Development appraisal for the AI era
The residual method hasn't changed. What changes is who checks the inputs: spreadsheets, appraisal software, and what an AI agent adds to an appraisal.
Every site looks promising until you run the numbers. A development appraisal is where a scheme stops being a picture in your head and becomes a column of figures: what the finished homes will sell for, what it will cost to build them, and what is left over for the land and for you.
The arithmetic of that column has not changed in decades. Gross development value, minus every cost, minus the profit a lender will insist on, leaves the residual. It is taught on the first day of every surveying course, and it fits on a single sheet.
So why do appraisals still go wrong so often, and so expensively?
Because the arithmetic was never the hard part. The hard part is the dozen numbers you feed it, and where each one comes from. A residual appraisal is a lever: a small error in the value of the finished homes, or in the cost of building them, moves the land value a very long way. A model is only as honest as its least-evidenced input.
This piece sets out how the traditional appraisal works, what spreadsheets and appraisal software did and did not change, and what an appraisal looks like when an AI agent gathers and argues the evidence behind every line, while deterministic calculators still do the sums.
1. The traditional appraisal, in one page
If you want the conventional method laid out cleanly, LandTech's Development Appraisals: A Complete Guide is a good traditional reference, and our own step-by-step guide to the residual method covers the same ground. The short version follows.
A development appraisal compares a scheme's expected income with its expected costs, to test whether a site is commercially viable. Every appraisal is built from the same handful of inputs:
- Gross development value (GDV): what the completed scheme is worth once every unit is sold or let. It sits at the top and drives everything below it. See how to calculate GDV.
- Land cost: the price of the site plus acquisition costs: Stamp Duty Land Tax, legal and agent's fees. In a residual appraisal, this is the unknown.
- Build costs: usually the largest line, estimated per square metre early on and firmed up later by a quantity surveyor's cost plan.
- Professional fees: architects, engineers, quantity surveyors, planning consultants, often taken as a percentage of build costs.
- Planning costs: the application, supporting reports, Section 106 obligations and the Community Infrastructure Levy, which vary enormously between authorities.
- Finance: interest and fees on development finance over the build and sales period. Timing matters as much as the rate, because interest accrues as money is drawn.
- Contingency: a buffer for what goes wrong, higher for conversions and riskier schemes.
- Sales and marketing: agent's fees, legal fees on sales, marketing.
- Developer's profit: treated as a cost, usually 15 to 20% of GDV or a similar margin on cost, because no lender backs a scheme with no room for error.
The same figures can be run two ways, depending on the question:
- The residual method asks what you can pay for the land: GDV minus all costs, including your target profit, leaves the residual land value.
- The profit method asks what a scheme returns at a known land price: GDV minus all costs, including the land, leaves the profit.
Schemes held for rent rather than sold, such as build-to-rent or commercial space, take their GDV from rental income capitalised at an investment yield. That makes the yield one of the biggest levers in the whole appraisal.
None of this is controversial. It is the agreed method, and it is the right one.
2. Spreadsheets vs appraisal software: the pre-AI era
For most of the last thirty years, the debate about appraisals has been a debate about tools.
Spreadsheets are flexible, familiar and free if you already own the software. You control every formula and can shape the model to any scheme. Their weaknesses show at scale: building a reliable model takes time, a broken formula can sit unnoticed for months, versions multiply as files are emailed around, and you source every build cost and comparable yourself.
Appraisal software trades some flexibility for speed and consistency. The calculations are built in, so formula errors fall away. Cost benchmarks and comparables can be pulled in rather than looked up. Reports come out in a layout lenders recognise.
Both are real improvements, and the traditional guides are right to say the choice depends on how many sites you appraise. But notice what the debate is about. It is entirely about the arithmetic and the presentation: fewer broken formulas, faster benchmarks, tidier reports.
Neither tool asks whether the inputs are true.
A spreadsheet will multiply an optimistic sale price by ten houses without complaint. Appraisal software will apply a benchmark build cost to a site with a sloping plot, a party wall and a contaminated corner, and format the result beautifully. The tool computes what you tell it. Judging whether what you told it is right is still entirely your job, and it is where most appraisals fail.
3. Where appraisals actually go wrong: the inputs
The residual is what is left at the bottom, so it absorbs every error above it. A worked example shows how hard that bites.
Take a scheme of eight houses. The figures are illustrative, chosen to show the mechanics rather than any real site:
- GDV: 8 houses at £400,000 = £3,200,000
- Build costs: 8,000 sq ft at £200 per sq ft = £1,600,000
- Professional fees: 10% of build costs = £160,000
- Contingency: 5% of build costs = £80,000
- Planning, Section 106 and CIL: £120,000
- Sales and marketing: 2% of GDV = £64,000
- Finance: £140,000
- Developer's profit: 20% of GDV = £640,000
- Total costs, excluding land: £2,804,000
- Residual land value: £396,000, before Stamp Duty and acquisition fees
Now move one input at a time:
- Sale prices 5% lower: GDV falls to £3,040,000, the profit and sales lines follow it down, and the residual drops to about £271,000. A 5% miss on values has taken almost a third off what you can pay for the land.
- Build costs 10% higher: with fees and contingency following, the residual falls to about £212,000, close to half.
Those are not exotic errors. A 5% optimism on new-build values, or a 10% overrun on build costs, are ordinary outcomes. And each input has its own way of being wrong:
- GDV is often taken from asking prices, from the best sale on the street rather than the typical one, or from new-build premiums that disappear on resale. The street has a ceiling, and a scheme priced above it needs a reason.
- Build costs taken from a national benchmark miss the region, the site and the specification. A rate per square foot cannot see a basement, a difficult access or a heritage frontage.
- Planning costs depend on the local CIL charging schedule, the authority's Section 106 practice and the affordable housing policy, all of which sit in documents, not databases.
- The site itself: floor areas, boundaries, flood zones, conservation areas, Article 4 directions, tree preservation orders and the planning history of the site and its neighbours each change what can be built and what it is worth.
The evidence for every one of these exists. It is spread across sold-price registers, planning portals, charging schedules, energy certificates, title registers, listings and cost data. Gathering it, reading it and reconciling it is what takes a good appraiser days. Skipping it is what makes a bad appraisal fast.
4. What changes in the AI era
Until recently, the only thing that could do the evidence work was a person. Calculators could do the sums; nothing else could read a planning decision, notice that a comparable was a repossession, or ask why a floor area on one certificate did not match another.
Large language models change that. An AI agent can be given the method an experienced appraiser follows, a set of data tools and a set of calculators, and asked to work through a site step by step, writing down what it found and why it believes it. The important design choice is the division of labour:
- The agent does the evidence and the judgement. It establishes what the site and the scheme are, finds the comparable sales, reads the planning record, judges condition from photos and floor plans, and decides what each input should be.
- Deterministic calculators do the arithmetic. The agent never does money in its head. Costs come from a priced cost plan with regional factors, tender-price indexation, preliminaries, contingency, overheads and VAT item by item. Tax comes from the actual Stamp Duty or Land Transaction Tax rules. Returns, finance, refinance and the maximum offer come from a returns calculator. The report's figures are copied from the calculators' outputs, never retyped.
- Every figure carries a receipt. Each query, each calculation and each document read is recorded with an identifier. A number in the report that does not trace to a receipt is refused.
That last point is what separates an agentic appraisal from a chatbot with a spreadsheet. The appraisal is not a paragraph of plausible prose. It is a chain of evidence, each link checkable.
Concretely, here is how each input changes:
- GDV starts from a fixed comparable set. Every nearby sale of the same form, filtered the same way every time: repossessions and bulk transfers out, quick flips out, developer first sales out, leases granted at sale out. Each sale is indexed to today and scaled for size by the same rule. The agent's job is judgement on top of that base, not a fresh search each time, so two runs on the same site start from the same evidence.
- The street has a ceiling, enforced. A finished value above the top of the local market is refused unless the agent names the reason, for example an extension that adds a bedroom or the best plot on the street. Optimism has to be argued, not slipped in.
- The site is identified before it is valued. The agent checks that the evidence describes the site it is meant to. An advert for a house on the same street is not the house. A floor area that disagrees between two certificates is a conflict to resolve, not a number to average.
- Build costs are a schedule, not a rate. The agent starts from a scope for the tier of works, edits it to what the evidence shows the site needs, and records every change with its reason, so a reader can see exactly why the cost is what it is.
- Planning is read, not assumed. Designations, constraints and the history of applications on the site and its neighbours are checked before a scheme is priced.
- Sensitivity is the default. Downside, base and upside scenarios are produced every time, with the maximum price at which the deal still meets its target profit.
The result is not a different method. It is the traditional method, done with the evidence work that the traditional tools left to the user.
5. Spreadsheet, software, agent
It helps to put the three side by side:
- Spreadsheet: you supply every input and every formula. Fast to start, easy to break, and only as good as your evidence.
- Appraisal software: you supply the inputs; the formulas and the report layout are built in. Faster and more consistent, still only as good as your evidence.
- AI agent: the agent gathers and argues the inputs from the evidence, the calculators do the formulas, and the report shows its reasoning line by line. You review the judgement rather than build the model.
The spreadsheet and the software answer "what do these numbers come to?". The agent answers the question that comes first: "are these the right numbers?".
6. Presenting it to a lender
Every guide to development finance gives the same advice: include all the information, evidence your assumptions, lead with the key metrics and be ready for follow-up questions. Lenders see a lot of appraisals, and a muddled one slows everything down.
An agentic appraisal is built to that brief, because the evidence is not an appendix assembled afterwards. It is the appraisal:
- The headline metrics come first: GDV, total costs, profit on cost and on GDV, the maximum offer and the verdict.
- Each assumption cites its source: the comparable sales behind the GDV, the priced cost plan behind the build costs, the planning record behind the planning assumptions.
- The follow-up questions are already answered: what happens if values fall, if costs overrun, if the programme slips, and why each comparable was kept or excluded.
The most useful thing a lender can be given is not an optimistic appraisal. It is one whose weaknesses are visible, because the lender is going to stress-test the numbers anyway.
7. What does not change
It is worth being clear about the limits.
- It is a desktop appraisal. An agent reads photos, plans and records; it does not walk the site. A survey, a quantity surveyor's cost plan and a measured building survey still matter before money is committed.
- Planning outcomes are uncertain. An agent can read policy, precedent and history and say what is likely. It cannot promise a committee decision.
- Judgement still needs a person. The value of written reasoning is that it can be challenged. A developer, a lender or a surveyor should read it, disagree where they know better and override it.
What changes is where people spend their time. The hours that went into gathering comparables, looking up rates and rebuilding models move to the decisions those numbers inform.
8. Conclusion
The appraisal method is not broken. GDV minus costs minus profit is still the right way to price land, and the traditional guides teach it well.
What the pre-AI era improved was the calculator. Spreadsheets made appraisals possible for everyone; appraisal software made them faster and tidier. Neither could tell you whether the numbers you typed in were true, and in a residual appraisal, that is the only question that really matters.
The AI era moves the work up a level. The calculator stays, exact and deterministic. Above it sits an agent that finds the evidence, argues each input, refuses what it cannot support, and shows its reasoning so a person can check it.
A spreadsheet tells you what the numbers come to. The agent tells you whether to believe them.
Written by the team at MarketCode, the operating system for UK real estate. Appraisal, our agentic refurbishment and development appraisal, is coming soon. To see the same approach on a single home today, read why an AVM is a prior, not a valuation and try an AI Valuation.
Further reading
- LandTech, Development Appraisals: A Complete Guide: the traditional method, with a worked residual example and a comparison of spreadsheets and appraisal software.
- MarketCode, How to run a development appraisal, step by step.
- MarketCode, Gross development value: what it is and how to calculate it.
- MarketCode, Planning constraints to check before buying a development site.
- MarketCode, Comparable evidence in property valuation.
Keep reading: Gross development value: what it is and how to calculate it, How to run a development appraisal, step by step and How to find out who owns land or a building in England and Wales.
Related guides
- How to run a development appraisal, step by stepA residual development appraisal tells you what you can pay for a site. The ten steps in order, a worked example for eight houses, and the sensitivities.
- MarketCode AI Valuer: the agent does the valuer's work, and shows itWhy statistical valuation stalls at the average home, how an agent settles the home, values it three ways and checks itself, and what is live and next.
- Gross development value: what it is and how to calculate itGDV is the value of a finished scheme as if sold or let on completion. How to build it from new-build sales, what affordable units do to it, and where it fails.