MASTER PLAN
What Indian law firms will operate on next.
Methodology, substrate, and operational integration that Indian firms will use to operate AI-native. The first design partnership is in motion; the substrate is in production with the agent that runs on it.
Written by Rohan Shiralkar · Founder · 1 May 2026
THE SHIFT
What just changed.
In 2025, long-running agents crossed from research into production. The architecture around the model (what runs it, what feeds it, what holds context across tens of minutes of work) became the load-bearing engineering. Anthropic published the pattern in November 2025 and industry consensus crystallised inside six months. The model is now commodity inside the architecture. A different kind of AI became possible.
In parallel, AI-native firms began emerging. Not “law firms that use AI” but firms architected around agents. Norm Law raised $50M from Blackstone. Crosby is a registered agentic law firm with malpractice insurance. Freshfields signed a multi-year deal with Anthropic across 33 offices, adoption growing 500% in six weeks. None work on Indian law. Indian firms haven’t reorganised yet; the structural opportunity is open in a way it isn’t in the US or UK. The moment for an Indian AI-native practice is now.
THE BIFURCATION
Two categories. Not one continuum.
Everyone in law is in the doc-Q&A chatbot era, treating AI as another SaaS tool. We treat it as a different category of work. Not two points on a line. Two categories, doing different jobs, for different buyers, under different economics.
The split is clearest in what each kind of system can actually do. Five tasks, in order of depth:
Section lookup
“What does Section 11(2) of FCRA say?” Chatbots handle this fine. So does Indian Kanoon’s free search. Table stakes.
Amendment summaryChatbots break
“Summarise the 2019 amendments to FCRA and what changed.” Chatbots break here. They hallucinate the chronology, conflate principal-Act sections with rule sections, miss notifications. The chatbot-over-paywalled-database failure mode begins here.
Single-proposition research memoClauseo today
“Research memo on whether a foreign-funded NGO can accept ₹X under the post-2020 amendments.” Synthesis across statute, amendment chain, caselaw, and regulatory practice. No single retrieval call surfaces it. This is what Clauseo today does.
Multi-proposition research memo
“Map every contested proposition in this 200-page brief, research each independently, synthesise across them.” Parallel investigation across many concurrent threads with structured intermediate state. Clauseo today is scoped down for this; the architecture supports it.
Continuous compliance research at firm scale
“Build me an internal compliance research agent that runs continuously across our firm’s 50,000 historical matters, surfaces drift from regulatory amendments, flags anything in active matters that needs revisiting.” What the architecture unlocks at full deployment. No chatbot can reach this.
The bet: that fifth task is what reshapes legal practice. Work no human team and no chatbot can produce today. The intermediate rungs fund the path.
THE BET
Four pieces. Each necessary. Each fails alone.
A research methodology built around long-running agents.
Long sessions where the agent works through a problem step by step, not in single shots. Parallel investigations that build on each other instead of running independently. Quote-with-paragraph-anchor citation. Gap-checking against legal-element checklists. The artefact is more thorough than any junior could produce, more traceable than any researcher would bother with, qualitatively different from any chatbot output.
A structured Indian legal substrate.
Built domain by domain. Regulatory orders need analytical enrichment. Statutes need a navigable structural representation that PDFs destroyed. Caselaw needs custom parsing and citation networks. Firm knowledge needs per-firm structuring. The methodology only matters because the substrate gives it terrain to walk on.
Frontier on every layer.
The most capable model the system can reach for, in a system built to make that reach economical. Each layer’s frontier shifts every few months; standing still on any one means doing what was possible a year ago, which is what every chatbot product does.
Operational integration with firms.
Outputs, methodology, substrate woven into how firms operate, not bolted on. The agent becomes part of the operational fabric the way a colleague would be: documents, communications, calendars, case management. Otherwise the system produces excellent memos that lawyers manually transplant into actual work.
THE SUBSTRATE
What the agent stands on.
Recover the structure PDFs destroyed.
Statutes are hierarchical and networked: chapters, sections, subsections; this section amends that one of that other Act. Judgements are structured discourse: facts, issues, submissions, reasoning, ratio, orders. PDFs annihilate both. Custom representations recover them. A wall of text becomes a navigable graph.
Extract analytical fields the source did not have.
Ground categorisation, party disposition, penalty amounts, market definitions, citation sentiment. Extraction compute amortises across every future query that filters on those fields. The agent filters against fields we extracted in advance, not raw text.
Map citations between documents.
Act A § 5 cites Act B § 12 as a machine-readable edge, not a string. CCI Order #823 cites NCLAT Appeal #117 with sentiment (upheld, set aside, remanded). The substrate becomes a graph the agent traverses, not a flat search index.
Track every version through time.
Point-in-time queries against the structural representation. Show me Section 11(2) of FCRA as it stood on 22 July 2019. Statutes are moving objects; every order, judgement, advice memo is written against a specific version. Point-in-time analysis becomes a primitive.
Each layer is independently valuable; together they make the economics work. The heavy structural and analytical work happens once, during substrate build; every session draws from the result. Compute the agent spends at query time goes into reasoning, not parsing. The compute that built the substrate amortises across every session that ever runs against it; the marginal cost of the next session is the agent’s reasoning, not the parsing it would have had to do without the substrate. Chatbots cannot match this depth at chatbot prices; we don’t need to.
PHASE 1 · DONE
Proof of architecture, in production.
The Clauseo agent at clauseo.chat.
Private beta. Sessions run 15 to 35 minutes typically, up to 60 when the question demands it, on Anthropic’s most capable model (Opus 4.7), inside the architecture above.
Indian competition law, structured.
2,800+ CCI orders and 425+ NCLAT appeals extracted, classified, linked appellate-back-to-original. Six analytical views designed around what a competition lawyer asks. 24 documented data-quality caveats. Free public layer at competition.clauseo.chat.
FCRA in Akoma Ntoso.
The Foreign Contribution Regulation Act and its rules in the international XML standard the UK Parliament built legislation.gov.uk on. Every amending notification merged. Viewer at fcra.clauseo.chat.
Paying customers and the first design partnership.
A handful of paying customers, all senior advocates at top-tier Indian firms; sessions billed at ₹150 to ₹1,500 land on the client invoice next to court fees. The first design partnership is in motion: a partner firm whose internal documents we are structuring.
We run AI-native ourselves.
Iris is a long-running agent on our team chat, with her own GitHub, Workspace, and Gmail. Before we ask Indian firms to operate AI-native, we’re answering the question for ourselves: what changes when an agent becomes a colleague rather than a tool. The internal proof for the operational-integration claim above.
RAW NCLAT FILING
- Case No.
- Competition Appeal(AT) - 4/2023IA/106/2023IA/107/2023IA/108/2023IA/425/2023Caveat/451/2022
- Filed
- 2023-01-09T18:30:00.000Z
- Status
- Disposed
- CCI Source Order
- —
- Penalty
- —
- Bench
- Justice Ashok Bhushan (Chairperson)Hon’ble Mr. Barun Mitra (Member (Technical))
- Sector
- —
STRUCTURED CLAUSEO DATABASE
- Case No.
- Comp. Appeal (AT) No. 4 of 2023
- Filed
- 9 January 2023
- Status
- Disposed · Appeal partly allowed
- CCI Source Order
- CCI Case 07/2020 · 24 October 2022
- Penalty
- ₹936.44 cr → ₹216.69 cr
- Bench
- Justice Ashok Bhushan, Mr. Barun Mitra
- Sector
- Mobile app store, Android OS
PHASE 2 · IN MOTION
The deep-investigation thesis.
Phase 1 proved the architecture on a single domain. Phase 2 expands the substrate, the methodology, and the operational-integration prototypes across more domains and more than one design-partner firm. Eighteen months.
Substrate expansion.
Regulatory: SEBI in full (~82,000 orders), then SAT, IRDAI, RBI. Statutes: 10 to 15 heavily-litigated central acts in Akoma Ntoso (Companies, Income Tax, IT, GST, the rest sequenced by production demand). State acts deferred to year two. In parallel, a deep re-enrichment of the existing 8K-document CCI/NCLAT corpus against sharper extraction primitives. Each body needs its own analytical schema.
Caselaw substrate work.
Beyond the Indian Kanoon fallback, structured judgement databases per court. Custom parsing extends; citation networks deepen. Sequencing partly determined by where lawyer demand surfaces in production usage.
Operational integration.
Frequent contact and fast iteration with the design-partner firm. Two to three additional design partnerships sought during Phase 2.
Task types beyond research.
Drafting, review, regulatory monitoring, case-management hooks. Each surfaced by partnership work, each its own structural problem. None claimed to fall out automatically.
The envelope.
Phase 2 is funded by a $1.5M / ₹15 Cr pre-seed round at 18-month runway. The round is sized to do Phase 2 properly without compromising on any of the four layers.
| Line | Amount | Notes |
|---|---|---|
| Round size | $1.5M / ₹15 Cr | |
| Runway | 18 months | |
| Team | 7 people | |
| Inference budget | ₹7.1 Cr | Dominant line. Substrate compute (₹3.24 Cr), product / session compute (₹1.92 Cr), and CCI/NCLAT re-enrichment all sit inside it. Sized to be unconstrained on which frontier models the system reaches for. |
| Cash-only spend | ₹5.43 Cr | Salaries, Bangalore office, operations. Seven people across 18 months at Bangalore market salaries. |
| Subtotal | ₹12.53 Cr | |
| Buffer | 20% | |
| Structure | Cash-funded compute | No compute credits banked. If GCP / Anthropic / AWS / Azure grants land, pure runway upside. |
Seven people. Founding team: Rohan Shiralkar (founder, engineering; built the research engine), Shamil Rasheed (engineering; ingestion, database, search), Harshvardhan Mudgal (legal; source-search, lawyer conversations). Joining: a senior engineer from month six, a senior in-house lawyer from month three, two legal interns from month three. Bangalore office. Salary set at sustainable, not heroic, levels; frontier commitment includes the founder.
PHASE 3 · CONDITIONAL
Categorical reshape.
Phase 3 is conditional on Phase 2 producing the compounding it suggests. If substrate expansion holds, the methodology generalises to additional task types as design partnerships surface them, and the first design partnership matures into a model other firms adopt, then Clauseo is positioned to cover enough substrate to be useful across practice areas, productise methodology and substrate access for firms beyond the design partners, and support firms operating AI-native with their own agent integrations on the shared substrate.
Task categories that do not exist today open here: continuous compliance research at firm scale, multi-proposition synthesis across hundreds of matters. If those signals hold, Clauseo’s relationship with firms changes shape from tool to infrastructure. The exact form gets shaped by the partnerships.
Three signals tell us Phase 3 is opening.
- Substrate expansion compounding (each new domain weeks, not months).
- Two of three additional design partnerships maturing into ongoing arrangements.
- Methodology generalising to a second non-research task type.
The path has conviction. The destination is something we will discover by walking it.
WHAT WE STILL DON’T KNOW
Open questions, named.
What productised firm-side infrastructure looks like.
Substrate-as-a-service, session-based agent access, per-firm bespoke integration. The exact contracting, packaging, pricing emerges from the partnerships.
The end-state shape of the AI-native Indian firm.
A senior-heavy core with agents replacing the junior layer (the obelisk) is the leading hypothesis. Pyramid-with-an-agent-per-lawyer, ambient-mediated, fully-dissolved-boundary are alternatives. The shape emerges from observation.
How distribution scales beyond the first cohort.
The Indian senior-advocate market is small, socially gated, slow to adopt. The first cohort came through direct relationships at top-tier firms; Phase 2 tests whether the same warm-referral channel carries to the next 50, then 100, then 500 lawyers, with month-12 volume of partner-referred design conversations and senior-advocate signups as the success signal. If those don’t materialise without paid acquisition, distribution gets rebuilt before Phase 3.
Pace at which non-research task types open.
Drafting, review, regulatory monitoring, case-management hooks each have their own structural shape. Order and pace depend on what partnership work surfaces as the highest-leverage move.
What competitive responses look like.
Indian incumbents cannot pivot without abandoning their existing business; global per-seat AI-firm products cannot enter India without rebuilding. The real risk is something new: an Indian AI-native firm built from scratch, or a global AI-native firm setting up here. Mitigation is depth of methodology, substrate, and partnership relationships, which take years. Path B (no compute credits banked) also means full price-exposure to model-provider pricing changes.
How responsibility shows up in operation.
The agent doesn’t change the standard of care. Citations link to source so a lawyer confirms in seconds. Every tool call and reasoning step is on the page so review is fast and what was missed is visible. How firms operationalise that (review checklists, second-passes, partner sign-off) gets shaped by the design partnerships. We don’t have a settled answer.
Two to three more design partnerships during Phase 2. Or come use the agent on a real matter. Either way: rohan@clauseo.chat.
Rohan Shiralkar, founder, Clauseo.
Last updated 1 May 2026