AI automation for small business: a practical checklist

Canadian businesses are taking this up fast: 19.2% used AI to produce goods or deliver services in the 12 months before Statistics Canada’s second-quarter 2026 survey, up from 12.2% a year earlier and 6.1% in 2024. In StatCan’s third-quarter 2026 survey, 25.2% planned to use AI in the next 12 months, up from 14.5% a year earlier.

For a small business that mostly means the admin between your inbox, phone, CRM and accounting tool: copying a lead into Jobber, chasing a quote, filing a receipt.

Step 1: Find the work worth automating

The first step in how to automate your small business is a one-week log, not a tool demo. For one normal week, note every repetitive admin task as it happens, yours and your staff’s. A sample from a small trades business:

Illustrative numbers. For your own log, copy the six column headings into a spreadsheet and add a row per task.

Multiply minutes by times a week. Invoices (160 minutes a week) come first; missed calls and custom pricing tie at 100, but only missed calls follow rules. Pricing a custom job needs judgment, so it gets AI help with a person deciding. Score your top five:

Step 2: 12 AI workflows for small business (start with one)

Every recipe below is a trigger, a few fixed steps, at most one AI step and a named person who checks the result. That follows Anthropic’s advice to start by “finding the simplest solution possible, and only increasing complexity when needed.”

None of the 12 needs an agent. Agents are worth the extra testing on the phone and in shared inboxes where each message needs a different next step (see what an AI agent is); for tasks a person starts on demand, the connectors in ChatGPT, Claude or Grok may be enough.

Start with one of these four. They cover the usual leaks from a time audit: missed calls, cold quotes, slow invoices and late payments.

• 1 Missed call to text-back Leads

  Within a minute, texts the caller your business name, “we’ll call you back” and a booking link; logs the lead and alerts the owner.

  Trigger

  A call goes unanswered (Twilio reports busy or no-answer).

  Tools

  Twilio or your VoIP provider, Jobber or HubSpot, Telegram.

  AI step

  Optional, a voicemail summary in the alert.

  Who checks

  The owner returns urgent calls the same day.

  If most missed calls come after hours, let an AI receptionist answer instead; see AI receptionist vs answering service and what one costs in Canada.

• 2 Quote follow-up on day 2, 5 and 10 Quotes

  Before each nudge, checks for a reply, an approval or a decline; any one ends the sequence. After day 10, marks the quote cold and creates a call task.

  Trigger

  A quote moves to Awaiting response in Jobber, or to sent in your CRM.

  Tools

  Jobber, HubSpot workflows or a platform.

  AI step

  Optional, one line from the quote notes, never prices.

  Who checks

  The owner works the day-10 call list.

  Jobber builds in up to two follow-ups from its Connect plan, but its version stops on approval, not on an email reply, so mark the quote before the next send or use a platform. The spec in Step 5 is for this recipe.

• 3 Job done to invoice draft (automated invoicing) Closeout

  Pulls hours, parts and line items into a draft invoice, with the tech’s note and photos attached. Prices come from your price list, never from the model.

  Trigger

  A job is marked complete.

  Tools

  Jobber, QuickBooks Online or Xero.

  AI step

  Turns “replaced cap + contactor, 1.5h” into line items a client understands.

  Who checks

  The owner approves and sends every invoice.

• 4 Payment reminders Money

  You may not need to build this. QuickBooks Online sends up to three automatic reminders, each up to 90 days before or after the due date, and Jobber up to two on past-due invoices (Connect plan).

  Trigger

  The due date, then 7 and 14 days past it.

  Tools

  Built into QuickBooks Online and Jobber.

  AI step

  None.

  Who checks

  The owner calls after the last reminder.

Eight more workflows to add later

In the order work moves through a service business:

On 10, ask every customer: Google bans selectively asking for positive reviews and offering incentives. On WhatsApp, a message outside the 24-hour window needs an approved template.

Recipes 1 to 4, 6 to 8 and 10 message customers, so read Step 4 before switching them on.

Step 3: Zapier vs Make vs n8n, and what each costs

Zapier, Make and n8n do the same core job: watch for a trigger, then run steps across your apps. What changes your bill is the unit each one counts.

Vendor pricing pages on October 14, 2026, before tax. Prices are per month, billed yearly, unless marked monthly. n8n prices in euros and says yearly billing saves 17%.

Zapier task

Each action that runs successfully. Triggers, filters, paths, and Formatter, Delay and Storage steps are free; an AI by Zapier step can cost more than one task, depending on the model tier.

Make credit

Covers most module actions, and the trigger uses one on every run, even a scheduled check that finds nothing. In Make’s own example, hourly checks use 720 credits a month before any work happens. Routers use none.

n8n execution

One complete run, however many steps it has.

Worked example: one workflow, three bills

Take recipe 7 as six steps: form trigger, AI extract-and-draft, find or create the contact, create the deal, save the Gmail draft, Telegram alert. It runs 500 times a month, and the AI step uses your own OpenAI or Anthropic API key, so each platform counts it as one ordinary step. Built-in AI, such as AI by Zapier or Make’s AI Provider, bills differently.

Bank of Canada rates for October 14, 2026 (US$1 = C$1.4263, €1 = C$1.5883), before GST/HST and card fees. Zapier’s pay-per-task overage costs 1.25 times your plan’s price per task, so 500 tasks over the US$49 tier add about US$15.31, for about US$64.31 a month.

The model is billed on top by the provider. At Anthropic’s API price for Claude Sonnet 5.5 (US$2 per million input tokens, US$10 per million output), a 2,000-token email and a 300-token draft cost about US$0.007, or US$3.50 for 500 runs.

Zapier charges for every action step, Make for every module plus every polling check, n8n once per run. Zapier gets a short workflow live fastest; for long workflows at volume, Make or n8n cost a fraction as much.

Check what you already pay for

Jobber sends quote and invoice follow-ups from its Connect plan, QuickBooks Online sends invoice reminders, and HubSpot includes workflows on Professional and Enterprise. Built in means no extra login and no extra vendor holding client data. Custom code makes sense at high volume or with unusual systems; Lane is one.

Step 4: Build Canadian rules into the workflow

CASL: consent for commercial email and texts

Canada’s Anti-Spam Legislation covers commercial electronic messages: email, SMS, WhatsApp and other text, sound or image messages. Each one needs consent, must identify your business with contact details, and must carry a free, working unsubscribe.

Those CASL rules don’t cover live two-way voice calls, faxes or voice recordings sent to a phone number. Calls fall under the CRTC’s Unsolicited Telecommunications Rules, which ban telemarketing calls using a pre-recorded or synthesized voice without express consent. An AI voice agent that dials out to sell is likely to be caught, so get that consent in writing or on a recorded call first.

Implied consent from a business relationship expires 2 years after a purchase or contract and 6 months after an inquiry; express consent, the box they ticked, doesn’t. If challenged, you must prove consent, and dates are where businesses get caught. In 2019 the CRTC imposed a $100,000 penalty on nCrowd’s former president and CEO after the company’s list showed express consent for over 1.5 million addresses on a single day, which the CRTC couldn’t verify.

• Store consent type (express, purchase or inquiry), date and source (form URL, call, invoice number) for every contact.
• Before each commercial send, check for express consent, a purchase in the last 2 years or an inquiry in the last 6 months. If none applies, don’t send.
• Put your business name, contact details and an unsubscribe link in every message. The link must work for at least 60 days after sending.
• Process unsubscribes within 10 business days at most. A single suppression list that every workflow checks before sending is safer than each tool’s own unsubscribe list. For texts, honour STOP replies too.
• Never load a bought or old list into a sequence.

A message that solely sends a quote they asked for, or confirms or completes a transaction they agreed to (a booking confirmation, an invoice), doesn’t need consent. It still needs your identification and an unsubscribe. Add a promotion and the exemption is gone.

PIPEDA: you’re accountable for your vendors

Your automation platform, AI provider and SMS provider all handle client data for you, and under PIPEDA you stay responsible for it. The Privacy Commissioner expects contracts that give comparable protection, and customers told their information may be processed abroad. List each workflow’s vendors and pass each step only the fields it needs.

CRA: six years, kept in Canada

Workflows that create invoices or file receipts are creating business records. Keep them six years from the end of the last tax year they relate to, in Canada unless the CRA gives written permission. Records stored outside Canada and accessed electronically from here don’t count, so know where your storage lives.

Quebec: fully automated decisions

If a workflow decides something about a person using only automated processing, such as auto-declining a job applicant, section 12.1 of Quebec’s private-sector privacy law applies. Tell them when you give the decision; on request, explain the information and main factors used and their right to have that information corrected; and let them submit observations to someone who can review it. If a person makes the final call, section 12.1 doesn’t apply.

Step 5: Guardrails before go-live

1. Draft mode first. For two weeks, every outgoing message waits for approval. n8n, for one, can hold an AI agent’s tool call for a person to approve in Slack, Telegram, Gmail or WhatsApp.
2. Least privilege. A separate login or API key per workflow, with only the access it needs. If it drafts, it can’t delete or forward.
3. A kill switch. One shared setting, such as a Google Sheet cell reading SENDING = ON, that every customer-facing workflow checks before sending. Set it to OFF and they all stop. Test it monthly.
4. A send cap. Each workflow counts its own messages in the last hour; at 20 it stops, pauses itself and alerts the owner on Telegram. If a broken filter matches all 600 past clients, 20 get a confusing email, not 600. Zapier Storage, Make data stores and n8n data tables can hold the counter.
5. Error alerts. A failed run pings a named person the same day, so a stalled follow-up gets noticed in a day, not a month.
6. Logs and test contacts. Record what ran and what the AI step received and returned. Test on your own phone and inbox before real clients.
7. Treat what the AI reads as untrusted. An incoming email can contain text aimed at your AI, so no AI step should trigger a payment, deletion or bulk send on its own. More in AI agent risks and guardrails.

Write a one-page spec first

Before anyone builds, put the workflow on one page. Here is recipe 2, built on a platform so it can watch the inbox:

Step 6: Measure the hours saved

You logged the “before” in Step 1. A month after launch, log the same task again, including time spent checking the automation’s work.

Hours saved per month = (minutes before − minutes after) × times per month ÷ 60 − review and fix time

Say a web-form lead took 9 minutes to read, enter and answer, 60 times a month, and now takes 2 minutes to check the draft and send. (9 − 2) × 60 ÷ 60 = 7 hours, minus 30 minutes reading logs: 6.5 hours a month.

Also track time to first reply, quotes that got every follow-up, and days to payment. Faster first replies and fewer days to payment are where the money shows up.

A real example: Lane, end to end

One job through Lane, the dispatch system we built for a Toronto home-services company, with each step tagged by who or what does it:

1. Automation

   9:42 PM, the phone rings. The business number connects through the phone provider’s API, so the call starts a flow instead of hitting voicemail.

2. AI

   The voice agent answers. An ElevenLabs agent works out the problem, books a window for the next morning and can hand the caller to a person. It’s the same setup as our AI receptionist service.

3. Automation

   The job record is created. Customer, address, issue, priority and call summary land in the CRM; repeat customers are recognized.

4. Automation

   One Telegram alert. The owner gets a single de-duplicated message linking to the job on the dispatch board. Contact details stay behind sign-in.

5. Person

   The tech closes out on WhatsApp. After the visit, the assigned technician replies “done” with a note and a photo.

6. Automation

   The job closes. Lane marks it complete and sends the review request.

AI handles one step, the conversation, because every caller is different. The rest is fixed steps that are cheap to run and easy to test, plus a person confirming the work is done.

Lane is custom code (Next.js and Supabase) because it runs several service lines, each with its own number, agent and records. A single-location business can get most of the flow from recipes 1, 7, 9 and 10. More for trades in AI for home service businesses.

A 30-day plan

1. Week 1

   Run the time audit, score tasks, pick one workflow and write its spec.

2. Week 2

   Build it in draft mode with test contacts, plus the send cap, kill switch and error alerts.

3. Week 3

   Go live with approvals on, and read every run.

4. Week 4

   Measure, relax approvals on steps that ran clean, pick the next workflow.

Our business process automation service covers intake, follow-ups, CRM and invoicing, and our AI automation agency work covers agents and voice.

FAQ