Twenty years after the AVM: why lenders still need a valuer

Finvise CEO Nikita Chirkov-Sakkej sits down with Mark Witherspoon, the person who brought automated valuations to the UK lending market, to talk data, risk and where valuations go next.

GuestMark Witherspoon HostNikita Chirkov-Sakkej 25 minVideo and audio October 2026
In this episode

From the first UK AVM to the desktop valuation

Twenty years ago, Mark Witherspoon helped build the UK's first automated valuation models, starting with a cooperative of lenders sharing their own surveyor data. Since then AVMs have become faster, cheaper and far more widely used.

So why do lenders still rely on a human valuer? Mark explains where AVMs perform well, where they fall short on the UK's varied housing stock, and why the skill, judgment and experience of a surveyor is still the hardest thing to replicate.

He also shares a bold prediction: within five years, around 80% of mortgage valuations will be done at a desktop by a human surveyor supported by an AVM.

80%

Desktop is the future

Mark expects around 80% of mortgage valuations to be done at a desktop, by a human surveyor supported by an AVM, within five years.

01

An AVM can still be wrong

Even with a high statistical confidence level, individual cases can be exposed to isolated errors. Interpretation still needs a human.

02

Humans and machines together

The real opportunity is the man-machine interface: AVM data combined with the skill, judgment and experience of a valuer.

03

Ranges, not single numbers

Estate agents asked for a range of values, not one figure, so they could explain to vendors where their home sat and why.

"The thing that is hardest to replicate is the skill, judgment and experience of a human being."
Mark Witherspoon, Ex-CEO, Hometrack, UKValuation and CoreLogic
Chapters

Jump to a topic

On the podcast

Guest and host

Mark Witherspoon
Guest

Mark Witherspoon

Ex-CEO, Hometrack, UKValuation and CoreLogic

Founding CEO of UKValuation (part of Countrywide Inc.), later CEO of Hometrack, and again at CoreLogic (now Cotality). Mark brought automated valuations to the UK lending market and now advises Finvise.

Nikita Chirkov-Sakkej
Host

Nikita Chirkov-Sakkej

CEO, Finvise

Nikita leads Finvise, the new generation valuation firm pairing RICS-qualified surveyors with proprietary technology to deliver Red Book reports in hours, not days.

Highlights

Short on time? Watch the clips

Read the full transcript

Edited lightly for reading: false starts and repeated words removed. Chapter times link to the video on YouTube.

0:00Introduction

Nikita: Hey, welcome to the podcast. I'm Nikita, CEO of Finvise, a new generation valuation firm. And today I'm joined by the person who brought automated valuations to the UK lending market, Mark Witherspoon. Mark is the former CEO of Hometrack, UKValuation and CoreLogic. Mark, great to have you here.

Mark: Thank you. It's good to be here.

0:32Twenty years after the AVM: the original problem

Nikita: So today's topic is twenty years after the AVM: why lenders still need a valuer. And I think it would be super helpful if you, Mark, could take us all the way back and tell us what was the problem you saw back in the day, and what were the challenges you were facing?

Mark: It's really interesting looking back, because in some respects the challenges haven't changed that much, but the solutions to those challenges absolutely have. And I think particularly in recent years there's been a lot of significant development, particularly in things like data availability.

1:30Where AVMs began: Countrywide and US tax records

Nikita: Amazing. And when it comes to the data, maybe you can elaborate a little bit more. Twenty years back, how did you even collect the data? It didn't even exist in the way we know it today, right?

Mark: Yeah, it's fascinating. So of course automated valuation models really originated in the US, and that was where I first came across them, whilst doing some work with Countrywide Inc., who at that time were the largest lender on the face of the planet. I think the singular advantage they had back in the day was that they had access to what's called the County Tax Record Office data, which was an annual appraisal, if you like, of the physical attributes of a property, much as we would recognise in the UK.

2:03Building the UK's first AVM data cooperative

Mark: So in fact, when we formed UKValuation as the first AVM provider in the UK, we had to go and find data. And bear in mind, at that time things like Land Registry data were not available. So what we did was create a cooperative with many of the mortgage lenders at that time, using the physical valuation data they were gathering from surveyors and others who were actually looking at property. That was a great start. It gave us an insight into which of a property's attributes were important, and what had the largest impact on value from a lender's perspective.

Mark: And then, through the chronology of the development of the market, Hometrack took that adoption one step further into the prime mortgage market, working directly with the lenders. What we were then able to do was contractually engage with lenders and say: the more data you can give us by way of a mortgage application, the better performance we can squeeze out of the AVM algorithms. That was a really good next step, and it led to many of the risk rules that are still applied today being built into, if you like, a sort of table format. How much is the property worth in terms of sale value? How much is the loan amount? Therefore, what's the loan to value ratio? Those kinds of things.

Mark: And it was really when we started doing the work with CoreLogic, now Cotality, that we began to develop a deeper and, I would say, better understanding of risk modelling. By that I mean we moved from a relatively static view, as at today's date this property is worth X, to a more dynamic view through time series data analysis. This property has changed in value over the last period, and therefore, being reasonably predictive, we could say how much that property might be worth, given static economic conditions, one, two, three or five years out. What happens is it becomes a virtuous circle, because the more you can predict, the more you want to ask new questions. There are new factors you want to take into account, which I'll come back to in a second.

4:50Zoopla and "how much is my house worth?"

Mark: What was very exciting, and I think one of the most significant developments, was when we launched Zoopla, the property portal here in the UK. One of the key planks of that was allowing consumers to have a view on how much their property might be worth. So there was a shift away from it being the preserve of the professionals, the surveyors, appraisers and estate agents, to putting it firmly in the consumer domain. The page on the site with some of the heaviest traffic was "How much is my house worth?" And it also allowed consumers to begin to develop an understanding of which factors and attributes affected value, making a property more or less attractive if they were thinking of bringing it to market.

Nikita: Fascinating.

5:52Land Registry and joining up property data

Mark: And of course, during that time, the big change we saw was the continuous improvement in access to data sources. Land Registry came online. You could take a licence to Land Registry data, albeit the attributes held by Land Registry are a relatively limited set: the address, whether it's a house or a flat, but not the subset, so whether it's a detached or a semi-detached house. So what then became the big requirement was being able to curate and bring together disparate data sources. Once you have that hook, the property address, you can begin to add things like environmental data, energy performance data and so on. So you begin to create a much broader view of what a property represents.

6:53What does AVM performance really mean?

Nikita: And I think you mentioned performance quite a few times, right? In the beginning, I guess performance was less superior to what we have today, and the definition of performance can vary as well. What does it actually mean for you? Is it speed, is it cost, is it consistency? What exactly do you think AVMs are doing well today?

Mark: Yeah, I think that's a key point. Yes, the algorithms have developed, and the developers themselves are continuously learning about what impact certain data points have on the accuracy and predictability of AVM outputs. But at the end of the day, it's quite simple. AVMs are accessible, they're relatively fast compared to manual processes, and they can be relatively inexpensive.

Mark: There are a couple of interesting factors in there, because typically an AVM cannot be manipulated by users. There are some versions, and in fact back in the early days of UKValuation we absolutely would allow qualified users, that is surveyors, to interact with the data informing the algorithms. Think of that as selecting or deselecting comparables, because a surveyor with local knowledge can spot that, although from a statistical standpoint a comparable looks like an outlier and should be taken out of the calculation, the reality is different. But any statistician would tell you that's almost a fruitless exercise, because if you have enough data, your output will be what is referred to as statistically valid.

9:47Why estate agents wanted a range, not a number

Mark: What became very interesting was that, as we developed AVMs, the number of use cases multiplied quite substantially. The initial uses had been how much is a portfolio worth at account level, or how much is an individual property worth for mortgage origination. We've talked about consumers. What then became just as interesting was the advent of valuation tools to help estate agents. Not that estate agents don't know how much a property is worth, but it allowed them to provide evidence to a prospective vendor and say: look, I have a valuation report from an AVM supplier. Here you can see we know about recent sales in the area, we know about properties currently on the market, and here is my guidance on what your property is worth.

Mark: Interestingly, estate agents did not want a spot value. They didn't want a single absolute number for a property's value. They wanted a range. And the reason for that was very sensible. They wanted some wiggle room to be able to explain to the vendor why, if that's the range of values for this property, their specific home was at the lower or upper end, because of the following factors. So it actually encouraged a dialogue between the agent and the vendor about what the property might achieve if it were brought to market.

Mark: And as I said earlier on dynamic data, the other thing AVMs do very well is identify market trends. It helps us start to ask some of the right questions. Why are property values declining in an area, or why are they improving? So we go and look for data evidence to explain what we're actually observing. It's the start of a journey.

10:53Where AVMs fall short in the UK

Nikita: Very interesting. And I would assume that if we still need a valuer today to sign the report, the AVM is falling short somewhere, right? There are some improvements that still need to be made. What do you think those improvements should be?

Mark: Yeah, it's fascinating. As I said earlier, the origins were in the US, where, simply because of history, we see far more homogeneity of housing stock compared to the UK. The UK's housing history goes through many interesting periods: the boom periods of housing, particularly around the late Victorian period into the turn of the 20th century, and then the post Second World War emergence of things like social housing. What it means is that in the UK we have a fairly heterogeneous housing stock.

Mark: AVMs do work better where they're dealing with properties that are typical of an area, in an area that is in and of itself homogeneous. The flip of that is that AVMs fall short if the subject property is unusual, either in its own right or atypical of a particular area. And maybe also where there's relatively thin data, so there are fewer comparables for the algorithms to work with. And indeed where markets are fast moving, in any direction, because there's sometimes a lag in the availability of data.

Mark: If we go back to what I was saying about Hometrack and surveyor data: by the time the surveying firm or lender has uploaded that data to the supplier, the supplier has loaded it into their database, and the database can be read by the AVM, you might have a lag of several weeks. So improvements in AVMs would come with better recency of data. And of course the one everybody is aware of is the lag on Land Registry data, where if you're using previous sales, you could be several months behind where the market is at the point you're actually trying to achieve a valuation.

13:23Can an AVM see your interior?

Nikita: Very interesting. I've been asked a few times: how can you do a valuation without leaving your office, using an AVM, when you're not even looking at what's happening inside my flat? How much money I invested in my interior, an amazing sofa, superb finishes. How can you do that without being physically in the property?

Mark: So this gets quite interesting. There are a couple of things around how AVMs work. AVMs don't just produce a number, they also produce a confidence score, which is a statistical expression of how confident the model itself is that it saw enough data and is familiar enough with what it's seen. Today we'd have to mention AI; it extends into AI. But essentially, how much information is available from which to compute that number? And to your point, one of the things an AVM typically wouldn't see is the interior condition of the property. That in itself is a little bit of an issue, and I'll talk later about some of the more recent developments where I think data helps enable the further development of AVMs, maybe in a slightly different context.

Mark: This also goes to one of the issues that sits around AVMs: there is no absolute regulatory framework underpinning AVM performance in the same way that a surveyor has one. There is an organisation called the European AVM Alliance that helps guide suppliers on best practice, for things like confidence levels, so that a user can compare suppliers A, B and C on a like-for-like basis. If you've got different constructions of confidence levels, that comparison in itself can be quite a challenge.

15:24Why lenders still need a human valuer

Nikita: Interesting, 100%. The next question I wanted to ask you: why do lenders still need a valuer? Why do we still rely on a human being to be there and put a signature on it? If we're talking about secured lending, banks, other financial institutions and specialist lenders may actually hedge the risk by offering different LTVs or other financial instruments around it. So why are they not relying only on AVMs with a high confidence level?

Mark: So let's address the last point first. Any statistician will tell you that even with a high statistical confidence level, an AVM can still be wrong. Statistically, if you looked at lots of outputs, it would be right, but on individual cases you can still be exposed to isolated errors. So to answer the question: even twenty, twenty-five years on in the development of AVMs in the UK, the thing that is hardest to replicate is the skill, judgment and experience of a human being. And by that I mean the interpretation of data and information.

Mark: Where that gets very exciting is the man-machine interface. What if I was able to distil down the best of both worlds: have the speed of data, the ability to process a vast amount of data, and present it with a proposed output? "Given everything I've seen," says the model, "here's where I think this should be." Then the professional user, the human user at a desktop, can say: "Yes, that's absolutely spot on." Or: "Nice try, but the data missing from your observations was this." And that interaction also asks: if data was missing, where do I get it from? How do I replicate the knowledge the human brought into that interaction? So it's a continuously developing process.

Mark: There are, by the way, a number of well proven processes by which AVM accuracy can be improved in isolation. One is using multiple AVMs, particularly where they have uncorrelated data sets, by which I mean different data sets, or data sets curated in different ways, or where the computation of value uses different modelling techniques. If you brought three AVMs together and they all aligned around a very similar number, you could be pretty confident you're in the right place.

18:42Imagery, 3D data and property condition

Mark: Just picking up on one of the points from earlier: what the human has brought to the physical valuation process is also the observation of physical characteristics. We talked about condition. There are now processes available where, through imagery, either still photographs or more likely video, you can extract data such as condition from those images. What I find mildly amusing is that when we talk about property, we typically talk about the square footage or square metreage of floor plates. We very rarely talk about things like ceiling heights or the volume of a room. But when you have 3D image data, you can very accurately calculate things like window reveals, doorways and ceiling heights. I've seen a lot of work done on things like energy performance reports, where that data is becoming more and more reliable, such that you could actually bring it into the desktop valuation process.

19:57Where valuations will be in five years

Nikita: And I think this brings us to the point where I'm going to ask you to predict the future. Where are we heading now? You mentioned that technology is improving, and that we now use a human in the loop, as they say in AI today. So the machines are learning from humans, and humans are learning from the machines. But what's next? Where do you see us in five years from now? It may be impossible to predict, but it would be very interesting to hear your thoughts.

Mark: Yeah, I could just ask one of the AI tools. No, I'd better give you a better answer. Here's the thing. Funnily enough, it goes back to a slide I had in a deck more than 20 years ago, where we looked at the throughput of valuation instructions, at that time typically to a surveyor. What we were trying to do was estimate what percentage of valuations would go to each of the processes of the time: AVM alone, AVM into a surveyor environment, and surveyor only.

Mark: At the end of the day, if it's a converted windmill in a very rural area, there's no AVM that is ever going to get close to an accurate valuation of that asset. We know from the get-go, as soon as I describe that property, that's a surveyor, that's a physical visit to go and look at it. And what I find fascinating is that my prediction now is that the distribution of valuation techniques will shift very, very significantly in the next five years, to the point where something around 80% of all mortgage valuations will be done at a desktop by a human surveyor supported by an AVM.

Mark: I think we're literally talking about the tails: those very unusual properties at one end, and the very upper end of the market at the other, where we're talking about very expensive, unique properties. Those kinds of assets do require a physical inspection to gather all the information available, and there would be very few, if any, comparables for them. But I think what would become true is that, across the vast majority of market activity in the valuation space, and let's go beyond mortgage valuations now, physical inspections on a portfolio of assets will also become relatively rare. It would be almost a sampling exercise, with the vast majority being modelled at a desktop.

22:57How Finvise combines AVMs, desktop and surveyors

Nikita: Amazing. Well, I'm very glad to say we are sharing the same vision, right? That's exactly what we are building at Finvise today. We have multiple products. We have an AVM that we've been training for quite some time now, and we're very proud of where it is. We need more data, and we're putting more resources into training it and making it better. Then we have desktop valuations, as Mark mentioned. So we're creating a triage system: when we get an enquiry from a client to assess a property, we filter and understand whether it's a very simple asset, say a new build in central London with a lot of comparables. That's a cookie cutter, and it's relatively easy to assign the number.

Nikita: Once a property is a little more complicated, we engage new technology and send our talent to do walkarounds. Here we're trying to do it as fast as possible, but without compromising on quality. The floor plan is very important, so you need laser measurement. At the end of the day, these are all the building blocks of the RICS reports we produce and sign off. And the third element, as you mentioned, is a very unique asset, like Buckingham Palace in London. There's a very good excuse to go there, visit and spend a good amount of time valuing it.

Nikita: So we're trying to find the right balance. There's no one silver bullet. We're big believers in technology, and from day one Finvise has always been a tech business. But tech sometimes doesn't solve a real problem on its own, and that's why we have an amazing valuation team who have been doing nothing else but valuing properties for fifteen, twenty years plus. We're pairing them with our tech talent, and together they create the solution you alluded to earlier, Mark, working together on assigning the fair value of the asset.

Nikita: I really enjoyed talking to you, Mark. Thank you very much for taking the time. Very, very interesting.

Mark: Thank you.

Don't let valuations hold the deal.

RICS-regulated valuations for UK lenders and brokers, in hours, not days.

Book a call →