Just talked with a new tech company called Vervus AI that is trying to utilize AI to change the entire process of a real estate appraisal.
This is not the “add-on” type stuff that we are hearing about that will help you determine (for example) what quality and condition to choose based on photo analysis, or that will help you choose the best comparables, or will run numbers to help you determine adjustments.
They are advertising that after they complete their setup process — analyzing some of your reports and working with you to understand your specific approach to an appraisal – a report can be completed in about 20 minutes (not counting the inspection).
I’m not surprised that someone is attempting this, or that it will be successful eventually. I recall having a discussion with an appraiser 15 years ago about the “art vs. science” argument. I asked him if he thought his process really differed so much on every report, or, if I had a team of PhD data scientists follow him around for a month or two, if they wouldn’t be able to replicate the way he went about completing an appraisal. I’ve seen way too many appraisals that still are adjusting $2,500 for a half bath regardless of the size/price of the house to know that most appraiser’s processes are absolutely replicable (even if they aren’t always accurate), myself included. Sure, I will do things differently for a complex property than I do for a production tract home, but from a general standpoint I’m looking to include proximate and recent sales that are of similar size and age, and then will add others as necessary to bracket features. So there is a general method to my comp selection. And I use the same process to gather data I will use to analyze features and determine adjustments; I figure out what I think is the reasonable market area, download that data, and use a custom spreadsheet to use several regression and group sales analysis methods to give me answers. Yes, there is still some of my own input into this; if I get data back that says an extra bathroom is worth $148,000 on a $300,000 house, I will check my data and analysis and probably use other methods to see if an adjustment is warranted and how much. But it is certainly easy enough to teach the AI what an over-the-top adjustment would look like and what to do next. The point is, this is all very possible theoretically.
“AI, we are going to create a new appraisal report. This will be an FHA purchase. Create the report based on my parameters for an FHA purchase report.”
“AI, get the tax record for (property address) and load the required information into the report.”
“AI, here is the sketch and photos I took on my site visit. Load them to the report.”
“AI, based on my sketch, photos, and tax record data, choose the 5 most comparable sales and 1 listing based on my standard comp selection process.”
“AI, using the large data set from the market area I downloaded from MLS, determine the proper adjustments to these comparables. Include charts and graphs for each adjustment showing how they were calculated, and write a description of which adjustments were made and how they were determined.”
“AI, convert this to a PDF and XML file.”
Here is the problem. At this point in time, AI isn’t reliable enough. I’ve been occasionally messing around with it for a year now, trying to figure out how to use it to shorten the time it takes me to complete a report. But even on simple reports, it usually takes more time than it is worth. Even with a list of prompts (which takes time to get right and aggregate) there are still way too many phantom references (check out Episode 191 at https://appraiserelearning.com/the-appraisal-update-podcast/), and incorrect data analysis (how can a computer be so bad at math sometimes??).
Additionally, I have read several reports that are saying that the “knowledge” available from these systems is plateauing and/or reaching a ceiling due to a lack of additional data to input. Much like we deal with in appraising, they are experiencing the theory of diminishing returns. I read one report where the author calculated that to significantly improve current AI systems one order of magnitude, we would need to feed them 3 times ALL OF THE AVAILABLE DATA CREATED SINCE MAN STARTED KEEPING WRITTEN RECORDS. Think about that; all of historical knowledge on every subject, and it not enough. In many ways, AI has a complete lack of understanding of the “Why” an appraisal is performed when it is responding to you.
We are calling the Large Language Models “AI” and it just isn’t really “intelligence”. We have to remember that ChatGPT and other LLM’s are at a very basic level just choosing the most appropriate next word to put into a sentence. As amazingly real — and occasionally helpful — as these conversations can be, it is not like you are talking to another knowledgeable appraiser and getting good advice. Much of the advice they have vacuumed up and are regurgitating to you may in fact be BAD advice.
That being said, it’s very easy to see the use of AI in appraisals coming. I’m not a Mathematics PhD, and neither are any of the appraisers I know. We work hard, we try to do our best, but a computer can always create something like an appraisal report faster and more accurately. (If you only knew how many words I have misspelled while writing this due to poor typing skills… 😊)
While it stinks from a professional standpoint, I was a fan of the “Property Data Collector” concept; have an appraiser — someone who understands what to look for — go measure a house and collect data and photos, and then turn that over to Fannie/Freddie and let them do all the math to determine a value.
Assuming they hadn’t immediately subcontracted that out to the lowest-cost providers regardless of knowledge or experience, that would have been great because,
(a) an appraiser could easily do +/-5 of those a day at $150 with no additional work needed, and
(b) would have no liability for the value since you are just inspecting it.
No more listening to realtors talk about why your price/SF is wrong, or a lender or borrower pleading “can we just get another $10k out of this?” “Sorry, you wanted it faster, so talk to the big Fannie computer about your value. Good luck!”
But AI appraisals will get here. My contact at Vervus Ai made the point that, “AI is about replacing labor costs. You can hire an assistant to type reports for you for maybe $30k/year and gain some efficiencies (assuming you have the volume right now which many don’t), or you can hire another appraiser at $___/year (or train them yourself at a high cost), or you can triple or quadruple your productivity with AI at minimal additional cost.
Their product starts at $1,500/month right now and I think there is also a charge for the initial setup, so if I were a single appraiser doing 7 reports a week (for example purposes let’s say at $500 each or $3,500/week) and could suddenly do 20 ($10,000/week), that is an easy ROI calculation.
My concern is the same as what I have historically called “appraisal mills” (like the guy in Colorado that is the subject of the FHA/Rocket Mortgage lawsuit, found here https://www.valuationreview.com/vr/articlesvr/appraiser-in-rocket-mortgage-lawsuit-has-license-r-93502.aspx). There will be some appraisers that just adopt it and kick out these reports with no regard to whether they are correct. They could have – based on current reports of AI outputs – completely made-up information; homes that don’t even exist. The AI may alter the original data to fit the narrative; “he likes a bracketing comp for garage spaces but I don’t have any, so will change Sale #4 to have 4-garage spaces!” The math and charts that the report contains may also not be correct.
The enforcement of good appraisal practice in this field is fairly minimal due to underfunded state agencies, so a bad actor could probably kick out enough work to retire on before they get caught, to the detriment of appraisers in that same market trying to do them the right way. And if I am doing appraisals in 20-30 minutes (not including the inspection), I might be willing to work for $250/report, which is also not good for the competitors not using the AI. It’s great for the large AMC’s which will still be charging the borrower $800 of course! And don’t think for a second that the AMC’s with staff appraisers aren’t going to be early adopters.
I also wonder about the data. Will companies like Vervus AI be sucking up all of your historical report data AND anything they gather from your MLS in order to resell it? I have no doubt that CoreLogic, which owns Alamode and has access to millions of completed appraisal reports, is using all of that data we have provided to enhance their own AVM’s and selling off that collected data.
I also wonder what this means for the forms providers like Alamode and Bradford and ACI. If my AI can recreate a government-mandated form AND populate it in the correct format, why do I need them?
It is a scary time in this field. I have traditionally counted on the fact that the government moves pretty slowly, and change would take a while for mortgage appraisals which make up the bulk of the business. The new UAD has taken the better part of 5 years just to design a form! And it’s not like the forms providers have ever really done anything exciting to their products so AI providing a better platform to complete an appraisal is low-hanging fruit. But right now if feels like we are doing an “Indiana Jones” where we are navigating an array of potentially lethal attacks from all sides.
I don’t know that appraisers can avoid the possibility of companies like the one I talked to; in order to remain competitive, we may have no choice but to use them so we can kick out reports in 30 minutes and accept a low fee to compete in the marketplace.
And remember, there will be a dozen companies like Vervus AI. Some will survive, some won’t. Will the one you choose make it, or will you need to start over in 18 months when the current one runs out of funding or gets bought out?
It is going to be an interesting couple of years while this shakes out…
What do you think? Leave me your thoughts on my LinkedIn page!