Land Records Digitisation Meets Retrieval AI

A vendor slide in a state revenue meeting showed a single search bar and the words 'natural-language land records'. An officer typed a village name and a father's name. The demo returned a clean khata. The officer who actually sat in that tehsil said the village had three hamlets, two survey systems, a pending partition, and a register that had been burnt in 1998 and reconstructed from oral evidence. The demo did not know that. The demo was not lying. It was from another state.

This is the first fact to write down. Land-record digitisation in India is a family of programmes, not a product. The Digital India Land Records Modernisation Programme and its predecessors funded computers, scans and some cadastral work. What citizens see as Bhulekh, Bhu-Naksha, or a state-branded portal is a local stack with a local legal meaning. Retrieval AI can sit on top of that stack as a finder. It cannot nationalise title.

This field guide is for revenue CIOs, settlement officers and NIC state teams. It is dated 17 August 2026. It is not legal advice and it is not a title opinion. If your state has already notified a digital-only mutation path, read that notification. If it has not, do not let a model invent one.

What 'digitised' actually means in a tehsil

A record can be scanned and still be wrong. It can be typed and still be stale. It can be on a portal and still be only informative. Officers already know this. Vendors sometimes do not. Before you index anything, classify each artefact: certified current record, historical scan, uncertified extract, cadastral image, mutation order, stay order, and oral or reconstructed register. Retrieval should know which class it is showing. A citizen-facing answer that mixes a 2004 scan with a 2026 khata is how families start fights.

Legacy quality is the silent killer in this domain, which is why this cluster has a companion teardown. Survey numbers that collide. Duplicate khatas. Transliteration of the same name into three scripts. Village codes that changed after a district split. If you embed that mess and then let the model 'resolve conflicts', you have asked a statistical system to perform a settlement. Settlement is a legal process. It has notices and hearings. It does not have a temperature parameter.

• Index what you have. Do not pretend the index is the title.
• Show the artefact class and date on every retrieval card.
• Keep mutation, partition and stay orders in a separate corpus with a red banner.
• Never let the agent write back to the system of record.

Mutation is an officer, not a PATCH request

Mutation updates the revenue record to reflect a transfer, inheritance, partition or similar event. The details differ by state. The constant is that a competent officer acts after a process the manual describes. An agent that prepares the file — death certificate present, registered deed hash, notice draft, field report skeleton — is useful. An agent that 'applies the mutation' because the documents 'look complete' has altered a public register without authority.

Write a hard prohibition into the standing order and into the tool layer. The land-records system of record accepts writes only from named official identities. The agent account is read-only. If a vendor needs write access 'to demonstrate closed-loop AI', they are asking to be the tehsildar. Refuse.

Retrieval jobs that belong in a tehsil, and jobs that do not.

Job	Allow	Refuse
Find khatas matching a name and village	Yes, with class and date shown, duplicates visible	A single 'the' record when the index is messy
Draft a mutation noting	Yes, from the local manual	Auto-sanction or auto-write to Bhulekh-like tables
Explain a stay order	Retrieve the order and quote it	A model summary that softens the stay
Cadastral 'AI measurement'	Only as an aid the surveyor signs	A new boundary the portal publishes as official
Citizen chatbot on title	Point to the certified extract path	A chat that says 'you own this'

Two tehsils after the same digitisation grant

Objections you will hear — and what to do with them

If it is on the portal it is already public, so any model is fine

Public to inspect is not public to train on, not public to copy into a foreign host, and not a licence to emit a new statement of title. DPDP still cares about identifiable people. Your revenue law still cares about who may certify.

We will never finish mutation if officers must keep signing

Then staff hearings and fix document completeness, which is the same lesson as scheme eligibility. Mutation delay is usually notices, field reports and stays, not a shortage of machine writes.

A national land API is coming, so we should wait

You should not invent one, and you should not freeze a tehsil waiting for a product you cannot name. Index what you hold. Keep write paths closed. If a later national service appears, you will have clean artefacts to connect. If it does not, you still have a finder.

OCR confidence is 99 percent so we can retire paper

Ask what 99 percent means on survey numbers. A one-percent error on a number that identifies a plot is not a rounding error. Keep the parent image. Let the officer accept the extract.

A four-week playbook for one tehsil

1. Week 1: inventory artefact classes. Write which ones are certified. Write which village-code version is live after the last district split.
2. Week 2: build a retrieval index officers can use. Disable any write credential the SI proposed. Prove the disable with a refused test write.
3. Week 3: run fifty known-ambiguous name searches. If the system returns a single card, it is not ready. Ambiguity is the honest output.
4. Week 4: attach the agent to mutation packet assembly only. Sample ten packets. Write the standing order that the chatbot, if any, may point to the certified-extract path and may not say 'you own'.

File note you can paste

Subject: Retrieval assistance on land records — no mutation by software.

This department notes that land-record systems in this state are not interchangeable with similarly named portals elsewhere. An agent may retrieve and display artefacts with their class and date, and may draft mutation notings for a competent officer. It shall not write to the system of record, shall not issue a statement of title, and shall not send records to a public model API. Contested or multiple matches shall be shown as multiple. This note does not alter the land-revenue manual.

This note is an internal aid. It is not a title opinion or legal advice.

Prcept AI will keep the agent read-only on this class of system. We will not train on your registers. If a competitor offers 'auto-mutation with human oversight' and the oversight is a nightly email, they are asking to operate your title register. Ask them to show the refused write.

What to tell a citizen who asks the chatbot

If you insist on a public surface, script it as a pointer, not as a title lawyer. The allowed sentences are few. Here is how to obtain a certified extract. Here is the office hours of the tehsil. Here is the list of documents the mutation manual currently names. Here is a warning that a chat answer is not a record of rights. The forbidden sentences are also few and more important. You own this. Your name is on this khata. The stay does not apply. The neighbour's claim is weak.

When the index returns two cards, the chatbot must return two cards. A single 'most likely' owner is how a family feud becomes a government statement. If the artefact is reconstructed or uncertified, say so in the first line, not in a footnote a feature phone will crop.

Log every citizen query as a potential personal-data record. Do not keep the chat longer than the retention map. Do not use chats to fine-tune a public model. Land already makes people angry. A leaked chat about a pending partition will not make them calmer.

This article is informational field guidance for Indian public institutions, not legal, revenue or title advice. Confirm against the state's land-revenue law and manual, DILRMP instructions as they apply locally, DPDP, and counsel before you file it.