// Blog / Guide

Telling customers when AI is involved

Share

Canada's privacy commissioner made a point this autumn that is worth building a practice on: transparency is not achieved by making information exist. It has to be understandable, and it has to reach people at the moment they are making a decision. Measured against that, the standard approach, a sentence about artificial intelligence in paragraph nine of a privacy policy, is not disclosure. It is a paper trail.

There is no Canadian law today that requires you to label AI use in most circumstances. There is a consultation, a direction of travel, and a set of customer expectations moving faster than either. This is the practical version.

Start with the only distinction customers care about

Not which model, not whether it is generative. The question a customer is actually asking is: am I dealing with a person right now, and did a person check this?

That splits your AI use into three buckets, and they need very different treatment.

AI as a tool you used. You drafted an email with help, summarised a document, cleaned up your own writing. A human wrote, directed and checked it. This generally needs no disclosure, the same way you do not disclose using spellcheck or a template.

AI as the thing they are talking to. A chatbot, a voice agent, an automated reply. This always needs disclosure, immediately and unmistakably.

AI as the thing that decided or recorded something. An automated decision about their application, a transcript or clinical note, an automated action on their account. This needs disclosure and usually a route to a human.

Where disclosure actually belongs

In the first message, not the footer. If someone opens your website chat, the opening line says it is an automated assistant and how to reach a person. Not a badge in the corner, not a tooltip. The first thing they read.

On a call, in the first ten seconds. "Just so you know, I am an automated assistant" or "we use an AI tool to take notes on this call, is that alright" belongs before the conversation, not after. The second one is also where recording and consent rules start applying, which is a separate obligation from courtesy.

At the point of decision. If AI is involved in assessing an application, say so where they apply, not in the terms they accepted months ago.

The test is simple and worth applying honestly: if the customer found out later, would they feel told or would they feel caught? If the answer is caught, the disclosure is in the wrong place regardless of whether it technically exists.

Always offer the human

The single most common complaint about business AI is not that it exists, it is being trapped with it. A chatbot that cannot be escaped generates more anger than no chatbot at all, and the anger attaches to you rather than to the tool.

Make the route out visible from the first message, make it work during the hours you claim, and do not make people argue with a bot to earn it. For a small business this is also a competitive advantage, because reaching a human quickly is something you can actually offer and your larger competitors mostly cannot.

What does not need labelling, honestly

Some of the pressure to label everything is overcorrection, and over-disclosure has its own costs: it trains people to ignore the notice, and it makes ordinary tool use sound furtive.

You do not need to label a blog post you wrote with AI assistance and then edited, a summary you produced for your own use, an internal document, or a photo you tidied up. The underlying obligation is the same as it always was: whatever you publish, you are responsible for, and it needs to be accurate. AI changes how you produced it, not who answers for it. We cover the accuracy side in using AI for marketing content.

The lines not to cross

These are not disclosure questions, they are honesty questions, and they do not have a notice that fixes them.

  • Never fabricate a person. A testimonial, review, case study or quotation from a customer who does not exist is deceptive regardless of any disclaimer, and in Canada misleading representations carry real consequences under competition and advertising rules.
  • Never present an AI persona as a named human employee. Giving a bot a friendly name is fine. Giving it a fake headshot and a job title is not.
  • Never imply human review that did not happen. "Reviewed by our team" has to be true.
  • Do not use AI-generated imagery to depict your actual premises, staff or work. Illustrative is fine. Representational is a misrepresentation.

Three sentences you can copy

For a website chat, as the opening message: "Hi, I am an automated assistant. I can help with common questions, and you can reach a person any time by typing 'human' or emailing us."

For a call with note-taking: "Before we start, I use an AI tool to take notes on calls so I do not miss anything. It keeps a written summary. Are you comfortable with that?"

For a proposal or report where AI assisted: nothing. You wrote it, you checked it, you are accountable for it. If you feel the need to add a line, that feeling is usually telling you the review step was thinner than it should have been.

What this means for your business

Disclosure done properly is cheap, improves trust, and gets you ahead of a rule that is visibly coming: AI interaction was one of the five areas named when Canada consulted on AI transparency this autumn, which we covered when the submissions were published. Done badly, as a clause nobody reads, it protects you from nothing and irritates the people who eventually find it. The work is deciding which of your three buckets each use falls into, putting the notice where the decision happens, and always leaving the door to a human open.

Not sure where your AI disclosure belongs, or whether your chatbot leaves people a way out? We help small businesses get this right in a way that builds trust rather than covering themselves.

Talk to us

Related