Why AI Comps Can't Replace Agent Judgement
By Matt Basedow
A $50 million deal nearly died over a chatbot conversation.
Not because the price was wrong. Because two people, on opposite sides of the same negotiation, asked an AI tool what the number should be, and got two confident, opposing answers.
That story went viral for a reason. It's not really about one deal. It's about what happens when everyone in a transaction has access to a market model, but nobody's checking whether that model understands the deal in front of it.
When a Chatbot Gets a Seat at the Table
Here's roughly what happened, as the agent involved told it. At the last minute, the seller asked ChatGPT whether the price was right. The tool told him no, the property was worth more than the offer on the table. Then the buyer did the same thing from his side, asking if he was overpaying. ChatGPT said yes, and backed it up with comparables (CNBCCNBC).
Two AI sessions. Two confident answers. Both were wrong for the same reason: neither one knew the buyer's intentions, the seller's emotional state, which other buyers were circling, or the off-market comparables that actually explained the number (CNBC).
The agent salvaged it by doing the thing AI couldn't do: explaining, to both sides, why the model's output wasn't the whole picture. His summary of it has become the line worth remembering: AI can model a market. It can't model a deal.
Every Buyer and Seller Now Has a Second Opinion in Their Pocket
This isn't a one-off. It's the new normal for every listing you run.
Buyers and sellers already have AI open in another tab before they've said a word to you. They're pricing your listing, sanity-checking your advice, and forming opinions before you've had the chance to add context. You're not just negotiating the deal anymore. You're often negotiating against a chatbot's homework.
AI can model a market. It can't model a deal.
Agents feel this tension already. RPR's 2026 survey of NAR members found that 63% of agents rank accuracy of AI outputs as their top concern, ahead of compliance issues and misreading the market. That's not agents rejecting AI. It's agents who've watched a client walk in with a number a model gave them, and had to do the work of explaining why the model missed something a human wouldn't have (HousingWire).
Where AI Actually Earns Its Place in a Listing
The mistake isn't using AI. It's asking AI to do the one job it was never built for: sitting in the room and reading what's actually going on.
AI is genuinely good at the parts of the job that are pattern matching. Pulling comparable sales. Spotting a trend across a suburb. Drafting a listing description in seconds. Turning a set of photos into a polished, branded video without booking a videographer or waiting three days for footage. That's the lane where AI adds real leverage, because it's handling volume and speed, not judgement calls.
It's a bad fit the moment the job requires context that a model can't see: what the seller actually needs from this sale, what the buyer is willing to stretch for, what's happening off-market that never made it into a dataset. That's agent territory. It always was.
This is exactly why PropertyVideos.ai stays out of pricing and negotiation entirely. It's built by videographers to solve one specific problem: agents needing a professional, branded listing video without the cost or lead time of a full production crew. It doesn't touch the deal. It helps you show up looking like the top agent in the market while you handle the part AI genuinely can't.
How to Use AI Comps Without Letting Them Run the Conversation
Treat AI pricing output as a starting point, not a verdict. Use it the way you'd use a rough CMA pulled by a junior team member: useful, unverified, yours to check.
Ask what data the model didn't have. Off-market activity, buyer intentions, and condition issues from a walkthrough. If it's not in the dataset, it's not in the answer.
Get ahead of it with clients. If you know a buyer or seller is likely to check a number against ChatGPT, address it before they bring it up. It builds trust instead of putting you on the back foot.
Keep the human judgement calls with you. Pricing strategy, negotiation tactics, and reading a counterparty's motivation are not tasks to delegate to a model, even a good one.
Common Questions
Will AI comps eventually get accurate enough to replace agent pricing advice?
They'll get more accurate at reading public data. They won't get access to the things that actually move a negotiation: what a seller needs by when, what a buyer has seen and rejected, and what's happening in conversations that never touch a listing portal. Accuracy on data isn't the same as judgement on a deal.
Should I stop clients from using AI to check pricing?
You can't, and trying to usually backfires. The better move is getting ahead of it: tell clients upfront that AI tools are useful for a general sense of the market but miss the specifics you're being paid to handle.
Isn't this the same argument every industry makes about AI replacing them?
Not quite. Nobody's disputing that AI can model comparables faster than a person. The Serhant story shows the actual failure mode: not "AI got the market wrong," but "AI got the deal wrong because it never had visibility into the deal." That's a different, narrower claim, and it holds up.
The Bottom Line
The lesson from a $50 million near-miss isn't that AI is dangerous or that agents should avoid it. It's that AI is extremely good at one thing and genuinely bad at another, and the two are easy to confuse when a client hands you a printout from a chatbot with total confidence.
Your job was never just producing a number. It's knowing what the number doesn't say. That's still true no matter how good the model gets.