1,220.
Monthly saving: $2,400 − $1,220 = $1,180.
One-time setup: $6,000 for configuration, mapping to your accounting system and testing (assumed).
Payback: $6,000 ÷ $1,180 = about 5.1 months.
Now test it. If the error rate is 10% instead of 4%, error fixing becomes 60 × 20 minutes = 20 hours, or $800, and the saving drops to $700 a month, so payback stretches to about 8.6 months. If volume halves to 300 invoices, the saving roughly halves too, and the case gets weak. The error rate and the volume are the two numbers to verify.
Step 3: test accuracy on a sample of 50
Never decide on a demo with the vendor's own examples. Take 50 real items from last month, including messy ones, and run them through.
- Pick the 50 deliberately: different suppliers, handwritten notes, credit notes, multi-page bills.
- Decide in advance which fields count (supplier, date, invoice number, subtotal, GST/HST, total).
- Have a person record the correct answer for each field.
- Compare and count. If 48 of 50 invoices come through with every field right, that is 96%.
- Look at the failures. Two random errors are manageable; two failures on the same supplier format show a pattern you can fix or route to a person.
Set your pass mark before the test. For invoices going into the books, many owners want 95% or better with every exception flagged rather than silently guessed.
Privacy and the rules for AI use
Canada's privacy rules apply to AI tools the same way they apply to any other system that handles personal information. We summarise them here; we do not give legal or compliance advice.
The Office of the Privacy Commissioner of Canada explains that PIPEDA sets the rules for private-sector organizations that collect, use or disclose personal information in commercial activity. Alberta, British Columbia and Quebec have their own private-sector laws that are considered substantially similar, while PIPEDA still covers information that moves across provincial or national borders.
In December 2023, Canada's federal, provincial and territorial privacy commissioners issued principles for responsible, trustworthy and privacy-protective generative AI. The nine principles cover legal authority and consent, appropriate purposes, necessity and proportionality, openness, accountability, individual access, limiting collection, use and disclosure, accuracy, and safeguards. They are written mainly for developers and organizations using these tools, and they are a sensible checklist for an owner too. As of October 2026, confirm the current guidance on the commissioner's site.
A one-page AI use policy outline
A short written policy prevents most problems. Ask your privacy lead or lawyer to review it before you adopt it.
- Approved tools: which AI tools staff may use for work, under which accounts.
- Data rules: what may never be entered (for example, health details, banking information, employee files) and what may be entered only in approved tools.
- Human review: which outputs must be checked by a person before they reach a customer or the books.
- Disclosure: when customers are told that a reply or document was drafted with AI.
- Ownership: who in the business is accountable for each automation and reviews it quarterly.
- Incidents: what staff do if something goes wrong or data is entered by mistake.
Step 4: roll out in stages and keep measuring
Pass the test, then expand gradually so a problem surfaces while it is still small. A staged rollout over about six weeks works for most first projects.
- Weeks 1–2: run the automation alongside the manual process on one supplier group. Staff still key everything; you compare.
- Weeks 3–4: switch that group over, with a person reviewing every item.
- Weeks 5–6: add the remaining suppliers and move to reviewing only the items the tool flags as uncertain, plus a random 10% sample.
- Monthly after that: record volume, review minutes, error count and fixes, and compare with the payback model.
If the monthly saving falls well short of the model for two months running, find out why before adding a second automation. Usually it is review time creeping up because staff don't yet trust the output, which is a training and threshold question rather than a reason to abandon the tool.
Common mistakes to avoid
Most disappointing AI projects fail on scope or measurement, not technology.
- Automating a broken process. If the manual process is inconsistent, automation repeats the inconsistency faster.
- Ignoring review time. A tool that saves five minutes but needs four minutes of checking saves one.
- Starting with the hardest task. Proposals and judgement-heavy work are poor first projects.
- No baseline. If you never measured minutes per item before, you can't prove the saving after.
- Tool sprawl. Five staff using five different AI tools with company data is a governance problem, covered on our digital governance page.
How we help
Our AI adoption and automation engagement scores your candidate list, builds the payback model, designs the 50-item test and drafts the policy outline with you. If you are still deciding where AI fits among other projects, the guide to prioritising digital projects shows how to rank it fairly. Fees are on the pricing page; a text to (604) 632-4959 naming the task you'd like to automate is a good start.
Frequently asked questions
What kinds of tasks are good candidates for AI or automation?
Tasks done many times a week, in roughly the same way, where a person can quickly check the result. Examples include pulling data from supplier invoices, drafting first replies to common customer questions, sorting incoming requests and summarising long documents for review.
How long should an AI automation take to pay back?
For a small or mid-sized business, we look for payback within about 12 months on a realistic estimate that includes review time and error fixing. If the honest math needs more than that, the task is usually better left for later.
Can we put customer information into AI tools?
That depends on the tool's terms, where data is stored and the privacy laws that apply to you. Canada's privacy commissioners have published principles for generative AI, and we can summarise them, but whether a specific use is allowed is a question for your privacy lead or lawyer.
Do you build the automations?
We help you choose the tasks, define the test, select tools and set the rules. Building can be done by your team, the tool vendor or a developer you hire directly; we oversee it from your side if you want.