AI Property Development Feasibility: A First-Pass Appraisal

Build a residual appraisal in a spreadsheet: estimate the gross development value from comparable sales, subtract construction, fees, finance, selling costs and your target profit, and read what is left as the most you can pay for the land. Use AI to structure the inputs and challenge them, not to supply or calculate the numbers.

This is a screening tool, not a valuation, and it does not replace a quantity surveyor, a planning consultant or a lender's own appraisal. Its job is to tell you in an afternoon whether a site deserves their fees. The worked example is six houses on a cleared site. Every figure is illustrative, so rebuild it from your own sources before relying on any of it.

Follow me on Instagram@sagnikteaches

What a first-pass appraisal answers, and what it cannot

A first pass produces three things. The first is the maximum land price at your target profit. The second is the profit you would make at the price the seller is asking. The third, and the most useful, is which assumption the answer depends on most. It cannot tell you what the site is worth, what a lender will lend or whether planning consent will be granted. It also leaves out what a funder will model properly: a monthly cash flow, a drawdown schedule and tax. Treat anything it cannot answer as a question for the consultants, and write the questions down as you go.

Connect on LinkedInSagnik Bhattacharya

Dedicated feasibility software exists. Civil Intelligence, for example, describes itself as an AI property feasibility and zoning tool that returns zoning details, environmental screening and a feasibility score for a property, with plans listed from $149 a month for 10 searches (as of November 2026). That kind of tool speeds up the site-constraint side. The model built here handles the money side, which is where your own sales evidence and cost sources have to go, and it works with a general assistant and a spreadsheet you already have. If you try a dedicated tool, check whether its cost and sales assumptions are yours to edit.

Subscribe on YouTube@sagnikteaches

The inputs table: one source and one owner per line

Start with a table in which every input has a value, an evidence grade and a person who verifies it. The grades are simple. A means a quote or a signed document. B means published data or comparables you have read yourself. C means your own estimate. D means a number an AI tool suggested, and D is not allowed into the final model.

The assistant earns its place here, at the start. Give it the scheme and ask for the inputs table back with the values blank, plus the lines that first-pass models usually forget. It must not suggest values.

I am a small developer screening a site before paying consultants.
Scheme: six three-bedroom houses of about 1,100 sq ft each on a cleared
site, planning status [STATUS].

1. Build an inputs table for a residual appraisal: Input | Unit |
   Value (leave blank unless I gave one) | Evidence I need.
2. List the cost and risk lines a first-pass appraisal for a scheme like
   this usually misses.
Do not suggest any values.

The "missed lines" it returns are worth a minute of your attention. For this scheme a typical list would include ground investigation and any remediation, utility connection capacity and upgrades, access works and adoption of new roads, agreements with neighbours on boundaries, structural warranty and site insurance, sale incentives, and the fees to discharge planning conditions. Each one either becomes a line in the inputs table with a source, or goes on the consultants' question list. None of them gets an invented number.

Gross development value: comparables, not the chatbot's guess

Gross development value (GDV) is the total you expect to receive from selling the finished scheme, and it is the biggest number in the model, so it deserves the most scrutiny. An assistant has no live sold-price records and will happily produce comparables that sound right. Collect the sales yourself, from sold-price records and agents, and use the assistant only to tabulate them and to challenge them.

I will paste five sold comparables that I collected myself. Do not add
any others and do not use sales data from memory.

COMPARABLES (type | area in sq ft | sold price | months ago | notes):
[paste]

1. Show price per sq ft for each one, with the working.
2. Flag any comparable that differs from my scheme in type, condition,
   age of sale or size by more than 15%, and say why it matters.
3. Give the range and the average of the comparables you consider closest.
4. List questions I should ask a selling agent. Do not propose a price.

Illustrative output for five comparables the developer collected:

The assistant's summary says the closest comparables are 1 and 2, with a range of $403.57 to $409.52 and an average of $406.55. Its questions for an agent include whether buyers pay a premium for new-build houses of this size, what incentives they now expect, and how long comparable new homes are taking to sell. The assistant has proposed no price, as instructed. Choosing one is the developer's job, and here is the reasoning in the example: the new-build premium in comparable 5 is unproven, because incentives may be hidden inside that sale price, so the developer takes no premium and shades the closest average down by about 1.6%, to $400 per sq ft. That gives a GDV of 6,600 sq ft at $400, or $2,640,000, which is $440,000 a house.

Two checks stay with you. Recompute every per-sq-ft figure and the average on a calculator, which takes two minutes. And phone two selling agents who work the area, giving them the scheme and asking what they would expect each house to sell for. Their opinions are free, they are grade B evidence when written down, and they are the first thing a lender's valuer will compare with your number. A selling agent can also tell you whether the new-build premium in your market is real, which no comparable table can.

Costs: where the assistant helps and where it only guesses

Ask a chat assistant for the build cost per square foot of a house and it will answer in a confident range. It cannot know your specification, your site, current labour and material prices, or what builders are quoting this quarter, so its figure is grade D however plausible it sounds. In the example the assistant offered $210 a square foot. Without a source it could not go into the model, and the verification log below shows it being rejected. The figure that goes in comes from a quantity surveyor's indicative budget or two builders' rough prices from the same drawings, here $160.

The assistant is useful around the cost lines instead of inside them. It can draft the request to a quantity surveyor for an indicative budget, with the drawings, the specification level and the questions you want answered. It can read the budget that comes back and list which lines are the largest and least supported. And it can explain, in plain words, what a contingency covers (unknowns in construction, not design changes) so that you keep the two apart. For a broader view of where AI estimating helps and where it fails, see whether AI can produce accurate estimates for a small building firm.

Finance deserves a sentence of its own, because the first-pass method simplifies it. Here, build costs are assumed to be drawn evenly over a 12-month build and then carried for 6 months of sales. That averages out to about one full year of interest on the build total, so finance on the build costs is the build total times 8.5%. The land is carried for the full 18 months. Real lenders draw down differently and charge fees, so the 8.5% in the example is meant to include them, and the real number comes from the lender's indicative terms. The simplification is deliberately crude and quick to explain to a consultant.

The residual calculation, line by line

The residual method works backwards. Start with what the finished scheme will sell for, deduct everything it costs to deliver, deduct the profit you need, and what is left is the most you can pay for the site. Land is the balancing figure, and that is why it swings so hard when anything else moves. Here is the base case, using the inputs above.

The last line needs a word of explanation, because the land price sits on both sides of the sum. Acquisition costs are 5% of the price, and the land, with those costs, is financed at 8.5% for 18 months, which is 12.75%. So every dollar of land price costs $1.05 once acquisition costs are added, and then 12.75% more in interest. Dividing the residual by 1.05 and then by 1.1275 gives the highest price at which the whole scheme still delivers 20% of GDV. At $371,659 the numbers close exactly.

Now the AI's part. Ask the assistant to build this in a spreadsheet with every result as a formula that refers to the inputs sheet, and nothing typed in. Then check three cells by hand: GDV, the build-related total and the residual. If those three match your calculator, the structure is probably right. For help with the formula side, the ChatGPT and Excel guide covers asking an assistant for formulas and checking them.

Build a residual appraisal in a spreadsheet. Put every input on one
sheet with a label, a value and a note on its source. On a second sheet,
write every calculation as a formula that refers to the inputs. Do not
type any result.

Lines: GDV; construction; external works; contingency; professional
fees; planning and statutory; build-related total; finance on build
costs; sales costs; target profit; residual before land; maximum land
price = residual / ((1 + acquisition %) * (1 + finance rate * land
years)); profit at an asking price.
Write each formula in words in the column beside it. Do not round
inside formulas.

Models can slip on arithmetic, and a residual model punishes it: one wrong line moves the land price by thousands without looking odd. That is the argument for formulas over chat calculation, and for a hand check on the base case before anything else is built on it. Whether ChatGPT can calculate property yield reliably makes the same case for a related calculation.

At the asking price: the first go or no-go

The seller is asking $420,000. The model's maximum at a 20% profit is $371,659, so the asking price is about $48,000 above what the target allows. Put the asking price into the model and see what profit is left.

At the asking price the scheme earns 17.8% of GDV: below the 20% target, though inside the 15% to 20% of GDV that published planning-viability guidance treats as a suitable developer return. Whether it is enough is your funder's call, not the guidance's. That is a negotiating position, not a verdict. The options are to offer around $372,000, to find savings in the build, to accept a thinner margin, or to walk away. Which one is right depends on your funder, your appetite and your alternatives, none of which the model knows. It also depends on the next section, because a margin only means something once you know how easily it disappears.

Sensitivity: the two numbers that can sink the scheme

A single answer is fragile, so rerun the model with the two biggest assumptions moved. Sale prices and construction costs account for most of the movement in any residual appraisal. This grid shows profit as a share of GDV at the seller's asking price, with sale prices down by up to 10% or up by 5% and construction and external works costs up by 0% to 15%. Contingency and fees follow the cost change, and the planning costs stay fixed. The figures come from the spreadsheet, not from the assistant's head.

Read it from the middle row. A 5% fall in sale prices alone takes the margin from 17.8% to 13.7%. A 10% rise in costs alone takes it to 12.1%. Both together leave 7.7%, which is about $192,600 of profit on a scheme costing more than $2 million. A fall of 10% in prices with a 15% rise in costs produces a small loss. None of these moves is extreme for a build lasting a year or more, and that is the finding: at the asking price the scheme has little room for error.

The same exercise on the maximum land price is even starker. It falls from $371,659 to $285,806 if prices are 5% lower, to $244,816 if costs are 10% higher, and to $158,963 if both happen. A 5% change in the sales line removes almost a quarter of the land budget, because land is the balancing figure and the profit and selling costs shrink with GDV. When you speak to a seller, that is the argument for a lower offer, or for conditions that protect you if the sums move.

The assistant can help set the grid up, by writing the spreadsheet's data table or telling you how to build one, and it can name the assumptions it considers most fragile when you show it the inputs table. It should not produce the grid's numbers itself. If you want to know how far prices can fall before profit reaches a floor, use the spreadsheet's goal-seek tool. For a general method of thinking in best, likely and worst cases, scenario planning with AI is a useful companion.

The planning-risk checklist

A model that ignores planning risk assumes the scheme can be built, so planning belongs next to the numbers. Collect the documents that bear on the site: the policy sections for this type of site, any pre-application notes, and the decision notices for two or three nearby schemes. Then ask the assistant to build a checklist from those documents alone. It must not predict whether consent will be granted, because it cannot know.

Below are extracts I copied: (1) planning policy sections for this type
of site, (2) notes from a pre-application meeting, (3) two decision
notices for nearby schemes.

Build a risk checklist for my six-house scheme with these columns:
Risk | What the documents say (quote the line) | Rating: Low, Medium,
High, or NOT IN DOCUMENTS | Who must confirm it.
Use only the documents. Do not predict whether consent will be granted.

Illustrative output, five of the rows:

The rows marked NOT IN DOCUMENTS matter as much as the rated ones, because they are the questions your documents did not answer. Add the rated items that move money to the model: the tree risk may change the layout and the number of houses, and the contributions line is the $14,000 per house in the inputs table, which needs a written source. Time is a cost too. At the asking price, land finance runs at about $3,124 a month, so a six-month planning delay adds roughly $18,700 of interest before any overheads. Put the delay in the sensitivity thinking as well.

What the assistant cannot know is how the planning body will weigh the scheme, whether policy has changed since your extracts were written, or what pre-application officers tell you off the record. Read the decision notices yourself, in full, rather than relying on the assistant's quotations, and check each quoted line against the source. For the wider checks on a purchase, AI due diligence on real estate covers leases, surveys and searches.

A verification log: every number gets a source, a grade and a name

Before the model is shown to anyone, keep a log that records where each input came from. It is a table you complete as you go, and it turns "I think it was about that" into a trail a consultant or funder can follow. Here is the log for the example, including one number that did not survive.

Then use the assistant as a sceptic. Give it the summary and ask it to name the weakest assumptions. It should not recalculate anything. Its list becomes your consultants' question list.

Below is my appraisal summary: inputs with evidence grades, the
residual calculation and the sensitivity grid. Do not recalculate
anything.
1. List the five assumptions you consider least supported, worst first,
   and say why.
2. For each, say which consultant or document could firm it up.
3. List anything the summary treats as certain that depends on a
   decision by someone else (a planning body, a lender, a seller).

For this example a sensible answer would put construction cost first, since it rests on one indicative budget and a second price would firm it up. Next comes the sale price, which has no new-build evidence beyond comparable 5. Then the planning contributions, which rest on notes and need a written estimate from the planning body. Then the finance rate, which is indicative rather than offered. And last the 12-month build programme, which no contractor has confirmed. None of those is a surprise to an experienced developer, but written down in order they turn a vague unease into an agenda.

A profit-basis mix-up that moved the land price by $74,000

Here is the error to guard against, described as an illustration. The inputs sheet says "target profit 20%". The formula the assistant wrote treats it as profit on cost, while the funder meant profit on GDV. A 20% margin on cost is the same as 16.67% of GDV, so the model deducts $440,000 instead of $528,000 for profit. The residual before land rises by $88,000 to $527,998, and the maximum land price jumps from $371,659 to $445,992, a difference of about $74,300. Against the $420,000 ask, the model would say you have $26,000 of room when in truth you are $48,000 over.

It shows up when a second person reads the formula-in-words column and sees "profit = 20% of total cost" beside a label that says "on GDV". The fix is cheap: state the basis in the input's label, require the formula to be written out in words, and recompute the profit line by hand. The same trap lies in every percentage in the model, which is why every percentage in the inputs table should say what it is a percentage of. It is also the kind of error that sounds too technical to matter until it moves a purchase price.

What to hand the consultants once the site passes

A site that clears the first pass earns a fee for the people who can firm up the numbers. Send each of them a one-page pack: the inputs table with grades, the base case, the asking-price case, the sensitivity grid, the planning checklist with its NOT IN DOCUMENTS rows, the verification log and the five fragile assumptions. Ask the quantity surveyor for a second look at the construction budget and any missed lines, the planning consultant for their view of the high-rated risks and the likely timetable, and the agents for a written opinion on the sale prices. Ask your lender for terms and for their own margin requirement. The assistant can draft each cover note from the summary, using only the summary.

A first pass like this takes an afternoon. Inputs and the missed-lines list take about 30 minutes, the comparables and the agents' calls 45, the cost sources 45, the spreadsheet 45, the sensitivity 15, the planning documents 45 and the log 15, which is roughly four hours (illustrative). The aim is not a precise answer but an early, cheap and honest one, with every number sourced, so that the money you spend on consultants goes to sites that can survive their scrutiny. For the commercial version of the same idea, an AI first-pass underwriting model for commercial property follows a similar pattern.

Developer questions about AI and first-pass appraisals

Can ChatGPT tell me the gross development value?

Not reliably. A chat assistant has no live sold-price database and can state comparables that do not exist, with confidence. Use it to organise sales you have collected yourself and to flag outliers. Take the value itself from sold records and the opinions of selling agents who work the area.

What profit margin should I use?

Published planning-viability guidance treats 15% to 20% of gross development value as a suitable developer return, other guides quote different figures, and some measure profit on cost, which gives a higher percentage for the same profit (20% of GDV is 25% on cost). Use the figure your funder requires, write down its basis, and test lower margins too.

Is this a valuation I can show a lender?

No. It is a screening model built on your own assumptions. Lenders commission or require their own valuation and cost review, and a model whose inputs have no sources will not satisfy them. Its value is deciding which sites deserve consultants' fees.

Should I let the AI do the calculations in chat?

Have it write spreadsheet formulas, or use a tool that runs code, and check the base case by hand. Chat models can slip on arithmetic, and one wrong line in a residual appraisal can move the land price by tens of thousands of dollars without looking odd.

Further reads

• AI Real Estate Investment Analysis: Vet a Rental Deal Step by Step — The same discipline applied to a rental purchase.
• Using AI for Property Investment: Screening a Deal Step by Step — A screening sequence for deciding which deals to look at.
• How Can AI Help Quantity Surveyors? BoQ Checks, Takeoffs and Cost Reports — What a quantity surveyor will do with your first-pass costs.
• Best AI for Real Estate Wholesalers: Underwrite a Deal Before You Call — Quick underwriting before you call a seller.
• How to Catch Made-Up Figures in AI-Drafted Proposals — A method for catching figures a model invented.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Civil Intelligence's product and pricing page; government planning-viability guidance on developer return as a share of GDV; published guides to residual land value; the worked example is illustrative.