Guide

How to find what's worth automating in your business

Start with the work, not the tool. Find where the hours actually go, check it against evidence, and only then decide what a machine should do, what it should prepare for a person, and what should stay human.

Updated · 6 min read

AI use is climbing fast. In Statistics Canada's survey for the second quarter of 2026, 19.2% of Canadian businesses reported using AI to produce goods or deliver services over the previous 12 months, up from 12.2% a year earlier and 6.1% in 2024 (Statistics Canada). For most owners the question is no longer whether to use it, but where.

Why the first automation is often the wrong one

Most teams pick their first automation by what is most visible or most annoying. That is rarely what costs the most. The work that quietly eats a week is usually small and frequent: retyping the same details into a second system, chasing an approval, rebuilding a document from last month's version.

Those steps rarely appear in a process document, and the owner's picture of how a task gets done often differs from what actually happens at the desk. So the method below starts from evidence and only then asks people to confirm it.

Step 1: List each role and what it produces every week

Don't start with tasks. Start with outputs: quotes sent, invoices entered, schedules published, reports delivered, payments chased. For each role, write down what leaves their desk every week and roughly how many.

RoleWeekly outputsRoughly how many
Office managerNew enquiries logged, site visits booked, quote follow-ups25 / 15 / 20
EstimatorEstimates priced, supplier price requests, quote documents12 / 8 / 12
BookkeeperSupplier invoices entered, invoices sent, payments recorded60 / 10 / 15

An illustrative example for a small contractor. Your roles and numbers will differ.

Step 2: Gather evidence before you ask anyone

Observe first. Before any interview, look at what your systems already show:

  • Recurring calendar events. A weekly "payables review" or "schedule meeting" usually hides a workflow.
  • Inbox patterns. Who sends what to whom, and what arrives every week: supplier invoices, requests for quotes, timesheets. Subjects and senders are enough; you don't need to read the messages.
  • Recurring files. Templates, documents rebuilt every month, and spreadsheets used as databases.
  • The apps each role uses. Including desktop software and private web dashboards that no one thinks to mention.

Each pattern becomes a hypothesis about a workflow, not a conclusion.

Step 3: Confirm, don't interview

Turn each hypothesis into a question that takes seconds to answer:

"Every Monday you open the payables spreadsheet and then email it to the owner. Is that your weekly payables review?"

Save open questions for what data cannot show: decisions, exceptions, workarounds and the reasons behind them. When you do ask, ask "walk me through the last time you did this", not "how do you usually do it". The last real time includes the exceptions; "usually" leaves them out.

Keep conversations short, 5 to 15 minutes, and spread them over several days. People remember more when they have just done the task.

Step 4: Write each workflow down as steps

For every step, capture the same fields so steps can be compared later:

FieldWhat to write
TriggerWhat starts it: an email arrives, a date, a status change
InputsWhat the person needs in hand: a PDF, a price list, last month's file
Action and decisionWhat they do, and any choice they make along the way
ApprovalWho has to sign off, if anyone
Output and systemWhat comes out, and where it lives afterwards
Frequency and timeHow often, and minutes per occurrence
ExceptionsWhat goes differently, and how often

Then cross-check the handoffs. If Tom says he sends the estimate to Karen, Karen should describe receiving it. When two descriptions disagree, flag it rather than picking one. Anything you could not confirm stays marked unknown.

See the result on a real-looking example: our example workflow map follows a renovation company from quote request to paid invoice.

Step 5: Class every step: automate, assist or human

ClassUse it whenTypical example
AutomateThe step moves data or applies a clear ruleCopying invoice details from an email into the accounting system
AssistAI can draft or prepare it, but a person should check before it countsA follow-up email drafted from the quote thread, approved before it is sent
HumanIt needs judgment, a relationship or a signatureSetting the margin on an unusual job

Expect a real mix. A map where everything is "automate" was not looked at closely enough.

Step 6: Score value against feasibility

Value is volume × time × cost: how often the step happens, how long it takes, and what that time costs. Feasibility asks three questions: is the data already digital, does the system connect (an API, an export, an email-in address), and what would a mistake cost?

StepPer weekMinutes eachHours / yearData digital?Cost of a mistakeClass
Entering supplier invoices into the accounting system604192PDFs by emailMedium, caught at approvalAutomate
Drafting quote line items from a customer request1240384Partly (notes, photos)High, it sets the priceAssist
Compiling the weekly operations report19072SpreadsheetsLowAutomate
Deciding the margin on a complex job43096In the estimator’s headVery highHuman

Illustrative numbers. Hours per year = per week × minutes ÷ 60 × 48 working weeks. For example, 60 invoices × 4 minutes is 4 hours a week, or 192 hours a year.

The quote line items carry the most hours, but a mistake sets the wrong price, so they belong in Assist with an approval. The invoice entry is the cleaner first automation: high volume, digital data, a system that connects, and an approval step that catches errors. To turn hours into money for your own team, use the ROI calculator.

Step 7: Start with one workflow

Pick the opportunity with high value, high feasibility and a low cost of mistakes. Then decide, for every step of the automation, what level of human control it needs:

CheckpointWhat happensTypical example
AutoRuns on its ownFiling an invoice PDF in the right job folder
NotifyRuns, and a person is told"Progress invoice sent to the owner for the Main St job"
ApproveWaits until a named person approvesA payment run, or any message to a client
Human taskA person does the work; the automation waits, then continuesSigning a contract

Measure the workflow for a normal week before you change anything, and again after. Without the "before", nobody will agree on whether it worked.

Good and poor candidates at a glance

Good candidates

  • Happens every week, many times.
  • The same information is typed into a second system.
  • Inputs arrive digitally: email, PDF, form, spreadsheet.
  • A clear rule decides what happens next.
  • A mistake would be caught before it reaches a client or a bank.

Poor candidates, for now

  • Happens a few times a year.
  • Depends on judgment nobody has written down.
  • Lives on paper or in a system with no way in or out.
  • A mistake goes straight to a client, a regulator or a payment.

Doing it yourself or with us

You can run this method with a spreadsheet and a few weeks of the owner's attention. Our AI automation assessment runs the same seven steps for you, with the evidence attached to every step, and hands you the ranked plan. If your team will be part of the review, read employee privacy when you automate first.

Get started

Start with one workflow.

Book a free 20-minute build call. You'll leave with a clear view of what to automate first, and a proposal the same day.

Book a 20-min build call hello [at] intelligence33.com