Inside AI Document Extraction: How Software Reads an Indian Sale Deed

Think of what happens after you hand a title verification tool three papers: a sale deed, an encumbrance certificate and a Record of Rights. Something must read them. Skimming will not do. Each party name, survey number and boundary has to be read, then compared across documents that may cover thirty years. This post follows one made-up sale deed from 1998, from coastal Andhra Pradesh, starting with a blurry scan and ending with the finished legal opinion, and it admits where the machine goes wrong.

What AI Extraction of Property Documents Means in India

In AI property document extraction, scanned or photographed records are turned into structured fields a computer can use. Those fields are then tested against one another and against outside registers. Feed it the image of a sale deed and it gives back labelled values for the seller, the buyer, the survey number, the extent, the consideration, the stamp duty paid, the registration number and the date. Encumbrance certificates, Records of Rights, link documents and tax receipts get the same treatment.

"Extraction" is a modest word for the task. A scanner turns ink into pixels, but extraction turns pixels into meaning. The system must recognise "Sy. No. 142/3A" as a survey number and "Ac. 0-20 cents" as a measure of extent. It also has to notice that the Telugu lines over the English block say the same clause again. Each value then goes into its proper slot, ready for a later step that compares it with the matching value in another document.

LegiScore takes over 150 separate data points from a full document set. Some are plain, such as the sale consideration. Others look minor yet decide the outcome, like whether the executant's father is named the same way throughout the chain. A human examiner would normally copy each of these into a checklist by hand, going through the documents one at a time, eyes tired, at 9 p.m.

The Real Condition of Indian Property Papers

A model can only read a deed once it copes with what the paper looks like, and Indian property paperwork is rarely tidy.

A typical set for verification is a pile of photocopies of photocopies. Our 1998 sale deed started life typed, was registered at the sub-registrar office, and got a handwritten violet endorsement on the back of each page. During twenty-eight years someone copied it for a bank loan, then for a mutation application, then for the present sale. Every copy loses contrast, and the registration number on page 6 has become a grey blur when the set arrives.

Language is a second layer. Twenty-two languages appear in the Constitution's Eighth Schedule, and regional scripts are used for property records. The test deed has two languages: the operative clauses are in Telugu with an English translation next to them, while the endorsement written at the bottom of every page is Telugu only. Papers from Tamil Nadu, Karnataka or West Bengal would bring Tamil, Kannada or Bengali.

Then there are physical flaws. A stamp impression spreads over two pages, a revenue stamp covers part of the consideration, and a page photocopied at the wrong scale loses the boundary schedule. Such things happen all the time, and the extraction step must deal with exactly them.

Reading the 1998 Deed Step by Step

Now the fictional deed. Koteswara Rao, son of Venkata Subbaiah, sells to two brothers who buy together. The land is Sy. No. 142/3A with an extent of Ac. 0-20 cents, in a Krishna district village. The price was Rs. 2,40,000, and the sale was registered in 1998 with the local sub-registrar.

Pass one is about the image. No character is trusted until the system has measured resolution, skew and contrast for each page. A page below the legibility threshold, like page 6 here, is flagged immediately and not once a wrong value is already in the report. The pages are straightened and their contrast evened out so faint typewriter ink can be read.

Pass two covers language and text recognition. The model works out which script each block uses, applies recognition suited to it, and pairs each Telugu clause with its English twin. If the two versions state one fact, that match adds to confidence.

Pass three is field extraction, in which pixels become the 150 data points. The system lifts from the deed who the parties are and whose children they are, the survey number with any sub-division, extent, all four boundaries, price, stamp duty, document number and date of registration. A confidence score and a reference to the spot on the page travel with each value, which lets a reviewer inspect the source.

What Each Document Adds

Extraction does not mean one model on one file. Each kind of document fills in a different part of the picture.

The sale deed tells you about the deal itself: seller, buyer, subject, price and the boundaries marking the parcel. From the encumbrance certificate you get a transaction list, every registered entry against the property in the period searched, and with it you rebuild the chain of owners and spot any mortgage or attachment. In Andhra and Telangana the Record of Rights (RoR, or pahani) names the recorded holder and gives the land classification and the extent according to the revenue department.

Each answers its own question. A sale took place, says the deed. The seller was entitled to sell, says the EC, since the chain leads to him. Revenue records agree, says the RoR. Three matching documents mean a clean parcel; a disagreement means a finding. Extracting that transaction list from an EC by hand is slow and prone to mistakes, and a machine can do it without boredom.

Comparing Documents: Where Machines Beat Tired People

A document may be read perfectly and still conceal a flaw that shows only against other documents. Here, steady consistency counts for more than brilliance, which favours software.

Three checks carry most of the load. Name drift comes first. Across thirty years "Koteswara Rao" becomes "Kotaiah" in one deed, "K. Rao" in a bank record and a slightly altered Telugu spelling in the RoR, and a person reading late in a long day may miss that one individual stands behind all of them. The system lines up each spelling from every document and flags the drift so that a person can accept or reject it.

Second comes extent reconciliation. Ac. 0-20 cents is the figure in the 1998 deed but the RoR gives 0-19 cents. A one-cent difference might come from rounding, a survey correction or an encroachment. The system puts the mismatch in front of you and leaves the meaning open. Third is boundary conflict. Suppose the old deed names a neighbour on the northern side who, according to a later document, also claims part of that survey number; the system raises the conflict. In documents we have handled from AP sub-registrar offices, differences of a few cents between deed and pahani are common and nearly always deserve another look.

How Anomaly Detection Works and What It Flags

This layer asks "does this look wrong?" in place of "what does this say?" Once the fields are extracted and compared, the system looks for patterns linked to title risk.

A stamp duty that looks too low for the price and the year gets flagged, because under-stamping may make a document inadmissible. So does a document number that does not fit that office's format for the year. A chain with a gap is flagged too, as when the EC moves the property from A to B to D without any recorded step between. The same applies to a property with one survey number on page 1 and a different one in the schedule.

A flag is never a ruling. It points and says a person should look. Spotting that two numbers disagree is the system's strength; deciding which is right is not. The legal reviewer makes that call, using the context the extraction produced.

Why You See and Can Edit Every Field Before the Report Runs

A black box and a usable tool differ in one design choice. Nothing is concluded before the user has seen each extracted field, and each one can be changed.

There are three reasons. The first is accuracy. Suppose the model read "142/3A" as "142/3B" because a smudge joined the characters. The person who has the original document corrects it in a click, before the wrong survey number touches a legal conclusion. The second is trust. A reviewer who sees the extracted value beside its place on the page, with its confidence score, decides personally whether to rely on it. A number that comes from nowhere earns no trust, while one you can trace and correct does.

Audit is the last reason. Each edit is logged, which means the opinion depends on values a person confirmed and not on silent model guesses. A bank operations team weighing an AI tool will ask if it can be defended to a regulator, and this settles it. By default LegiScore includes the review step: extraction proposes, a human disposes, and the report is built from confirmed inputs.

Is AI Able to Read Handwritten Property Documents?

In part, and be wary of anyone who claims completely. Nothing is harder to read dependably than the handwritten registration endorsement, in the violet ink that the sub-registrar adds under the Registration Act, 1908. Recognition fails more with cursive Telugu or Tamil that has faded for decades and been photocopied repeatedly than with printed text.

Structured handwriting in a known place works well. An endorsement always has a registration number, a date and the sub-registrar's notation in a predictable position, so the system knows where to look and what form the data takes. Free-form notes written sideways in the margin work badly. Our firm answer is that AI reads printed bilingual deed text with high reliability, reads structured handwritten endorsements with moderate reliability, and sends the rest to a human reviewer with the unclear area highlighted. Pretending otherwise yields confident wrong answers, and in title work these are worse than an honest "please confirm this."

Handling Poor Input

Real document sets come with missing pages, blur and cropping. A system that works only on clean input is a demo and not a product. The right behaviour is to spot the gap and ask for better input instead of guessing across it.

Imagine page 6 holds the consideration and stamp duty but cannot be read. No figure is made up. The field is labelled unreadable, the user is told the page and field, and a better scan of that page is requested. Or imagine an EC for 2005 to 2024 against a deed from 1998: the system states that the search period falls short of confirming the chain and requests the earlier EC period. If a regional format breaks an assumption, such as a Telangana pahani laid out unlike an AP one, the field returns with low confidence and goes to the human review queue and is not filled with a wrong value. What you want is a report that rests on legible material and names its gaps, not a polished opinion resting on guesses.

Manual Review Compared With AI Extraction

• Time for a full set: manual review takes 5 to 15 days for a title search and 2 to 4 weeks for full due diligence; AI extraction gives a full opinion in under 15 minutes
• Cost per opinion: manual review runs Rs. 15,000 to 50,000 (industry range); LegiScore says AI extraction costs a fraction of that
• Consistency: manual results vary with the examiner and the time of day; AI applies the same logic every time
• Cross-document checks: done by hand and limited by attention in manual review; AI compares every field automatically
• Fatigue errors: these increase late in long documents for manual review and do not occur with AI
• Language coverage: manual review is bounded by the examiner's languages; AI handles several Indian scripts in one pass
• Volume: manual review means one examiner with one set at a time; AI takes 1000+ pages and 500MB uploads per set
• Auditability: manual review keeps notes in a file; AI traces every field and logs each edit

Manual time and cost come from the published range for title due diligence in India, while speed and volume are LegiScore platform figures. None of this says people are poor at the job. Humans and machines have different strengths, so let the machine copy and compare and leave the judging to the human.

The Arithmetic Behind Under 15 Minutes

Add up the deed, link documents, decades of EC entries, RoR extracts and tax receipts, and the document set for a disputed or long-held parcel can exceed 1000 pages. For this reason LegiScore takes uploads of up to 500MB. One person working through 1000 pages, pulling out each field and cross-checking by hand, explains why the schedule runs to days.

Parallel processing makes it fast. The pages are handled together, the fields come out in one structured pass, and comparison across documents runs over the assembled values. Copying and comparing, the slow manual steps, shrink the most. People keep the work that should stay with them: looking over flagged fields, confirming corrections and forming the opinion. A process that took days thus yields an opinion in under 15 minutes, and a human still reviews every value that counts.