I am a firm believer that someone will create a great AVM that will rapidly replace human input into the valuation process. As we’ve seen with advances like self-driving cars, computers are better at assessing risks, reacting to local conditions, and, in many cases, making judgement calls than humans are. That’s why I was so intrigued by Redfin’s recent AVM launch.

To test it out, I looked up the Redfin estimate of my parent’s 2 bedroom suburban Maryland home. Their AVM estimates the home’s value at $573, 414.

My parent’s house. Yes, that is an antennae. They didn’t get cable until 2009.

One thing I know with 100% certainty is that whenever they sell there home, it will not be  for $573,414. Maybe it will be for $573,000 or $573,500, but definitely not $573,414.

This highlights the perils of letting data scientists control a product that needs to appeal to homeowners. As Redfin proudly highlights, their AVM has the lowest median error rate of any AVM. That could absolutely be true, but it is also totally irrelevant to the average homeowner. The only error rate that matters to a homeowner is the error rate on the valuation on their own home, and every homeowner is going to assume a valuation that ends in $414 is incorrect. This basic flaw also counters the premise that this AVM will be useful as a lead gen tool – If I immediately assume that my home estimate is wrong, what do you think my likelihood is of inquiring further?

So what would a useful, consumer-driven AVM look like? The key insight is that an AVM should expedite the process of listing and selling a property, not provide the most accurate possible number. At the end of the day, only the market will dictate what a property is worth, so trying to generate the most accurate number is akin to calculating the “right” stock price for a company. It’s more important to be directionally correct, because you’re unlikely to be exactly right.

Based on our work with home sellers, I’ve identified 4 characteristics that would define a truly valuable and revolutionary AVM:

1. The AVM result has to realistic as a listing price

When it comes to listing a home, we all know that homeowners focus on one number: the listing price. As such, a useful AVM must deliver a realistic listing price, not just the minimum error rate. In my parent’s example, I would be much more intrigued by an AVM price that looked plausible ($575,000 or $569,900) rather than an extremely specific number that I would never list my home at.

2. The result must be accompanied by a natural language explanation

Redfin’s AVM currently provides a number with literally no explanation as to how it was developed. I can guess it has something to do with a $/Sq Ft, price trends in the neighborhoods, days on market data, local schools, etc, but that’s mostly because I’m in the industry. The average homeowner won’t have a clue where the number came from.

In a typical model, the agent would explain their listing price recommendation during a listing presentation. A valuable AVM has to simulate that experience in order to quickly instill consumer confidence in the number and drive more interest in the product. On this note, it seems that Redfin has a ways to go on their natural language abilities, since their description of my parent’s house calls it a “townhouse” twice (which it is clearly not from the picture above).

3. The result must account for the psychology of the purchase process

The difference between $601,000 and $599,000 may be tiny mathematically, but it’s huge psychologically, for the same reason that most retail products are priced ending with $0.99 rather than on the whole dollar. If an AVM result is to be used as an actual listing price, it needs to account for this factor in order to attract buyers, and also highlights why having natural language explanations are so important. If the “minimum error” result comes back as $610,000, a great AVM would identify whether trends indicate that number is likely to rise in the near term, in which case that might be the right listing price, or if prices in the area have stagnated, in which case the property should be listed at $599,900 to capture all those buyer who search for the “$500k – 599k” range in Zillow and would otherwise miss the property entirely.

4. The AVM should actually reduce the effort involved in the listing process

If I were to believe Redfin’s estimate of my parent’s home, what are my next steps? Call or email their local agent, who will then do a home valuation and recommend a listing price. The exact thing any other agent would do, AVM or not. The technology has saved me exactly 0 time as a consumer. Worse, what if the agent comes back with a recommendation of $550,000? I’ve already been primed by the $573,414 number, so any significant deviation is going to turn me off the agent.

A great AVM would replace the home valuation step and enable the homeowner to just move on to the next step towards listing without having to schedule a wasteful in-home listing presentation. Until an agency has come up with this solution, AVMs are much more prone to do harm than good both to consumers, who don’t trust the results, and agents, who are forced to contend with homeowners who have been primed by inaccurate information.