Somewhere between the eleventh and twelfth browser tab, most Singaporean home seekers hit a wall. There are the portal listings, the WhatsApp screenshots from a friend, the HDB resale price query, the URA caveat search, the mortgage calculator, and the nagging feeling that the unit you like is either 5% overpriced or a bargain nobody has noticed yet. In the last two years, a new tab has quietly joined that stack: a chat window.
AI property search in Singapore rarely announces itself. Nobody calls it that. It looks like pasting a listing link into ChatGPT and asking "is this a good price?", or prompting a model to rewrite a listing description in plain English, or asking an assistant to summarise what a neighbourhood is like at 7am on a weekday. These are small, unremarkable behaviours — and collectively they are reshaping the first, most consequential stage of the buying journey: the shortlist.
The shortlist is where deals are won and lost. By the time you message an agent, you have already decided which twenty units out of two thousand are worth your Saturday. Change how that filter works, and you change outcomes: which projects get traffic, which asking prices get challenged, which districts young buyers even consider. That is the quiet part. The loud part — AI-written marketing copy, AI-generated listing photos, AI chat assistants on agent websites — is mostly noise.
This piece is about the useful middle. What large language models and automated valuation tools genuinely do well with Singapore property data, where they fail in ways that cost real money, and three concrete workflows for using them without trusting them.
Why AI Property Search Took Off So Quietly
Singapore is unusually well-suited to this shift, for one structural reason: the transaction data is public, granular, and free.
- HDB publishes resale transaction records with block, floor range, flat type, floor area, lease commencement date, and price.
- URA publishes private residential caveats — project, unit size band, floor range, price, and date — refreshed regularly.
- The Master Plan, plot ratios, zoning, and URA's real estate statistics sit alongside them.
That is a lot of ground truth for a model to be pointed at. Compare this with markets where comparable sales are locked behind a valuer's subscription, and it becomes clear why Singapore is a natural testbed.
The second reason is friction. A Singaporean shortlist typically requires stitching together four or five sources: the portal for listings, HDB's resale price query for the block, URA's caveat data for the project, a loan calculator for the monthly number, and a mental model of school and MRT catchments. Each step has its own interface and its own vocabulary. Language models are extremely good at one thing above all: collapsing many small, fiddly tasks into a single conversational one.
The third reason is generational. Buyers aged 25 to 40 are already using AI tools for work — drafting emails, summarising documents, debugging spreadsheets. Applying the same reflex to a $700,000 decision is not a leap; it is the default. The question is not whether the reflex is there. It is whether the reflex is calibrated.
The new search stack
Read that diagram again, because the arrow that matters is the one pointing back. The feedback loop between AI output and verified source data is the difference between a tool that sharpens your judgment and one that flatters your existing bias toward a nice-looking photo.
What AI Genuinely Does Well in a Singapore Property Search
Let us be specific and fair. There are five categories of task where current models perform well enough to change your shortlist for the better.
1. Comparable-unit analysis at scale
Asking a human to pull twelve comparable transactions across three projects is a twenty-minute job with several tabs. Asking a model that has been given access to the transaction data is a thirty-second job. When it works, it works well:
- Grouping by unit type and size band. A 990 sqft three-bedder in a 500-unit development has a small natural comp set. Models are good at finding it.
- Filtering by recency. Restricting to the last 9 to 12 months, which matters enormously in a market where sentiment can shift within two quarters.
- Spotting the range, not just the average. A good output tells you the transaction spread — for example, that a stack has traded at $1,750 to $1,880 psf over the last year — and identifies which end the current asking price sits at.
This single capability probably accounts for most of the genuine value AI adds today.
2. PSF normalisation and like-for-like adjustment
Raw PSF is a blunt instrument. A low-floor unit facing an expressway and a high-floor unit facing greenery will not trade at the same PSF, even in the same stack. Models can be prompted to reason about adjustments explicitly — floor level, facing, stack, renovation condition, time on market, and, with the right data, remaining lease.
Ask for the adjustment logic in writing. If a model says "adjusted to $1,820 psf" and cannot explain why, treat the number as decorative.
3. Neighbourhood and amenity summaries
This is where language models are strongest in raw capability and weakest in accountability. They can synthesise a description of what a district feels like: the walk to the MRT, the wet market, the coffee shop cluster, the school catchment, the noise profile of a particular road, the age and character of the surrounding blocks.
The catch: much of this is generated from training data that may be years old, or from the listing copy itself. A model summarising a neighbourhood from the listing description is summarising the seller's marketing. Nine times out of ten that produces a pleasant, generic paragraph that tells you nothing you could not have guessed.
4. Decoding floor plans and listing jargon
Floor plans are a genuine pain point. A plan with a 12-metre frontage, a long internal corridor, a "study" that is legally a balcony recess, a yard that eats your kitchen, or a service yard that doubles as the only ventilation path — these are things a buyer discovers on site, if at all.
Multimodal models are now reasonably capable of reading a floor plan image and commenting on efficiency: circulation space, wasted corridors, dead corners, room proportions, natural cross-ventilation, and how a given furniture layout will actually feel. It is not a substitute for standing in the unit, but it is a genuine screening tool that did not exist three years ago.
The same applies to jargon.
- "Squarish layout" — sometimes true, sometimes a euphemism for a boxy bedroom with a skewed wall.
- "Rarely available" — unverifiable, and often simply means the last transaction was a while ago.
- "High rental yield" — usually means a small unit, a short walk to a business hub, and a rent you should verify against actual lease evidence.
- "Well-maintained" — check the age of the development and the state of the sinking fund.
5. Drafting and screening communications
The unglamorous win. Models can turn your messy criteria into a structured list of questions for an agent, draft an enquiry that gets you a straight answer instead of a brochure, and summarise a long WhatsApp exchange into a three-line action list. For buyers who dislike confrontation — which is most of us — a model that drafts the awkward question ("what is the actual remaining lease and has the flat been through HIP?") is quietly valuable.
A quick verdict table
| Task | AI performance | What you must still do |
|---|---|---|
| Pulling comparable transactions | Strong | Confirm the comps exist in a primary source |
| PSF adjustment reasoning | Moderate | Demand the adjustment logic in writing |
| Neighbourhood summaries | Mixed | Cross-check against current, first-hand sources |
| Floor plan reading | Improving fast | Visit the unit or a showflat |
| Listing jargon translation | Strong | Verify every claim against a document |
| Instant valuation estimates | Mixed | Stress-test with your own comp set |
| En bloc speculation | Weak | Speak to someone who reads the Master Plan for a living |
| Lease decay analysis | Weak | Manually check remaining lease and financing rules |
Where AI Still Fails — and Fails Confidently
This is the section that will save you the most money. Every failure below has a common root: the model does not know what it does not know, and it will still answer.
Failure 1: Lease decay
Singapore's residential stock is dominated by 99-year leasehold. Two units in the same development, identical in every visible respect, can have 50 years of remaining lease between them if one block was completed in 1974 and another in 2024. Their prices should reflect that. Models frequently do not.
The consequences compound:
- CPF usage collapses. For HDB flats with 20 years or fewer of remaining lease, CPF savings generally cannot be used at all. Between 20 years and the threshold where the lease covers the youngest buyer to age 95, CPF usage may be pro-rated rather than full. This is a hard rule, not a soft guideline, and it changes your cash requirements overnight.
- Bank financing gets harder. Lenders become progressively more cautious as remaining lease shortens. Reportedly, many grow uncomfortable below roughly 30 years, and the loan tenure offered can be capped by the lease rather than by your age.
- Exit risk rises. Your future buyer faces the same constraints, plus a shorter remaining lease. The pool of buyers who can pay cash for the CPF shortfall shrinks over time.
Remaining Lease Needed for Full CPF Usage, by Youngest Buyer's Age
The chart is derived from the CPF rule that the flat's remaining lease must cover the youngest buyer to age 95 for full CPF usage. It is a rule, not a market forecast — and it is exactly the kind of mechanical arithmetic that models tend to gloss over in favour of a confident-sounding paragraph about "value".
An AI that tells you a 1978-built flat is "priced attractively at $520 psf below the block average" without flagging a 45-year remaining lease is not wrong about the PSF. It is wrong about everything that matters.
Failure 2: En bloc potential
Ask a language model whether a development has en bloc potential and you will usually get one of two answers: a vague "it could be redeveloped in the future" or a confident restatement of news from several years ago. Both are close to useless.
Real collective sale feasibility depends on factors that live outside the listing and often outside the model's training data:
A site with an unexhausted plot ratio, a long remaining lease, good geometry, and concentrated ownership is a genuine candidate. A site with a high plot ratio but only 45 years left on the lease may need a top-up so expensive that the economics never clear. A development that has already used its entire plot ratio has nothing to sell.
Consent thresholds also matter and are commonly misremembered: the LTSA requires 80% by strata area and share value for developments under 10 years old, and 90% once the development is 10 years or older. Models routinely mix these up or omit them.
The honest position: most en bloc attempts fail. Treat any AI-generated en bloc commentary as a prompt to go and read the Master Plan and the site's development baseline yourself — or to ask someone who does that for a living.
Failure 3: Transaction nuance
Transaction data is not as clean as it looks. These are the distinctions a model will flatten into a single PSF number:
- Subsales versus resales. A subsale is the resale of an uncompleted unit by the original buyer. It can price above or below the developer's recent sales for reasons that have nothing to do with market direction.
- Non-arm's-length transfers. Transactions between related parties, or in the context of a divorce or estate settlement, can appear in the data at prices that distort the average.
- Units sold with tenancy. A tenanted unit with a below-market rent rolls into the price. The PSF looks fine; the yield does not.
- Mortgagee sales. Usually priced to move, and usually not a true comparable.
- Caveat versus actual price. Caveats can be lodged late, amended, or cancelled. A "current" dataset may be missing the last six weeks of activity.
- Lodgement lag. HDB resale data and URA caveat data are published on different cadences and with different lag profiles. Mixing them without accounting for that creates phantom trends.
- Option date versus completion date. HDB resale transactions typically complete several weeks after the option is exercised. Private resales take longer still. A "recent" transaction may reflect a price agreed two or three months before its publication date.
- Progress payments. New launch pricing works on a different payment schedule. A $2,000 psf new launch and a $1,800 psf resale in the adjacent project are not directly comparable.
- Floor area definitions. URA's harmonisation of floor area definitions, phasing in for new development applications, means that strata area measured in older projects may include components that newer projects exclude. A square metre is not always a square metre.
None of this is exotic. All of it is routinely missed by an AI answering the question "what's the market rate for this project?"
Failure 4: Hallucinated comparables
The single most dangerous failure mode. Ask for comparable transactions without giving the model a verified dataset and it may produce plausible-looking entries: a right project, a slightly wrong unit size, a price that is close but invented, a date that feels right. It reads exactly like a real comp. If you act on it, you have made a six-figure decision on a sentence a model made up.
The rule is simple and non-negotiable: a comparable that cannot be traced to a primary source is not a comparable.
Failure 5: Financing blindness
AI can do mortgage arithmetic. It is much less reliable at the rules that determine whether you can borrow at all.
| Buyer profile | ABSD on first property | ABSD on second | ABSD on third and beyond |
|---|---|---|---|
| Singapore Citizen | 0% | 20% | 30% |
| Singapore PR | 5% | 30% | 35% |
| Foreigner | 60% | 60% | 60% |
Additional Buyer's Stamp Duty Rates by Profile (%)
These are the rates set under the April 2023 cooling measures; trust funds and entities sit higher still. They change. Verify before you commit.
Then layer on the loan-to-value limits, which depend on how many housing loans you already have outstanding:
Maximum Loan-to-Value by Outstanding Housing Loans (%)
On top of that sit the Total Debt Servicing Ratio at 55% of gross monthly income across all debt, and — for HDB flats and new Executive Condominiums — the Mortgage Servicing Ratio, which caps your property loan instalment at 30% of gross monthly income. Add in the income ceilings for HDB flats and CPF housing grants, and you have a rulebook that decides your budget before any PSF comparison is relevant.
A model that helps you fall in love with a $1.6 million condo when your MSR-governed budget is $950,000 has done you no favours at all.
Failure 6: Echoing the listing
If a model reads the listing to describe the property, it inherits the listing's framing. "Bright and breezy" becomes fact. "Mins to MRT" becomes three minutes when it is twelve. The fix is boring and effective: never let the same document be both the subject and the source.
Three Verified Workflows for Combining AI With Hiva's Transaction Data
Here is where this becomes practical. Each workflow takes under an hour, uses AI for what it is good at, and anchors every number to verified transaction data.
Workflow 1: The Twenty-Minute Shortlist
Goal: turn an overwhelming portal scroll into a defensible list of 10 to 15 units.
Step 1 — Define your constraints in writing. Before opening any AI tool, write down: budget range, monthly instalment ceiling, minimum remaining lease, flat type or unit size band, maximum walk to MRT, and districts you will and will not consider. This document is your filter; everything else is execution.
Step 2 — Let AI structure the search. Feed your constraints to a model and ask it to produce a search specification: which districts, which size bands, which project age ranges, what to exclude. You are using the model as a research assistant, not an oracle.
Step 3 — Verify against transaction data. For every unit that survives, check three things in verified data before you message anyone:
- Block-level or project-level recent transactions — at least five, in the last twelve months, in a comparable size band.
- Remaining lease — from the lease commencement date, not from the listing.
- MOP status — for HDB units, whether the minimum occupation period has been served, and whether it is a Prime or Plus flat carrying a 10-year MOP with subsidy clawback and rental restrictions.
Step 4 — Score and cut. Rank what is left and drop everything that fails a hard constraint. Hard constraints beat soft preferences every time.
The twenty minutes is not the AI step. It is the verification step, and it is the only one that changes your outcome.
Workflow 2: The Valuation Stress Test
Goal: answer the only question that matters at offer stage — is this price defensible?
Step 1 — Get the AI estimate. Ask for a valuation range, not a point estimate. A range of $1,780 to $1,860 psf is more honest than "$1,820 psf".
Step 2 — Build your own comp set. Pull eight to twelve transactions from the same project or block, same size band, last nine to twelve months. Record price, PSF, floor range, date, and any flags: subsale, tenanted, mortgagee sale, related party.
Step 3 — Adjust, and write down why. For each comp, note the adjustment you are making. Higher floor: up. Facing an expressway: down. Older completion date: adjust for market movement. Longer remaining lease: up. Larger unit: usually a lower PSF.
Step 4 — Check the project's own trend. Is this project's PSF rising, flat, or falling over the last four quarters? A single high transaction in a softening project is not a valuation; it is an outlier.
Step 5 — Sanity check with the lender. A bank's indicative valuation is a second opinion that costs nothing. If your adjusted range and the bank's valuation are far apart, dig into why before you sign anything.
Step 6 — Decode the price with the data, not the narrative. If the asking price is 8% above your adjusted range, ask the agent a specific question: which transaction supports this price, and what adjustment gets you there? Vague answers are informative.
Workflow 3: The One-Page Decision Memo
Goal: stop yourself from buying on vibes, and give yourself a document to revisit in three years.
Ask a model to draft a structured memo with these headings, then fill each one with verified figures:
- The unit — project, block, size, remaining lease, completion year, tenure.
- The price — asking, your adjusted range, the gap, and the justification for the gap.
- The comps — a table of five to ten verified transactions with dates and flags.
- The financing — loan amount, LTV applied, monthly instalment, TDSR and MSR position, ABSD payable, CPF usage permitted given the remaining lease, cash required at completion.
- The bear case — three specific reasons this could be a mistake. New supply in the area, an upcoming MRT line that shifts demand, a lease that shortens your exit pool, a development with an unresolved maintenance issue.
- The exit — who is the likely buyer in five to ten years, and what will they care about?
- What I still do not know — the honest section, and the one that prevents most bad purchases.
The bear case and the "what I still do not know" sections are where AI earns its keep. Models are reasonably good at generating counterarguments when asked directly, and reasonably bad at volunteering them.
The Verification Cheat Sheet
| If AI tells you… | Verify it by… |
|---|---|
| "Comparable units sold at $X psf" | Tracing each transaction to HDB resale or URA caveat data |
| "This is a good price for the area" | Building your own adjusted comp set |
| "The lease is not an issue" | Calculating remaining lease from the commencement date, then checking CPF and financing rules |
| "This project has en bloc potential" | Reading the Master Plan, plot ratio, development baseline, and remaining lease |
| "You can afford $X" | Running TDSR and MSR against your actual income and debts |
| "The yield is around 3.5%" | Checking actual lease evidence, not the listing |
| "It is a five-minute walk to the MRT" | Walking it, at 8am, on a weekday |
| "This neighbourhood is up and coming" | Checking the Master Plan and actual transaction trends, not adjectives |
Food for Thought
1. If everyone shortlists with the same model, do prices get more efficient — or more correlated? AI tools compress the research process. That should make markets more efficient in theory. But if thousands of buyers are nudged toward the same "undervalued" projects by the same underlying data, the effect may simply be faster convergence on a narrower set of units. The question worth asking is whether your AI-assisted shortlist looks like everyone else's.
2. What happens to the value of an agent when the buyer arrives with a better comp set? The agent's traditional edge was access to information. That edge has largely evaporated. What remains — negotiation, handling the option and completion process, knowing which developer will actually move on price, spotting an issue at the viewing — is real, but it is a different job from the one most agents trained for.
3. Is remaining lease now the single most under-priced variable in Singapore residential? It is mechanical, it is knowable, and it materially affects CPF usage, financing, and exit liquidity. Yet it is routinely absent from the first conversation. If a market systematically under-adjusts for a variable this consequential, that is where the mispricing lives.
4. Does AI make you more likely to buy a home you will regret? Language models are optimised to produce confident, fluent answers. Fluency feels like expertise. If a model describes a 45-year-remaining flat as "well-priced for the district", the sentence is grammatically perfect, tonally reassuring, and financially dangerous. Which of your own decisions in the last year were made on fluency rather than evidence?
5. When the model is wrong and you cannot tell, whose responsibility is it? No regulator certifies an AI valuation. The agent's duty runs to the seller. The bank's valuation exists to protect the bank. Structurally, the only party with an incentive to get this right is you — which is an argument for building your own verified reference point, not for abandoning the technology.
The Shortlist Is Still Yours to Make
The technology is not the story. The story is what it changes about the moment you decide which ten units to visit. AI has genuinely improved that moment: comps come faster, floor plans are easier to read, jargon is easier to see through, and the boring administrative work of a property search takes minutes instead of weekends.
What it has not changed is the burden of proof. Lease decay is arithmetic, not opinion. En bloc potential is a development feasibility question, not a vibes question. Transaction nuance lives in the footnotes of the data, not in the summary. And financing rules are decided by policy, not by a well-written paragraph.
The right posture is the one you would take with a brilliant, fast, occasionally careless research assistant: let it do the legwork, then check the numbers yourself before anything is signed. AI is very good at producing a shortlist. It is not good at producing accountability.