Train It Before You Use It

AI for a small business, from a real estate agent who teaches it — train the assistant on your business, mine what you already have, build one phase at a time, and the Canadian rules the interview doesn’t mention

A small-business guide to using AI as more than a search engine, built on Lauren Lucas’s October 2026 interview on RealDealCrew: training ChatGPT or Claude on your business, finding missed leads and customer questions, avoiding AI slop, the comment-to-DM funnel, and building in phases. Every claim checked, with what Canada’s anti-spam law (CASL) and privacy law add. Checked October 2026.
Author

Behzad Samadi

Published

October 28, 2026

A paper-craft shop counter under a striped awning: a laptop glowing warm, a stack of envelopes with a few marked by amber tabs, an open appointment book, a coffee, shelves of labelled boxes, three steps rising to the counter, and the five-dot AI glyph above. Title: Train It Before You Use It.

Most small-business owners use AI the way they use Google: type a question, read the answer, move on. Then they wonder why everything it writes sounds like everyone else. In an October 2026 interview on the RealDealCrew podcast, Lauren Lucas, a licensed real estate agent who now teaches AI to agents and investors, explains the difference between asking AI questions and putting it to work in a business. Her examples are from real estate. The lessons fit a plumber, a bookkeeper, a hair salon or a physiotherapy clinic just as well.

The interview: Lauren Lucas on RealDealCrew, 27 October 2026, 39 minutes. Watch on YouTube.

This article takes her advice, checks each claim against the vendors’ own documents, and adds what a Canadian business needs and the interview doesn’t mention: the anti-spam and privacy rules that apply the moment AI starts sending messages or reading your customers’ emails. The interview is also a pitch for her paid community and the host’s consulting, so the numbers in it are the speakers’ own.

NoteIn short
  1. Train it before you use it. Spend 15–30 minutes letting the AI interview you about your business, and give it samples of your own writing. Without context, it can only be generic.
  2. Start with what you already have. Unanswered emails, forgotten quotes, past customers and the questions people keep asking you are worth more than new leads, and AI is good at finding them.
  3. Don’t automate chaos. Fix the process first, then automate one step, look at the results, and let the next problem tell you what to automate next.
  4. Use your own voice and your own footage. Generic AI images and posts now cost trust: one dictionary made “AI slop” its 2025 word of the year.
  5. In Canada, automated follow-ups are regulated. Instagram and Facebook DMs, texts and emails all fall under CASL. A lead who only asked a question more than six months ago can’t be “reactivated” with a sales message, and even asking for permission by message counts as one.
  6. Expect about C$50–85 a month for one AI subscription and a DM tool, with free tiers for the rest. Most of these tools bill in US dollars, so the Canadian price moves with the exchange rate.

Train it like a new hire

Lucas’s central point: people get “very bland, very vague” answers because they never told the AI who they are, what they do, who they serve and where. The host’s analogy is the right one: you wouldn’t hire a new assistant and say “go market” without showing them how you work.

What to do, in 15 to 30 minutes:

  1. Let it interview you. Start a conversation with something like:

    I run a small renovation company in Barrie, Ontario. Before you help me
with anything, interview me about my business, my customers, how I talk
to them and what I want help with. Ask one question at a time. When you
have enough, write a one-page summary I can reuse.

  2. Give it samples of your own writing: a few emails you’re proud of, a quote, a social post, your website’s About page. It learns your tone from them.

  3. Let it ask before it guesses. Add “ask me if anything is unclear” to your requests. An assistant that asks questions learns faster than one that predicts what you want.

Where to save it so you don’t repeat yourself. Both major assistants can remember this. In Claude, the free plan already includes memory across conversations, account-wide instructions, and up to five Projects, each with its own instructions and files. ChatGPT has custom instructions, memory and projects too. Put the one-page summary in your instructions or a project, and every conversation starts from it. If you work from a folder on your computer, a short rules file does the same job; see Let It Open the Folder.

Learn it the fast way

Lucas spent 18 months learning from YouTube. Her advice to anyone starting now is to skip most of that: put the videos into a notebook tool and ask it questions instead. The tool she names, NotebookLM, was renamed Gemini Notebook by Google on 16 July 2026. Its free tier allows 100 notebooks of up to 50 sources each.

Two limits she doesn’t mention: for a YouTube video, the tool reads only the transcript, not what’s shown on screen, and the video must be public, captioned and at least 72 hours old. Ask it “I run a small cleaning company and know nothing about AI. From these videos, what are the five things I should learn first?” and you have a reading list.

Start with what you already have

The open loops in your inbox

Her first quick win: have the AI go through your email and texts for “open loops”: questions you never answered, quotes you never followed up, drafts you never sent, people who said “let me think about it”. This is business you already earned and let slip.

Two cautions she doesn’t give:

  • Connecting your inbox sends it to the AI company. Claude’s Gmail and Calendar connectors can read, and also send email and create events, asking your approval before each action by default. Under Canadian privacy law (PIPEDA) you stay responsible for customer information you hand to a service, including one outside Canada, and you should tell customers it may be processed abroad. Canada’s privacy commissioners advise using de-identified information in prompts where you can. If you only want a list of who to follow up, an export with the names removed is often enough.
  • Following up is regulated. See the anti-spam rules below: a sales follow-up to someone who only made an inquiry is allowed for six months, not forever.

Your existing customers

“If I have 100 people in my database who already know, like, and trust me, why am I over here chasing thousands and millions?” Lucas suggests having AI look at what you already record (which emails get opened, which posts get shared and saved, where your work actually came from) and tell you what to do more of. Her example: if half your business came from referrals, a thank-you event for past clients may beat another round of advertising.

When you ask an AI to analyse numbers, check that it ran code rather than writing a figure into the text. Claude, and OpenAI’s code tools, run calculations as real code in a sandbox and can show the steps; a number the model simply writes down is a prediction, not a calculation. Ask it to use code and show the formula, then check one row by hand.

The questions your customers actually ask

Most small-business marketing talks to everybody, “and in turn, they’re really not talking to anybody”. Lucas’s alternative: find the exact questions your customers ask, then answer them.

  • Your own inbox and phone notes are the best source: the questions you answer every week.
  • Reddit and local forums, where people ask what they won’t ask a business.
  • AnswerThePublic, which lists the questions people search for. Its paid plans start at US$20 a month; secondary sources say a free account gets three searches a day.

A renovation company that answers “what does a permit cost in our town?” and “can I stay home during a kitchen renovation?” in plain words is more useful, and more trusted, than one posting “just finished another project!”

Don’t automate chaos

The most important line in the interview: “automating chaos is just going to cause more chaos.” The host’s version: AI amplifies what’s good about a business, and what isn’t.

Before automating anything, write down how the job is done today: how a lead arrives, who answers, how a quote is sent, how a booking is confirmed. Ask the AI to find the gaps and duplicates in that description. Lucas calls this her biggest time-saver: having AI look at her current systems to “get as lean as possible” before adding anything.

Build in phases

Her funnel was not designed in a weekend. She describes it growing one problem at a time:

  1. She offered something free (a guide, a short webinar) in exchange for contact details. People took it, but nothing followed.
  2. So she added a link to book a call. People booked, but she didn’t know what they wanted.
  3. So she added a short form before the call. Now she knew, but was typing their answers into her customer database by hand.
  4. So she automated that.

“Every problem that you’re solving with AI should create another problem,” which is your next task. That is the method: automate one step, watch what happens, and let the result tell you what to do next.

Your own voice, not AI slop

Generic AI images and posts, all in the same style, are now common enough to have a name. Australia’s Macquarie Dictionary chose “AI slop” as its 2025 word of the year; Merriam-Webster chose plain “slop”. Lucas finds it is “hurting consumer trust”.

What she does instead:

  • Real footage, AI-written words. Short clips from her actual day, with one or two sentences of text on top, written by the AI in her voice from a real customer question, and a longer caption below.
  • Talk to the customer, not to your peers. Investors post for other investors; tradespeople post for other tradespeople. The person who hires you is a worried homeowner, a busy parent, someone dealing with an estate. Name their situation, with some empathy.
  • Be specific. Her example is a woman in her community in Massachusetts who posts answers to probate questions over a simple video of a staircase, and now closes deals with people across the US who inherited property in Massachusetts.

The comment-to-DM funnel

The automation Lucas credits most:

  1. Something free and useful, a “lead magnet”: a checklist (“five mistakes to avoid before…”), a short webinar, a local event.
  2. “Comment the word GUIDE” on a post, and a tool (she uses ManyChat) replies to the comment and sends the guide by direct message.
  3. The guide ends with a booking link, so people choose a time themselves.
  4. A short intake form before the meeting.
  5. The answers flow into a customer database (CRM), which handles the follow-up.

She says this one automation booked “over 200 consultations” in the past 12 months, and that 187 of those people showed up, because they had chosen to be there. Those are her figures; we couldn’t verify them.

What the platforms require, from Meta’s own developer rules:

  • The customer must start the conversation. An automated DM can only answer a comment or message.
  • One private reply to a comment, within 7 days, until the person replies; after that, a 24-hour window for most messages.
  • Bots must say they’re bots at the start of an automated conversation.

What it costs

The interview says “like 40, 50 bucks a month”, in US dollars, which is about C$56–70. Here is a realistic set-up priced on 28 October 2026. Most of these tools bill in US dollars; we convert at that day’s Bank of Canada rate (US$1 = C$1.4036). Prices are before GST/HST, and your card may add a foreign-exchange fee.

Tool Listed price In Canadian dollars
Claude Pro (one AI subscription) $20 a month, or $17 a month billed yearly about C$28, or C$24
ChatGPT’s paid plan (a second subscription, optional) widely reported at US$20 a month; not confirmed on OpenAI’s site about C$28
ManyChat (comment-to-DM) Free up to 25 contacts; Essential US$17; Pro US$39 a month; not confirmed on ManyChat’s site C$0; about C$24; about C$55
HubSpot CRM Free: 2 users, 1,000 contacts C$0
Cal.com or Calendly (booking) Free for one person C$0
Gemini Notebook Free: 100 notebooks, 50 sources each C$0
Set-up Per month
Claude Pro and ManyChat’s free plan (testing the idea) about C$28
Claude Pro and ManyChat Essential, free CRM and booking about C$52
Claude Pro and ManyChat Pro (which adds AI replies), free CRM and booking about C$83
Both AI subscriptions and ManyChat Pro about C$111

So the interview’s figure is about right for one AI subscription and a basic DM tool. Paying for both assistants, as the host does, takes it past C$100. ManyChat changed to pricing by “active contacts” in March 2026, so check how many people your posts will reach before choosing a plan. Claude’s price was read from its own pricing page; the ChatGPT and ManyChat figures come from their help pages as reported in search results, because both sites refused our automated checks. Confirm them at checkout.

Other quick wins from the interview

  • Turn your spreadsheet into a decision tool. Investors upload the spreadsheet they use to judge a deal and say “this is how I analyse a deal”; then they type in a new address and numbers and get yes or no. The same works for quoting a job or pricing a product. Make it run code (see above), and check its first few answers against your spreadsheet.
  • Connect it to your tools, once it knows you. With email and calendar connected, “email Jack and see if he’s free for coffee Friday at 2, and if he is, add it to my calendar” becomes one instruction. Only after the training step: an assistant that doesn’t know your business shouldn’t write to your customers.
  • Hire admin help before sales help. Her first hire was another agent, which “create[d] more admin work”. Free your own time first.

The rules the interview leaves out

Canada’s anti-spam law (CASL)

CASL covers any electronic message that encourages someone to buy from you: email, text messages, and direct messages on Instagram, Facebook or WhatsApp (the law covers “an instant messaging account”).

  • You need consent. Express consent (they opted in) lasts until they withdraw it. Implied consent is time-limited: six months after someone makes an inquiry or asks for a quote, two years after a purchase.
  • “Reactivating” old leads has limits. A person who only inquired more than six months ago and never bought has no implied consent left. A friendly “are you still looking?” that promotes your service is a commercial message, and so is a message asking for permission to keep in touch: the law treats a consent request as a commercial message too.
  • Every commercial message must identify you, give a way to contact you, and include an unsubscribe that works within 10 business days.
  • You must be able to prove consent, so keep a record of how and when each person gave it.
  • Penalties reach $1 million per violation for individuals and $10 million for businesses.

A request for a quote is an exception: sending the quote someone asked for needs no consent, though the identification and unsubscribe rules still apply. Past customers within two years, people who opted in, and inquiries within six months are the safe list to “reactivate”.

Privacy

  • Under PIPEDA you stay accountable for customer information you give an AI service or CRM, even one in another country. Read the vendor’s data terms, and say in your privacy notice that information may be processed outside Canada.
  • Strip what you don’t need. Canada’s privacy commissioners’ 2023 principles for generative AI say to prefer de-identified information in prompts.
  • In Quebec, Law 25 goes further: a privacy impact assessment and a written agreement before personal information leaves Quebec, and a named person responsible for it.

Keep a person in the loop

AI makes mistakes; Anthropic’s own help pages say not to rely on Claude “as a singular source of truth”. Anything that goes to a customer — a quote, a price, a promise — gets a human look first.

Is it urgent?

Lucas’s opening claim is that anyone in her industry without AI “by the end of 2027” will be like an agent without a cell phone, and will “never” catch up. That is an opinion, and a useful prompt rather than a deadline. The evidence in the interview points somewhere calmer: the gains come from small, boring steps — answering the emails you missed, knowing what customers ask, fixing one part of the process at a time. Those are worth doing whether or not the deadline is real.

A first month, one phase at a time

  • Week 1 — train it. Let the assistant interview you; save the summary as instructions or a project. Feed it five samples of your writing.
  • Week 2 — your own data. Export last quarter’s inquiries (names removed if you can), and ask it which ones were never answered and which questions came up most. Follow up with the ones CASL allows.
  • Week 3 — one piece of content. Answer your most common customer question in your own words, over your own footage or photo.
  • Week 4 — one automation. Pick the most tedious step in your process, write down how it works today, and automate only that. Watch it for a month before adding the next.

Sources

Notes recording what was read in full, what came only from search results, and what could not be checked (including every ChatGPT plan detail, because OpenAI’s pages refused automated access) are in the accompanying folder.