The durable moat in Indian legal AI is not the model, it is the data. As frontier models converge and inference costs collapse, the lasting advantage shifts to whoever owns the structured, refreshed, entity resolved court data that every model has to call. In India that substrate has to be built from scratch, and it compounds with every use.
The model layer is commoditising in slow motion and most of the people building AI products are watching the wrong race. The story was supposed to be Claude versus GPT versus Gemini. The actual story is that frontier capability is consolidating into a small number of foundation models that all hit roughly the same ceiling on roughly the same tasks. The model wars matter less every quarter. What matters more is what those models can call out to.
For vertical AI categories, especially in regulated, jurisdiction specific markets like Indian law, that shift moves the durable moat away from the model and toward two things. Proprietary structured data the model has to call. Distribution that makes the model worth opening. This post is about the first of those, what a real data moat looks like in court data, and why the team that locks one in first becomes the substrate every legal AI agent in the country eventually depends on.
Why the model layer is commoditising
Three signals point in the same direction.
Frontier capability is converging. The gap between the top models on the standard benchmarks has shrunk to single digit percentage points on most reasoning, coding and instruction following tasks. Each major release does not feel like a generational leap any more. It feels like a six month incremental.
Inference costs are collapsing. The cost per million tokens has dropped by roughly an order of magnitude in recent years across all major providers, and the decline is continuing. Open weight models are catching up on the parts that matter for vertical applications, including legal reasoning, long context handling and citation work.
Switching between models is becoming routine. Most serious AI applications now run on routing layers that can swap Claude for GPT for Gemini based on cost, latency or capability. The infrastructure layer treats models as interchangeable.
None of this means the models are not improving. They are. It means the differentiation per dollar is shrinking. For a product company sitting on top of the model layer, the implication is direct. Building a moat at the model layer is no longer realistic for a non frontier company. The moat has to live somewhere else.
What “data moat” actually means in this market
The phrase data moat gets used loosely. In a regulated vertical like Indian law, it has a specific shape. The team that owns the structured, refreshed, entity resolved, vernacular ready, cross court legal corpus has a moat that compounds with use and does not erode with model improvements.
Break the moat down into its components.
Coverage. Every court that matters. The Supreme Court, 25 High Courts and their forty plus benches, seven hundred plus district court complexes, the major tribunals, the specialist forums. In raw scale that layer now holds more than twenty eight crore structured records: over a hundred crore orders, sixty seven crore litigant records, twenty nine lakh advocate profiles and seventy seven thousand judge profiles, all linked and refreshed as new orders publish. A product that covers only the appellate tier is not useful to the large majority of the bar that argues at the district tier. A moat is the team that covers all of it.
Freshness. Daily refresh at minimum. Sub hour refresh on cause lists for the next working day. Cause list data goes stale within a few hours. Order uploads happen continuously. The team that has the pipeline to keep everything current owns the freshness moat. The team that scrapes weekly does not.
Structure quality. Raw HTML and PDFs are not data, they are signal that has to be turned into data. Structure quality means consistent field schemas across courts, parsed bench composition, parsed section citations, parsed relief granted. Without structure, the LLM has to do the parsing work itself, which is slow, expensive and error prone.
Entity resolution. One advocate is one entity, not three. One company is one entity, not eight. Across all courts, across spelling variations, across name changes, across mergers and demergers. This is hard, never finished, and absolutely the difference between a useful API and a noisy one.
Vernacular OCR. Hindi, Marathi, Tamil, Telugu, Bengali, Kannada, Gujarati, Odia, Punjabi. Scanned PDFs that go back decades. The team that has invested in vernacular at production grade serves the entire bar. The team that has not, serves the metro tier only.
Historical depth. New filings are interesting but the bulk of legal value is in the backlog. A counterparty’s litigation history matters more than its most recent case. A judge’s order pattern over five years matters more than today’s listing. Depth is built only by going far back, and the longer the platform has been pulling, the deeper the corpus gets.
A real moat has all six. The combination is what we wrote about in The Operating System for Indian Law. The honest cost of building each of these to production grade is what makes the moat real.
Why the moat is harder to copy than people think
The first reaction from anyone hearing this for the first time is that scraping is easy and AI makes data extraction trivial. Neither is true at scale in Indian law.
Scraping does not survive the captchas, the rate limits and the schema variations across twenty five plus portals. Anyone who has tried to run a scraper across the Indian court ecosystem for a quarter has learned that the maintenance burden compounds faster than the data accumulates. Most teams give up after the second portal change.
AI extraction works on the document level. It does not solve the fragmentation problem. You still need a pipeline that reaches the document, retrieves it under captchas and rate limits, and turns it into a record linked to the case, the bench, the parties, the advocates, the prior orders and the next listing. The AI does some of the OCR work. The rest of the pipeline is engineering. There is no shortcut.
The honest replication cost for a second team to build a comparable structured layer in Indian court data, starting from zero today, is twelve to eighteen months of focused engineering work with a small specialist team. We made that estimate in earlier posts and it has not moved. By the time that hypothetical second team finishes, the first team has another year of depth, another year of paying enterprise contracts, another year of agent traffic and another year of vernacular coverage. The moat is widening while the copy attempt is being built.
Why the moat compounds with use
This is the part that gets missed in the model layer conversation. A data moat is not a static asset. Every API call that an enterprise buyer makes, every MCP (Model Context Protocol) invocation that Claude or GPT makes, every case file that a lawyer uploads to the AI Clerk surfaces edge cases that improve the schema. Every new judge name learned, every new advocate spelling resolved, every new tribunal added is permanent value that the next user inherits.
The model layer does not work like that. A frontier model that is six months newer is better. A frontier model that has more users does not get better from the use. The data layer is the opposite. More use means more coverage, more freshness signals, more entity resolution, more quality. That is the textbook definition of a network effect at the substrate level.
We wrote about how this looks operationally for the in house counsel who runs portfolio monitoring through Claude in Litigation Portfolio Monitoring for General Counsel: A Claude + eCourtsIndia MCP Playbook. The General Counsel never sees the data layer. The Claude window is the only surface. But every query she runs makes the substrate underneath sharper for the next user.
Why this matters more in India than in the US
The American legal AI category, dominated by Harvey, can afford to underweight the data layer for one reason. The US already has Westlaw, LexisNexis and PACER. The structured legal corpus exists. Harvey can plug into it and focus on the application layer.
India has no such substrate. The work of building it is, in effect, a precondition for the application layer. Whoever builds the application layer in India without owning the substrate underneath is dependent on a substrate that someone else has to build, sell to them and price. That dependency is not where the value accrues over the next decade.
This is also why we keep coming back to the point that Harvey and Legora will not build for India. We covered the structural reasons in India’s Legal AI Will Be a Ten Billion Dollar Category. The short version is that the substrate is the whole game in India in a way it is not in the West.
What the moat looks like in five years
Project forward. In five years, the Indian legal AI category will have one or two clear platform leaders and a long tail of agent layer products built on top of one or two structured data providers. The agent layer will be fragmented and competitive. The substrate layer will be concentrated and structurally profitable.
This is the same shape that played out in payments, where NPCI sits underneath PhonePe and Paytm and Google Pay. Same shape in identity, where the Aadhaar stack sits underneath every fintech and insurance KYC flow. Same shape in tax, where GSTN sits underneath every accounting product. The platform layer wins through scale and consistency. The application layer wins through user experience and distribution. Both can be valuable businesses. Only the platform layer captures the network effect.
The team that owns the Indian court data substrate in five years is the team that started early enough to lock in the data moat before the category caught on. The window to be that team is measured in months, not years, and it is closing.
What this means for eCourtsIndia
The model layer is commoditising. The model wars are moving from differentiation to convergence. In Indian legal AI, the moat that compounds and survives is at the data layer. Coverage, freshness, structure, entity resolution, vernacular and depth. We are building that substrate. Every agent built on top of us, from Claude to in house GC tools to fintech risk systems, deepens the moat. That is the durable position we are playing for.
TL;DR
- Frontier model capability is converging. Inference costs are collapsing. Routing layers treat models as interchangeable. The moat at the model layer is not realistic for a non frontier company.
- In vertical AI, the durable moat lives at the data layer. For Indian law, that means coverage, freshness, structure, entity resolution, vernacular OCR and historical depth.
- Replication cost for a second team is twelve to eighteen months of focused engineering. By the time the copy is built, the original has compounded another year of depth, freshness and use.
- Data moats grow with use. Every API call and MCP invocation sharpens the substrate. The model layer does not work that way.
- Five years out, the Indian legal AI category will look like payments, identity or tax. Concentrated substrate layer underneath. Fragmented agent layer on top. The substrate is where the network effect accrues.
Sources
- Public benchmark convergence across frontier model families, 2025-2026
- Anthropic, OpenAI and Google pricing pages for inference cost trajectory
- Tracxn India LegalTech 2026 funding report
- TransUnion CIBIL data layer valuation reference
- All court coverage figures verified against the eCourtsIndia structured data index
Read next: The Operating System for Indian Law and MCP and the Agent Layer.
Frequently Asked Questions
Why is the AI model layer becoming commoditised?
Frontier capability is converging to single digit benchmark gaps, inference cost per million tokens has dropped by roughly an order of magnitude in recent years, and routing layers now swap Claude, GPT and Gemini freely. Differentiation per dollar keeps shrinking, so a non frontier company cannot build a durable moat at the model layer. We unpack this in The Operating System for Indian Law.
What does a data moat actually mean in Indian legal AI?
It is ownership of a structured, refreshed, entity resolved, vernacular ready, cross court legal corpus. The post defines six components that a real moat needs: coverage of every court, daily freshness, structure quality, entity resolution, vernacular OCR and historical depth. You can explore that structured corpus directly through eCourtsIndia case search.
How long would it take a competitor to copy the data moat?
The honest replication cost for a second team starting from zero is twelve to eighteen months of focused engineering with a small specialist team. By the time that copy is built, the original has added another year of depth, freshness, paying contracts and agent traffic, so the moat widens faster than it can be matched. More context is in India’s Legal AI Will Be a Ten Billion Dollar Category.
Why does a data moat compound with use?
Every API call, every MCP (Model Context Protocol) invocation by Claude or GPT, and every case file uploaded surfaces edge cases that improve the schema. New judge names, advocate spellings and tribunals become permanent value the next user inherits. The model layer does not improve from use, the data layer does. See it for yourself with an advocate lookup.
Why does the data moat matter more in India than in the US?
The US already has Westlaw, LexisNexis and PACER, so the structured corpus exists and application products plug straight in. India has no such substrate, so building it is a precondition for the application layer, and whoever owns it captures the network effect. Research a judge order pattern to see the depth in practice.