What does "automate a business with AI" actually mean?
It means a system does the office work a person is doing by copying, chasing and remembering, and AI is used only for the steps that need reading or writing. A rule can move a web form into a CRM; a rule cannot read a customer's email and work out which job they mean, pull the fields off a scanned document, or draft a reply in your voice. Those steps use a language model. Everything else stays a plain rule, because rules do not have off days. If someone sells you "AI" for a step a rule could do, you are paying for a word.
Step 1: which job should you automate first?
The one the money arrives through. For most owner-run businesses that is a request by email, a form or a phone call that a person has to notice, answer and type into something. Pick by three signs: it arrives often enough to count (twenty a week or more), it is handled the same way most of the time, and it costs something when it slips.
Do not start with the hardest job or the one with the most systems. Start with the one you can measure.
Step 2: how do you measure it?
Count a week. For the job you picked, count how many came in, how many got an answer, how long the answer took, how many were answered twice, and how many were never answered. A spreadsheet and a patient afternoon with the sent folder will do it; so will a free discovery, which reads one shared mailbox for a week and sends back one page with those numbers. The point is a baseline. Without one you will never know whether the automation worked, and you will not be able to tell a vendor what you need.
Step 3: can you write the rule in one sentence?
If you cannot, stop here. The sentence is: "When X arrives, do Y, unless Z, and show W to whom." For example: "When a quote request arrives, create one card with the customer, the job and the date, draft a reply from our price list, unless the customer is already quoted this week, and show the card to the person whose name is on it." If three people in your office write three different sentences, the job is not ready. Agree the sentence first. Most of the value of automating is in this step, and it costs nothing.
Step 4: how do you build it without losing control?
In your own accounts, in shadow mode. Your own accounts means the automation tool, the email access and any code sit under logins you hold, so you can revoke access in a click and nothing stops working if the person who built it disappears. Shadow mode means the system reads and drafts but a person sends: every reply, chase and invoice is a draft until you say otherwise. Choose the tool for the job rather than the tool you have heard of: a visual tool such as Zapier or Make for a simple chain between common systems, n8n when you want it on your own server, and plain code with a language model behind it for the steps that read a PDF or write an email. One mailbox to begin with, not all of them.
Step 5: when do you switch it to automatic?
After a month of drafts being right, one email type at a time. Read the drafts for the first week every day, then a sample. Count how many you changed before sending and why. When the change rate for one type of message, say the booking confirmation, is close to zero for a month, switch that type on and leave the rest as drafts. Quotes and anything with a price are usually the last to go automatic, and in some businesses they never do, which is fine; the time was saved in the drafting, not the sending.
Step 6: what comes after the first job?
The next job, and one page every morning. The morning page lists what came in, what was answered, what is still waiting with a name beside it, and what was missed. Once the first job is stable, add the one that touches it next: quotes, then the paperwork for the jobs you won, then the invoice, then the review request after it is paid. Each is a separate build with its own sentence from step 3. Businesses that try to automate everything in one project usually switch it all off.
Worked example: a trucking office
Assumptions, not a client: three dispatchers, one shared dispatch inbox, about 40 rate requests a day from brokers, three a day get no reply and two a day get two replies. Step 1 picks rate requests. Step 2 counts the week: 200 requests, 185 answered, 15 missed, 10 answered twice, median reply 48 minutes. Step 3's sentence: "When a rate request arrives, make one card with the broker, lane, dates and equipment, put the dispatcher who covers that lane on it, draft a reply with our lane rate filled in, unless the same load was quoted in the last 48 hours, and show the owner the misses at 7 a.m." Step 4 builds that beside the transport software, reading only the dispatch inbox. Step 5 keeps the dispatcher on the send button, because a rate is a negotiation. Step 6 adds the proof-of-delivery chase. The arithmetic on the 15 misses: win one in four, $350 of margin a load, about $1,300 a week, or roughly $65,000 a year over 50 working weeks, from requests that were already in the inbox.
Worked example: a tax office
Assumptions, not a client: a two-partner practice with 300 personal tax clients and one shared office inbox. Step 1 picks document collection, because March and April are lost to it. Step 2 counts a week in March: 140 client emails with attachments, 60 "did you get my T4?" questions, 45 chases sent by hand. Step 3's sentence: "When a client sends a document, file it to their engagement and update the checklist; when a client asks what is missing, draft the answer from the checklist; when a client is still short three days after the date we set, send the chase in our words, unless a partner has paused that client, and show the partners who is still short every morning." Step 4 builds it inside the firm's own Microsoft 365, reading one mailbox. Step 5 lets the first chase go automatic after a month and keeps replies to questions as drafts. The arithmetic: 45 chases a week at four minutes each is three hours a week of a senior's time in the eight busiest weeks of the year, before the phone calls that chase the chases.
What should you not automate?
Decisions, negotiations and anything nobody can describe. The dispatcher choosing which driver takes the load, the partner deciding how to treat a transaction, the owner deciding whether to take a job at that price: those stay with people, and the automation should hand them the facts and get out of the way. Do not automate a job that arrives fewer than twenty times a week; the build costs more than the job. Do not automate a complaint; a person answers those. And sometimes the fix is a process change, not a system: three personal inboxes merged into one shared inbox removes half the duplicates before any software is bought. A good vendor will tell you that in discovery and lose the sale.
Should you do it yourself or hire it out?
Do it yourself when the chain is simple, the systems are common, and someone in the office has a few hours a week to maintain it; a web form into a CRM with a confirmation email is a Saturday with Zapier. Hire it out when a step has to read or write (an email, a PDF, a messy spreadsheet), when more than two systems are involved, or when nobody has the hours to fix it the day a supplier changes their invoice layout. Either way, insist on owning the accounts and on a one-page handover written for your office manager, not a developer.
Where does Apna Automation fit?
We do steps 2 to 6 for owner-run businesses in the Toronto area and across Canada, inside the Microsoft 365 or Google Workspace you already pay for. Discovery is free and produces the week's count from step 2. Every build is quoted after it, runs in shadow mode first, and the first automation is live in five business days from the day access is granted. What it costs in the wider market, with sources, is in the cost post.