The algorithm just got better at pricing. Your job did not change.

Two moves that made AVMs better

Two things happened a week apart in May 2026, and together they raise the floor on automated valuation.

On May 5, 2026, ATTOM launched an AI-first AVM. The headline claims are worth knowing precisely: a 2.9 percent median error, comp-free, built on 30 years of data. Comp-free is the interesting part. Most valuation work, human and machine, leans on comparable sales. An AVM that does not start from comps is a different design, and a lower median error than the rough-and-ready estimates buyers are used to.

On May 12, 2026, Clear Capital announced it acquired Restb.ai, a computer-vision firm, for valuation. Computer vision matters because it means a model looking at photos to read what a property actually looks like, not just its square footage and bed count. That is an attempt to close one of the oldest weaknesses in automated valuation: the model could not see the house.

You do not need to follow the vendor race. You need to absorb the result. The free or cheap number a client can pull is getting closer to defensible, which changes the conversation you walk into.

It helps to be honest about the size of the shift. A 2.9 percent median error is not the same as a 2.9 percent error on every home. Median means half the estimates are better and half are worse, and the worse half can be a lot worse on a property that is unusual. So the right read is not "the model is right now" and not "the model is still useless." It is "the model is good enough that clients will trust it, and uneven enough that you still have real work to do." Both halves of that sentence matter when you sit across from a seller.

The question you are about to hear more often

Here is how this shows up on a Tuesday. A seller, or a buyer, arrives with a number from an algorithm and asks the question every agent will hear more this year: the algorithm says X, why is your price different?

The wrong move is to dismiss the number. It is better than it used to be, the client knows that, and waving it away makes you look like the one protecting an old way of working. The other wrong move is to fold and match it, because then you have made yourself the slower, more expensive version of a free tool.

The right move is the middle one. You acknowledge the AVM is a real, improved input, then you do the thing it cannot. You explain, in plain terms, the specific reasons your number differs for this property. That is not a sales tactic. It is the actual work, and the better the models get, the more clearly you have to show it.

The three places an AVM still misses

A more accurate model is still a model. It is strong on the parts that live in data and weak on the parts that do not. There are three reliable gaps.

• Condition. An AVM, even one reading photos, does not walk the property. It does not feel the soft floor, smell the deferred maintenance, or see the renovation that photos flattered. Computer vision narrows this, as the Restb.ai acquisition shows, but a screen is not a walkthrough.
• Micro-location. Models work at the level of an address and a neighborhood. They are weak on the difference between the quiet side of the street and the side facing the bus stop, the unit with the view and the identical unit without it, the block that turns over fast and the one that does not. That granularity is local knowledge, not a dataset.
• Motivation. No model knows why this seller is selling. A relocation on a deadline, an estate that wants a clean close, a buyer who will pay a premium to stop searching: motivation moves price as much as the property does, and it is invisible to an algorithm.

When a client quotes the number, your answer lives in these three. Not "the computer is wrong," but "here is what it could not see on this specific deal." Said that way, you are not in a fight with the algorithm. You are completing it. The client keeps their faith in the number as a baseline and gains a reason to trust your adjustment on top of it, which is a far stronger position than asking them to throw the number out.

It also keeps you honest. Sometimes the model is close and your instinct to price higher is just hope. Reading the three gaps deliberately forces you to ask whether condition, micro-location, or motivation actually justify the gap, or whether you are talking yourself into a number the market will not pay. The same discipline that defends a price also stops you from chasing an overpriced listing that sits.

How to use the AVM without being replaced by it

The agents who lose to better tools are the ones who competed with them on their own ground. The ones who win use the tool and add the part it cannot.

Pull the AVM yourself before the listing appointment. Walk in already knowing the number the client will likely find, so you are never surprised by it. Then build your price as a story the number cannot tell: this is the model's read, here are the three specific things about this property and this moment that move it, and here is where I land and why. You have now used the AVM as input and kept yourself as the authority.

This is the same discipline that runs through every part of an AI-equipped practice, and it ties to a standing idea worth repeating: your judgment is the asset. The tool gets better; the asset is what you bring that the tool cannot. For the wider view of how this fits together, how real estate runs on AI lays out the pattern, and how real estate agents use Claude shows the daily-work side of it.

The skill under the number

Every year the automated number gets a little better, and every year the same lesson holds for the people who price property for a living. The advantage was never your access to data. The client can pull a number too, now more accurate than before. The advantage is reading that number, knowing where it fails, and defending a price with judgment the model does not have.

That skill survives every vendor announcement and every accuracy claim. If you want the structured version built for property professionals, The Leveraged Real Estate Series teaches how to use AI valuation as an input while keeping your judgment in front of it, and the two minute course quiz will point you to the right starting place.