Using AI for realtor cold outreach: what works and what gets flagged

AI makes it faster to write outreach. That's the upside. The downside is that it also makes it faster to produce emails that get flagged by spam filters, sound like every other AI-assisted drip sequence, and run afoul of CASL if you're not careful. The good news is that most of the problems are predictable. Once you know the patterns, they're easy to avoid.

The spam-trigger patterns AI tends to produce

Most realtors who hand a cold email task to an AI model and take the output as-is end up with the same set of problems.

The first one is urgency clustering. AI models are trained to sound helpful and persuasive. That tends to produce language like "don't miss this opportunity" or "reach out before it's too late" in the same email that opens with "I wanted to personally connect." Spam filters see that combination and score it up.

The second is vague personalization tokens. "Hi [First Name], I noticed your home has been on the market for a while" reads as a template because it is one. Recipients know it. Spam filters don't care, but your reply rate does.

The third is subject line structure. AI-generated subject lines tend to be either too polished ("A quick question about your property on Elm Street") or too vague ("Following up"). The first category sometimes triggers promotional-tab routing. The second category gets ignored.

The fourth, and the one people underestimate, is send velocity. AI doesn't help you think about cadence. A sequence that sends four emails in six days to someone who hasn't replied looks like a blast campaign, not a follow-up. Email service providers and spam filters flag it accordingly.

How CASL actually applies to realtor outreach

CASL governs commercial electronic messages. That covers email, and it covers SMS. If you're sending a message that has the purpose or effect of promoting a commercial activity, CASL applies.

The key question is consent. Express consent is the cleanest basis. The person signed something, clicked an opt-in, or otherwise clearly said "yes, contact me." Implied consent is narrower than most Ontario realtors realize. It exists in relationships where you've done business with someone in the last 24 months, or where the person has conspicuously published their contact information without a "no solicitation" note. It does not extend to a list of homeowners you scraped from property records.

AI doesn't change the consent requirement. What AI does is lower the friction of sending, which means the volume goes up, and the risk of non-compliant sending goes up with it. A tool that helps you send 50 emails a week instead of 5 is useful if your list is clean. It's a liability if your list has gaps.

If you're building outreach automation on GoHighLevel or FollowUp Boss, the compliance rules still sit with you, not the tool. Both platforms have CASL-related documentation worth reading if you're configuring automated sends.

The RECO layer

RECO's Code of Ethics requires registrants to be courteous, respectful, and honest in dealings with clients and the public. That's a broad standard that cold outreach can run into sideways.

Specific patterns that tend to cause problems: subject lines that imply familiarity ("Following up on our chat" when there was no chat), opening lines that suggest a relationship that doesn't exist, or language that implies you have specific knowledge of their financial situation. These are the things AI will produce if you prompt it to sound "warm and personal." The model doesn't know what's actually true about the relationship. You do.

I'd encourage any Ontario realtor building outreach sequences to spend 20 minutes on reco.on.ca before they automate anything, rather than after.

How to use AI to personalize without producing generic output

The output quality is almost entirely a function of input quality. If you give an AI model a specific observation, it produces specific copy. If you give it a category, it produces category copy.

Compare these two prompts:

"Write a cold email to a homeowner in Toronto whose listing expired."

versus

"Write a short, respectful follow-up to a homeowner whose listing expired in February in the Bloor West area after 60+ days on market. I want to acknowledge the experience of a stale listing without being presumptuous. Tone: direct, no fluff. Max 120 words. No urgency language."

The second prompt gives Claude or ChatGPT enough specificity to produce something that doesn't read like every other expired-listing drip in the market. The extra 45 seconds to write a tighter prompt is the work.

The pattern I see across most realtor outreach setups is that the AI is doing the personalization, when it should be doing the writing. You do the personalization by thinking carefully about what segment or situation you're addressing before you prompt. Then the model scales the writing, not the thinking.

A few practical inputs that consistently improve AI output for outreach:

• The specific trigger for the outreach (listing expired, ownership change, neighborhood sale, anniversary of purchase)
• The tone you want, stated explicitly ("not salesy", "peer to peer", "brief and warm")
• Word count or structural constraint ("two short paragraphs, no bullet points")
• What you do NOT want ("no urgency language", "don't assume I know them")

Building a sequence that doesn't look like a blast

Sequence structure matters almost as much as copy. A few things that tend to keep AI-assisted sequences from triggering filters or looking automated.

Send intervals with real gaps. Three to five business days between touches is a baseline. Two emails in two days to a cold contact is aggressive regardless of copy quality.

Vary the format across the sequence. Email one might be short and conversational. Email two might reference something specific that's changed (a new comparable sale, a market update). Email three might be the explicit ask. Same channel, different structure, different entry point.

Mix human and AI steps. Not everything in a sequence needs to be automated. A handwritten or personally composed note in the middle of an otherwise AI-assisted sequence tends to perform better precisely because it breaks the pattern.

And check your email infrastructure before you worry about copy. SPF, DKIM, and DMARC records configured correctly do more for deliverability than optimizing subject lines. Mailmodo has a deliverability guide that covers the technical basics.

What I'd actually do

If I were building a cold outreach system from scratch for a realtor practice today, the structure would be:

Start with consent. Build the list from people who have some legitimate touchpoint. Past clients, open house sign-ins, referrals, publicly listed contacts. Know which legal basis you're relying on for each segment.

Use AI for the drafting, not the sending. Claude or ChatGPT drafts the copy. You review it. A human sends or approves the send. Automated sequences are fine for contacts who have opted in and are expecting communication. Cold contacts warrant a closer review loop.

Prompt specifically. One prompt per segment, not one prompt for "cold email to homeowners." The more specific the context you give the model, the less generic the output.

Read the output before it goes. Not for every single send once you trust a template, but during the build phase, every single email. AI models produce plausible-sounding text that can still be wrong about the tone, the relationship, or the compliance details.

The tools make all of this faster. The judgment about what to send, to whom, and on what basis is still yours.

---

FAQ

Can Ontario realtors use AI to send cold emails? Yes, with conditions. Under CASL, sending commercial electronic messages requires consent. Express consent is the cleanest basis. Implied consent is narrower than most realtors assume. AI doesn't change the consent requirement. It just makes it easier to send at volume, which makes non-compliant sending riskier, not safer.

What spam-trigger patterns does AI tend to create in cold outreach? AI tends to cluster urgency language, use vague personalization tokens, produce subject lines that read as templates, and get sent on cadences that look like a blast rather than a follow-up. The patterns are predictable and fixable at the prompt level.

How do I personalize AI cold outreach without sounding generic? Give the model a specific observation before you ask for the draft. "This person listed in March and the listing expired" is a useful input. "Write a cold email to a homeowner" is not. The specificity of the input determines whether the output sounds researched or templated.

What does RECO say about unsolicited contact to homeowners? RECO's Code of Ethics requires registrants to be courteous, respectful, and honest. Unsolicited contact isn't prohibited outright, but misleading subject lines, implied familiarity, or misrepresenting the relationship to the recipient would likely conflict with that standard.

Is AI-generated outreach detectable by spam filters? Not reliably in 2026, and that's not the right question. Spam filters score on behavioral and structural signals. A well-structured, consent-based sequence drafted by AI scores fine. A poorly structured blast to a cold list scores poorly regardless of who wrote it.

What's the safest way to build a realtor outreach sequence with AI help? Start with a real consent basis. Use AI to draft, not to send. Personalize at the prompt level. Review every draft before it goes out. Set send cadences with gaps that reflect how a human would actually follow up.

---

Emma Pace — strategic marketing consultant, AI coach for realtors, keynote speaker. Realtor at Monstera Real Estate. Builds AI-operated marketing systems at emmapace.ca.