Key takeaways
- Automate the repetitive, rule-based work that sits between systems first, such as enquiry handling, scheduling and follow-up.
- Never automate the decisions that carry legal, financial or human consequences without a person approving them.
- Every agent needs an owner, a written scope, a log of what it did and a way to switch it off.
- Measure hours returned to the team and errors avoided, and stop any agent that is not clearly doing both.
Every owner has now been told that AI agents will run their business. Most have tried a tool, got a mixed result and gone back to how things were. The useful truth is narrower and more valuable. In a small team there is a specific set of tasks that agents do well today, a set they should never touch, and a way of running them that keeps you in control. Ali founded MyAutomation.ca, an automation company, in 2020, and this is the order we follow.
What to automate first
Start where the work is repetitive, rule-based and sits between two systems. These are the tasks a person does badly at five in the afternoon and an agent does the same way every time.
- Enquiry handling. Acknowledge every enquiry within minutes, ask the qualifying questions, book the appointment onto the calendar and log it in the pipeline.
- Scheduling and confirmations. Reminders the day before, rebooking when a customer cancels, a note to the coordinator when something needs a human.
- Follow-up on open quotes. The scheduled chase, in your voice, until the customer answers or asks to stop.
- Data movement. Copying what the field team recorded into the invoicing system, the CRM and the reporting page.
- Drafting. First drafts of procedures, job descriptions, customer replies and reports for a person to correct and send.
Each of these returns hours every week to people who have better things to do, and each is easy to check. Our AI agents and automation work for teams starts with this list and nothing else for the first month.
What never to automate without a person
Anything with a legal, financial or human consequence stays with a person at the point of decision. An agent can prepare the information and recommend. It does not decide.
- Hiring, firing, discipline and pay decisions. These involve employment law and people's lives, and your employment lawyer will tell you the same.
- Quotes outside the price book, refunds above a threshold and anything that changes a contract.
- Safety calls in the field. Whether a job is safe to proceed is a human judgement.
- Customer complaints that have escalated. The customer wants a person, and an agent that pretends to be one costs trust.
- Anything touching personal information beyond what your privacy obligations in Canada allow, which you check with your lawyer first.
How to run an agent responsibly
Every agent in the company has a named owner, a written scope of what it may and may not do, a log of every action it took that the owner can read, and a switch that turns it off in seconds. It is introduced to the team as a tool that belongs to a person, not as a colleague. Customers are told when they are dealing with an automated step where that matters.
Treat the first month of any agent as a trial with a person checking its work daily. After the trial, the checks move to weekly sampling and the log is reviewed in the operations meeting. An agent that cannot be explained in two sentences to a new employee is too complicated to run in a small team.
Measuring whether it is worth it
Two numbers decide whether an agent stays. Hours returned to the team each week, estimated honestly by the people who used to do the task, and errors avoided or caused, taken from the log. If an agent saves hours and causes errors, fix it or stop it. If it saves nothing measurable, stop it regardless of how clever it is.
Put the agents on the weekly page under the owner's name like any other process. That is how you avoid the drift where a company has twelve automations running and nobody can say what they do.
Where to start on Monday
Pick the one task your team complains about most that fits the first list. Write the rules for it in plain words. Build the simplest agent that applies those rules, run it beside the person for two weeks, and compare. Then move to the next task. Companies that do this steadily for a year end up with a small, understood set of agents doing the boring work, and a team that trusts them because they watched them being proven.
Our outsourced technology team and fractional CTO service build and run these for owners who want it done properly, including the logging and the switch-off that most tool vendors leave out.