Search for "marketing automation" and you get definitions written by software vendors, then ranked lists of tools. None of it shows what the thing looks like in a fifteen-person firm that sends one newsletter a month. And since every vendor added "AI" to its pricing page, the owner's question has changed: do you need an AI agent on top of the tool you already pay for, or is the tool already doing the job?
The answer comes down to one distinction. A newsletter tool reacts to gestures: a sign-up, a click, a date. An agent reads sentences. Whatever fires on a gesture, your tool already knows how to do, and most small businesses have never switched it on. Whatever arrives as text slips past every workflow.
The short answer, for the busy
Marketing automation for a small business starts in the newsletter tool it already pays for. Cyberimpact, Mailchimp and Brevo trigger emails on events: a form sign-up, being added to a group, a click, a birthday, a segment. A welcome email and a short series for new customers take no AI to set up, as long as you are on the right plan. What no workflow does is read text. An AI agent adds that layer: it reads replies to the newsletter and routes them, qualifies an inquiry from what the person actually wrote, drafts variants a human approves, and keeps the list clean by syncing it with your other software. Canada's Anti-Spam Legislation (CASL) does not change because an AI writes the message: consent you can prove, sender identification, and unsubscribe requests processed within 10 business days at most. The agent proposes; a person approves what goes out.
What your tool already does, and nobody switches on
In Cyberimpact, the Quebec email platform, a scenario is a path that starts on a trigger and moves forward through actions and conditions. The list of triggers in its documentation is telling: sign-up through a form, being added to a group, the contact's birthday, expiry of implied consent, a link click, an email open, a segment, a survey answer. The actions are just as plain: send an email, wait, add to a group or remove from one. Marketing automation is included only in the Plus and Pro plans.
Mailchimp calls the same thing "Marketing Automation Flows". Its comparison table caps the Essentials plan at four steps per flow and opens up to 200 flows on Standard and Premium. Brevo triggers a follow-up email on form sign-up from its Starter plan and keeps multi-step workflows for the Standard plan.
With those building blocks, three workflows cover most of what a service business needs: the welcome email, which restates the frequency promised on the sign-up form; a short series for new customers, triggered when someone joins the "customers" group; and the consent campaign, which we will come back to. None of them needs AI. If they are not running yet, start there, before paying for anything else.
Be wary of one trigger in particular: the open. When Mail Privacy Protection is on, Apple downloads an email's remote content in the background when the message arrives, not when the person reads it. An "opened, so follow up" workflow follows up with people who read nothing. Our piece on newsletter structure and frequency explains why clicks and replies are now the signals you can trust.
A trigger reacts to a gesture, not to a sentence
Read the trigger list again. Each one is an event the software observes by itself: a button, a date, a field. None of them looks at what a reply says. That is expected: a rules engine compares values; it does not understand a sentence.
In a service business, though, the useful information arrives as sentences. Someone replies to the newsletter to ask whether you also do commercial buildings. A contact writes that they changed jobs and gives a new address. A customer types "stop emailing me" instead of hunting for the link at the bottom. The "describe your project" field on the contact form tells you, in three lines, whether the inquiry is serious. To the newsletter tool, none of these messages exist: they land in an inbox that someone reads when they have time.
Vendors are adding AI, but on the sending side. Brevo includes a content generator that drafts subject lines and copy from its Starter plan, plus AI send-time optimization on Standard. Mailchimp offers Intuit Assist, which its pricing page describes as a beta available to certain Standard and Premium users, in select countries, in English only. For a Quebec business writing in French, that last detail settles it. These features help produce the message. None of them reads what comes back.
Four things an agent adds
An AI agent sits between the inbox, the website form, your CRM and the newsletter tool, and does four things no workflow can.
It reads replies and routes them. A question about a service goes to the right person with a draft answer. A change of address updates the record. A removal request becomes a real unsubscribe, through the tool's API, the same day. It is the mechanism behind the agent that triages incoming email, applied to what comes back from a campaign.
It qualifies an inquiry from what the person wrote. The form says "basement to redo, Laval, before summer": the agent pulls out the project type, the region and the timeline, and files the inquiry in the right CRM queue with a one-line note. It does not decide to drop anyone. It sorts; a human decides.
It drafts variants a human approves: three subject lines for the next issue, a paragraph rewritten for commercial clients, the French version of a send. Nothing goes out unreviewed, because the text carries your name.
It keeps the list clean by syncing it with your other software. This is the least glamorous job and the most useful one. Cyberimpact calculates when implied consent expires from the consent date entered at import: two years for an active customer, six months for an inquiry. With no date, the record shows "None" as the expiry, and the tool will let you keep writing indefinitely to someone whose consent may have lapsed. The real date, the last purchase, lives in your invoicing software. An agent that copies it over every week makes an existing automation work. Our piece on AI agents in sales prospecting applies the same logic to prospect lists.
A complete loop, for a service firm
Take a fifteen-person renovation company (a constructed example, not a client) that sends a monthly newsletter and gets its inquiries through the form on its website.
An inquiry comes in through the form, with a separate newsletter checkbox the person ticks themselves. The agent reads the project description, creates or updates the CRM record, classifies the inquiry and drafts a reply the estimator reviews before sending. If the box is ticked, the agent adds the contact to the "newsletter" group in Cyberimpact through the API, with express consent. From there the tool takes over: joining the group triggers the welcome scenario. Cyberimpact's own API documentation (Plus and Pro plans) suggests exactly this setup.
The next month the newsletter goes out, written by a person, with a subject line picked from the ones the agent proposed. Replies come back to a real inbox: the agent sends questions to the estimator, unsubscribes anyone who asks the same day, and flags dead addresses.
Once a week, the agent compares invoice dates in the accounting software with consent dates in the newsletter tool and corrects the ones that changed. Cyberimpact's scenario template for implied consent about to expire does the rest: it sends a series of emails starting 30 days before expiry to obtain express consent, and stops as soon as the person gives it. Timing matters, because the consent request is itself a commercial message: it has to go out while you are still allowed to write.
It all comes down to the division of labour: the tool runs whatever fires on a gesture, and the agent reads whatever is written. What goes out, a person signs off. For a business on Microsoft 365, the same question comes up between Power Automate and an AI agent.
What does not change because an AI writes it
CASL makes no distinction based on who drafted the message. The three obligations covered in our piece on CASL and your newsletter stand as they are: consent, express or implied, which you bear the burden of proving; identification of the business, with contact details valid for at least 60 days after sending; and an unsubscribe mechanism, with requests processed without delay and within 10 business days at most.
For an agent, that means three rules. It never creates consent: an address found in an email signature does not become a subscriber. It does not touch the proof, which the tool already records (date, source, method). And it processes faster, never worse. The CRTC's enforcement record shows that botched unsubscribes are what get expensive, and an agent that reads "stop emailing me" in a reply closes exactly that gap.
One more thing: the replies the agent reads contain personal information. Where the model processes them, and who can see them, gets settled before launch, not after.
Where to start
In this order. First switch on what your tool already does: welcome, new customers, consent campaign. Then check that your imported contacts do not show "None" as their consent expiry. Then open the inbox where newsletter replies land. If it is full of unanswered questions from the past three months, an agent will have something to read. If it is empty, the problem is upstream, in the newsletter itself. Those are the questions we start with, before any talk of an agent.
→ Tell us about your newsletter and what comes back
Written byXavier Peich