AI marketing agents: what a small business can safely hand over

What the demo doesn't show you

The demo takes eight minutes. The agent reads your Google reviews, writes three flawless replies, lines up a post and follows up on two quote requests. Everything works. The question the demo never asks: how many times out of ten does it work when nobody is watching?

You'll see plenty of demos. The share of Canadian businesses using AI to produce their goods or services went from 6.1% to 19.2% in two years (Statistics Canada, 2026). And the word "agent" now gets slapped on everything: Gartner estimates only about 130 of the thousands of vendors selling "agentic AI" are the real thing (Gartner, 2025).

Most advice pushes you toward autonomy: "let it run your social media." That's climbing the ladder from the top. For a small business, the value of an AI marketing agent is in watching and drafting. Autonomy is earned one task at a time, by counting.

Chatbot, automation or agent: who picks the next step?

It comes down to who decides what happens next.

• A chatbot (ChatGPT, for example) answers when you write to it, then waits: it's the consultant you call.
• An automation does the same thing every time a trigger fires: "when a form comes in, send this email." It's the timer on your thermostat: reliable, but blind.
• An agent gets a goal and access to your tools, then picks its own steps: it reads your numbers, digs in, writes, checks, tries again. It's an employee you've handed a set of keys; Anthropic, the company behind the Claude models, describes agents as systems that "dynamically direct their own processes and tool usage" (Anthropic, 2024).

When a vendor pitches you, ask one question: "What does your tool decide without me, and where do I see what it decided?"

What an agent does well, and what it won't do for you

With the right guardrails, an agent is excellent at the work you keep pushing to 10 p.m.: watching your numbers and flagging what moved, writing the weekly report in plain words, drafting review replies in the customer's language, turning your photos into posts, sorting incoming requests.

What it can't do, however good it gets:

• Know what you didn't tell it. The broken hoist, the late supplier, the customer who's also your brother-in-law.
• Take responsibility. When Air Canada's chatbot misstated a fare policy, the tribunal said it "should be obvious to Air Canada that it is responsible for all the information on its website" and ordered it to pay $812 (BC Civil Resolution Tribunal, 2024).
• Make the call on prices and brand. When Anthropic let its own AI run a small shop in its offices, the AI sold items below cost, got talked into handing out discount codes and told customers to pay into an account it had made up (Anthropic, 2025).
• Deliver reliable facts unchecked. In a Salesforce benchmark, the best agents completed about 58% of sales and service tasks in a single exchange, but only about 35% once the conversation stretched on (Salesforce AI Research, 2025).

Those numbers will age. The lesson won't: an agent is strong at looking and preparing, weak at judging.

The four-rung ladder

Hand over tasks, not "the marketing," and put each one on a rung:

1. Watch: the agent reads and alerts you. Nothing leaves the business.
2. Draft: it writes; you fix and send.
3. Do on your OK: it prepares and sends once you click.
4. Do alone: it acts and tells you afterwards.

The eraser test sets the ceiling. If the agent gets it wrong, can you erase the mistake before a customer, an employee or your bank account sees it? If not, the task tops out at rung 3. Prices, discounts, promised turnaround times and ad budgets: rung 2, at most.

The rule of three sets the promotion. Count the runs you approved without changes, in a row. Statisticians know the catch: after n runs with no errors, all you can claim, with 95% confidence, is that errors happen fewer than 3 times in n (Archives of Disease in Childhood, 2005). Ten clean runs? It could still go wrong 3 times in 10. You need 30 to get below 1 in 10, and 100 to get below 3 in 100. Hence our thresholds: 30 clean runs for rung 3, 100 for rung 4. One mistake a customer sees, and the task drops a rung.

Task	Starts at	Ceiling
Tracking numbers and reviews	1 Watch	1 (its best role)
Weekly report	2 Draft	2 (too rare for 100 runs)
Tagging requests, never discarding any	2 Draft	4 (after 100 clean runs)
Replying to a positive review	2 Draft	3 (it's public)
Replying to a negative review	2 Draft	2
Posting on social media	2 Draft	3
Prices, discounts, deadlines, ad budget	1 Watch	2

So rung 4 is narrow: mostly internal tasks. And a slick demo proves little: METR researchers found that the tasks agents complete 8 times out of 10 are four to six times shorter than the ones they complete half the time. That length doubles roughly every seven months (METR, 2026), so revisit your ladder twice a year.

Example: an October week at a Saint-Jérôme auto shop

Take a fictional six-person garage in Saint-Jérôme, north of Montréal, in the thick of tire season: Quebec requires winter tires from December 1 to March 15 (Government of Quebec). The owner connects an agent to Search Console, Google Analytics, Google reviews and the shop's Facebook page. All numbers are made up.

• Monday. The report (rung 2) shows the shop's appearances in Google searches for "pose pneus d'hiver Saint-Jérôme" (winter tire installation) tripled in two weeks, while form requests fell from 14 to 6. The agent flags the gap without knowing why. The owner tries the form on his phone: it hasn't sent anything since a plugin update.
• Tuesday. Nine new reviews, two of them in English. Eight drafts are approved as written. The ninth, to a customer annoyed by the wait, offered a 10% discount nobody had authorized. Rewrite, and the counter goes back to zero.
• Wednesday. 42 quote requests sorted: 25 complete, 11 missing the tire size, 6 outside the area. The agent drafts the follow-up asking for the size.
• Thursday. It recommends raising the Facebook ad budget. Bad idea: the shop is booked solid until November 20, which it had no way of knowing.

The time math. Before: 90 minutes to pull the report together, 54 for reviews (9 × 6 min), 126 for sorting (42 × 3 min), or 270 minutes. With the agent: 15 minutes of reading, 17 for reviews (8 × 1 min + 9 for the rewrite), 21 for sorting (42 × 30 sec), 50 for alerts (10 minutes a day), or 103 minutes. Close to 3 hours saved a week.

The real payoff is elsewhere. The broken form was costing at least 8 requests a week; if half become a $120 appointment (an assumption), $480 was leaking out every week. Watching, not drafting, is what paid.

Four risks, and how to supervise

Asked what limits their use of AI, Canadian businesses most often cite cybersecurity or privacy (13.4% of them), then cost (10.6%) (Statistics Canada, 2026). Those are two of the four risks to manage.

1. Confident mistakes. Require a source for every number in the report and check two of them each week in the original tool.
2. Your voice. Give the agent a one-page sheet: five words you use, five you avoid, three replies you're proud of, what it may never promise, this week's capacity.
3. Personal information. Quebec's private-sector privacy act, as amended by Law 25, requires a privacy impact assessment before personal information is communicated outside Quebec, and for any project to acquire an information system that handles it (LégisQuébec, ss. 3.3 and 17). A tool hosted outside the province that reads your customers' emails is, in principle, covered. A decision about a person made solely by the machine must be disclosed to them (s. 12.1). Quebec's privacy regulator, the Commission d'accès à l'information, publishes an assessment guide (in French). Give the agent as little data as the task allows, and check your situation with a lawyer.
4. Creeping costs. Many tools bill by usage, like Intercom's Fin agent, listed from US$0.99 per outcome (Intercom). In peak season, the bill follows your traffic, not your budget. Set a monthly cap and count your supervision time.

Supervision takes 20 minutes a week: open the agent's log, re-read five actions at random, update the counters. Don't hand that job to a second agent: in the next phase of Anthropic's experiment, the AI "CEO" brought in to oversee the shop shared many of its blind spots, and one of the most useful changes was forcing the shop agent to follow procedures, like double-checking prices before quoting them (Anthropic, 2025).

An honest limit: below five requests a week and a few reviews a month, supervising an agent will likely cost more than it returns. A calendar reminder to check your numbers on Mondays will do.

Your plan for the next 7 days

1. List the ten marketing tasks in your week and how long each one takes.
2. Run each through the eraser test and write down its ceiling.
3. Start two tasks at rung 2: the Monday report and review replies.
4. Open a log: date, task, approved as is, edited or rejected.
5. Write the one-page sheet.
6. Ask your vendor, in writing, where your data is hosted.
7. Set a spending cap and a review date 30 days out.

Finally, get ready for the other side of the counter. In the United States, Google already calls businesses on behalf of its users to check prices and availability, a feature not offered in Canada for now (Google, 2026). A customer's agent will rely on what you've published: this week, make sure your hours, starting prices and turnaround times are clear on your website and your Google Business Profile.