Inside a 30-Minute Title Search: How an AI Engine Builds the TSR

In Andhra Pradesh or Telangana, a manual title search usually needs three to seven working days. The advocate goes to the sub-registrar office and takes out index II, then walks to the tehsildar for mutation records. A clerk is told to fetch encumbrance certificates, and an RTI for revenue records may be filed when needed. After that come case searches on eCourts, a single court per search. The same workflow, handed to an AI title search engine, takes about 30 minutes and sometimes less.

There is no magic in it. The speed comes from parallelism and structured automation, and a lawyer signs off at the end. Here we open the bonnet and follow what happens from the moment you upload a sale deed to the moment a Title Search Report (TSR) reaches your inbox.

Defining the AI Title Search Engine

A title investigation has five traditional tasks: fetch documents, parse them, trace encumbrances, check litigation and flag risk. Software that automates all five is an AI title search engine. Reading is done with OCR (Optical Character Recognition), NER (Named Entity Recognition) and clause-extracting large language models. Facts come from simultaneous API calls to revenue, registration and court portals of the states. The Title Search Report is ready in minutes, not days, with the lawyer retained for the closing legal opinion and for judgement calls.

The word "engine" is deliberate. Nothing here is a chatbot that guesses at risk. The pipeline is deterministic and is made of parsers for documents, scrapers for portals, reconstructors of chains and detectors of red flags that follow rules, and AI enters only at the point where free-form language meets structured land records.

The Pipeline in Seven Steps Over 30 Minutes

What follows is the actual order of operations in a modern AI title search engine such as LegiScore. Stages run side by side wherever they can, which is why the total clock time stays short even with 20-plus portals involved.

Step 1: Ingesting Documents and Running OCR (0 to 90 seconds)

The user, a buyer or a banker, supplies the sale deed and its parent documents, the newest encumbrance certificate, the patta/khata, the survey settlement record, and link documents if any. PDFs, scans and JPEGs are all accepted, even the smudged photocopies that old Telugu-script deeds usually come as.

The OCR knows each language. A separate model handles each script, whether Telugu, Tamil, Kannada, Marathi, Hindi or English, so descriptions of the property, survey numbers and names of parties are extracted faithfully. The raw text then moves to a Named Entity Recognition layer, which pulls out the following:

• Names of vendor and vendee, with aliases, because "K. Rao" and "Kondaiah Rao" often turn out to be one person across documents
• Extent with survey numbers and sub-divisions, in acres, guntas, sq yards or sq metres
• Boundaries on the north, south, east and west
• Document numbers, the year, and the sub-registrar office of registration
• Entries of encumbrance, the holders of charges and the entries of discharge
• Schedules of the property: first, second and third

Step 2: Reconciling the Identity of the Property (90 seconds to 3 minutes)

This step is where most manual title searches quietly fail. A single plot goes by different names across the sale deed, EC, patta and local-body tax record. Say a sale deed of 1992 gives survey number 142/A. After a sub-division mutation in 2003 the same plot appears as 142/A1. The corporation's property tax record, meanwhile, identifies it by a door number issued in 2011.

A property identity graph is the engine's answer. Survey number, sub-division history, plot number, door number, patta number and mutation references are linked in it, which means later portal queries go out under every alias of the parcel, not only the one on the newest deed. Nothing produces more false-negative title searches in manual work than omitting this step.

Step 3: Querying Portals in Parallel (3 to 10 minutes)

After the property identity is fixed, the engine launches 20-plus queries at once against state and central portals. For a typical AP or Telangana search, the list of calls is as follows:

• IGRS (registration department): confirms that the deed was registered and pulls index II
• MeeBhoomi (AP) or Dharani (TS): revenue records, pattadar status, ROR-1B and adangal
• Sub-registrar EC search: encumbrances over 13 and 30 years
• Municipal property tax portal: present tax status and the trail of door numbers
• Corporation building plan approvals: sanctioned plans and any deviations
• eCourts services: case search in district courts, high courts and consumer fora
• DRT and DRAT: SARFAESI proceedings and recoveries that are pending
• IBBI database: insolvency cases against the seller
• MCA21: for a company seller, charges, CIRP status and director details
• GST portal: whether the seller's GSTIN is active, if a company
• ED, CBI and PMLA databases: attachment orders
• Lok Adalat and revenue court records: land disputes still pending
• RERA portal: project registration, when the property is in a notified project
• BIFR / NCLT: older proceedings against corporate sellers
• Income tax attachment lists, where they are published
• Local body records of leases and grants
• Cantonment board records, where applicable
• TSPCB / APPCB notices for industrial parcels
• Forest department records for parcels close to reserve boundaries
• Survey of India or the state survey department: to confirm the survey number is genuine
• Records of stamp duty refunds and adjudication

This stage turns days into minutes purely because of parallel execution. A clerk has to take the checks in sequence, as one person cannot be at 20 counters together. Software opens 20 sessions, beats captchas at machine speed and collects the replies concurrently, so the slowest portal alone sets the total time, not the sum.

Step 4: Searching eCourts for Litigation (parallel with Step 3)

The hidden defect with the greatest danger is a court case in progress, since registration records never show it. eCourts is queried by the engine for every bench in which the property, the seller, or the seller's spouse or company could reasonably appear as a party. A search of this kind spans courts of 28 states and over 600 districts, more than any manual advocate could cover in a day.

Matching is done on these items:

• The seller's name and known aliases
• The seller's spouse, and major family members who are named in any document
• Where relevant, the seller's company name with its CIN
• Survey and door numbers used as search strings against cause titles
• Holders of a power of attorney named in any document in the chain

Each case found is then classified, whether it is a suit for specific performance, a partition suit, an injunction, a writ petition, a criminal complaint, lis pendens or a SARFAESI proceeding, and marked with what it does to title.

Step 5: Rebuilding the Encumbrance Chain (10 to 18 minutes)

The encumbrance certificates collected in Step 3 are now stitched into one chain. The engine looks for these points:

• Every charge placed on the property, be it mortgage, hypothecation or lien
• Whether a matching discharge entry exists for each charge
• Breaks in the chain, for example a transaction registered in year X that refers to no parent
• Registration at the wrong sub-registrar office, which often signals forged deeds
• Registered sale agreements for the property that never became sale deeds

Where records allow, the chain is rebuilt back to a root deed 30 years old, and any break is flagged for the reviewing lawyer to examine.

Step 6: Detecting Red Flags (18 to 22 minutes)

Risk patterns defined in advance are sought in everything collected, by a rules engine working with an LLM-based clause reader. Cases the engine detects unaided include these:

• Two registered transfers with the link document missing
• A power of attorney used for a sale with no check that it was not revoked by death
• No discharge of an encumbrance although the mortgage is 15 years old
• A sale by one co-owner without the consent of named legal heirs
• Boundary mismatch: the deed and the revenue record describe the property differently
• Sub-division of a survey number with no matching mutation order
• A live SARFAESI notice against the seller on any property
• A pending writ petition that affects the approval of the parent layout
• A project advertised by the seller although RERA registration shows "lapsed"
• A cash component disclosed in the agreement to sell but absent from the registered deed (stamp duty risk)

Step 7: Assembling the Report and Human Sign-Off (22 to 30 minutes)

Findings are organised into a structured Title Search Report. The usual AI-generated legal opinion runs to 29 sections, which cover identity of the property, chain of ownership, encumbrance status, exposure to litigation, statutory compliance and regulatory clearances, and close with a rating. A qualified advocate then checks the draft, confirms the red flags, uses legal judgement on unclear findings, and signs the TSR. That human step is what makes the result a legal opinion instead of a mere data extract.

Limits of an AI Title Search Engine

A seller who presents an AI title search as a full replacement for a lawyer is overselling. The engine is superb at retrieval, parsing, parallel work and rule-based pattern matching. It cannot take the place of legal judgement in situations like these:

• Fraud patterns that are new and match no historical signature, such as a family settlement deed discovered after the sale agreement
• Customary rights, including tribal land, joint family Hindu property and agricultural ceiling questions that need statutory interpretation
• Advice for negotiation: whether a flagged risk should end the deal or be priced in
• Preparing indemnities, special conditions or escrow arrangements from the findings
• Appearing in court if a flagged case turns against the buyer
• Local custom, such as the unwritten practice of one sub-registrar office that only an old-timer lawyer knows

Pair the engine with a lawyer and do not put it in a lawyer's place. About 80% of the work, which is mechanical retrieval and pattern detection, goes to the engine. The other 20% needs judgement and goes to the lawyer, who now has time for it since the mechanical part is done.

What the Output Contains

Among the sections of a 29-section TSR from an engine like LegiScore are:

• Executive summary with the risk rating
• Property schedule with the identity reconciled
• KYC trace and profile of the vendor
• Chain of title listing transfers by date
• Findings on encumbrance, year by year
• Litigation findings arranged court by court
• Statutory compliance covering RERA, building plan and occupancy certificate
• Status of revenue records: patta, mutation and ROR
• Tax compliance: property tax, water and electricity
• Findings on power of attorney, if there is one
• Verification of boundaries against the revenue map
• List of red flags with severity
• Documents to obtain before registration
• Indemnities to include in the sale deed
• The advocate's final opinion and signature

Specialised checks fill the other 14 sections: SARFAESI, IBBI and MCA charge searches, lis pendens, family-tree validation, attachment orders and statutory dues.

Side-by-Side Timeline

• Document collection · By hand: 0.5 to 1 day (physical visits) · With the engine: 5 minutes by upload
• Reading and identity matching · By hand: Worked out mentally by the advocate · With the engine: 3 minutes, deterministic
• Fetching the sub-registrar EC · By hand: 0.5 to 2 days · With the engine: 2 minutes via API or scraper
• Fetching revenue records · By hand: 0.5 to 1 day · With the engine: 2 minutes
• Case search on eCourts, every relevant court · By hand: Rarely done for each court · With the engine: 5 minutes for all 28 states
• Remaining portals (RERA, MCA, SARFAESI, IBBI) · By hand: Often skipped · With the engine: 5 minutes together
• Encumbrance chain reconstruction · By hand: 0.5 day (table made by hand) · With the engine: 3 minutes
• Red flag scanning · By hand: Subjective and dependent on the advocate · With the engine: 2 minutes, rules + LLM
• Report drafting · By hand: 0.5 to 1 day · With the engine: 3 minutes
• Lawyer review and sign-off · By hand: 0.5 day · With the engine: 5 to 10 minutes
• Total · By hand: 3 to 7 working days · With the engine: Under 30 minutes

Note that lawyers in India already disagree 34% of the time on one property. Carelessness is not behind it: every manual search covers a different set of portals and uses its own subjective rubric. The engine, being deterministic, performs the same 20-plus checks on every run, and so its output is open to audit.

Parallelism: The Structural Advantage

One design choice gives an AI title search engine its speed, and it is parallel execution. At no single step is the engine faster than a person. It simply performs many steps at once.

Take a sub-registrar query of 90 seconds, an eCourts query of 4 minutes and an MCA21 query of 2 minutes. A person doing them in turn needs 7 minutes 30 seconds, besides travel. An engine running them together is done in 4 minutes, which is the length of the slowest query, and the gap grows with the number of portals. With 20-plus portals to query for each property, the difference is that between 30 minutes and three days.

This also explains why the engine finds court cases that manual searches miss. Under deadline pressure, a human advocate will check the district court where the property lies and maybe the high court. The 18,000-plus courts across 28 states, in which the seller may have been named as a respondent, are out of reach.

Where AI Title Search Engines Fit in the Workflow

Between the agreement-to-sell and registration, a homebuyer uses the engine, and the TSR decides whether to go ahead, renegotiate or walk away. A bank or NBFC places it before sanction, as a vendor input to the credit committee alongside the technical valuation. For a developer who is aggregating land, the run happens at LOI stage, ahead of earnest money. An NRI buyer in Dallas or Singapore swaps the old process, in which a relative runs around Hyderabad for a week, for an upload-and-report flow.

In each case the engine supports a lawyer and does not replace one. The lawyer signs the opinion, and is simply spared a week of chasing photocopies.

Related reading:

• Property due diligence checklist for India
• Title chain verification: 13 year vs 30 year
• How to check pending court cases on property via eCourts
• Encumbrance certificate: complete guide for India
• Red flags in property documents every buyer must spot