CounselCheck

CounselCheck - Designed around intelligent AI routing and cost optimisation — not cost reduction. Verification should be an independent or differently structured layer, using authoritative sources, deterministic checks and human review wherever required. Because an LLM is not inherently an independent verifier of its own output. In fact, asking the same model to check its own answer can reproduce or reinforce the same error.

**Issues addressed by Counsel Check **- AI Infrastructure + Cost + Use + Human Intervention AI COST = Cloud cost + Tokens + API Calls + Model choice + Context The Proposition - If lawyers use the most capable model for every task — research, drafting and verification, the results will be brilliant, but the usage and costs can increase dramatically. The question should not simply be "what is the smartest model available?” but "what is the least expensive model that can reliably solve this particular legal problem to the required standard?”

Why CounselCheck? Route simple tasks to smaller models. Reserve expensive models for complex reasoning. Control context and token use. Cache repeated requests. Escalate high-risk outputs to humans. The goal isn’t cheaper AI. It’s intentional AI spend.

The judgment that proves AI hallucination is a real and persistent problem On 2 July 2026, the Supreme Court of India set aside NCLT and NCLAT decisions in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. Six of the authorities the lower court relied on were fake, misattributed, or built on invented paragraphs. The Court called it a “zero-tolerance” matter and directed the Bar Council of India to act.

A wholly invented case is relatively easy to catch: any citation checker flags a case that doesn’t exist.** What’s genuinely dangerous is a real citation carrying a fabricated paragraph**. This was not a chatbot mistake in a blog post; it was hallucinated legal authority relied upon by a tribunal. This is where CounselCheck sits: as an assurance layer between AI-generated work and the lawyer who ultimately takes responsibility for it.

• A 2025 Manupatra survey shows where India sits within this pattern: ~60% of lawyers already use AI, but only 4.1% trust its output unchecked, 51.2% identify hallucination as a barrier, and just 11% operate under a formal AI policy. Lawyers are not avoiding AI; they are adopting it without a reliable verification layer. CounselCheck is designed to be that layer.

What CounselCheck does CounselCheck does not draft. It provides an independent verification architecture for AI-generated legal work, regardless of where that work originated.

1. Upload the memo, opinion, or submission, from any tool, any author.
2. Extract every claim and citation.
3. Verify each one against a dedicated Indian case-law and statute bank (not general web knowledge) inside a Trusted Execution Environment, so client data stays confidential even while it’s being processed.
4. Audit whether the cited paragraph actually supports the claim, and whether the authority is still good law.
5. Report: a concise, source-linked exception report, labeled Verified, Needs review, or Couldn’t confirm. Never a bare confidence score standing in for a lawyer’s judgment. Every flag links to the source. Nothing is taken on faith, including from CounselCheck itself.

Why the gap is real Adoption is accelerating, not plateauing. Cyril Amarchand Mangaldas has adopted Legora. Shardul Amarchand Mangaldas and Fox Mandal have adopted Harvey. Trilegal is running Lucio and Relativity aiR on sensitive investigations. As more AI-drafted work enters the pipeline, the real question stops being can we draft faster and becomes can we trust what we just drafted, fast enough to still bill for it?”

That shifts the bottleneck from drafting to verification. CounselCheck does not care where the work came from either Harvey, Legora, ChatGPT, Claude or a first-year associate - because the audit only means something if it is independent of the source.

Human Accountability CounselCheck is deliberately designed to keep the lawyer in the loop. AI can identify issues, verify sources and surface risks; it does not take professional responsibility for the final work product. Where the risk or complexity warrants it, CounselCheck escalates the matter for human review and deliberately leaves gaps that require legal judgment. The lawyer, not the AI, remains accountable for what goes out the door. This also means juniors continue to draft, question, analyse and build judgment rather than simply accepting an AI-generated answer.

CounselCheck prototype is built by me in just 3 days. The limitations below are intentional - they're what's next, not what's missing by accident.

I built CounselCheck around a simple fact: a law firm is a business, and its lawyers generate revenue by billing hours. Any AI tool that wants a place in that business has to make sense on both sides of that equation - for the firm's economics, and for the lawyer actually doing the work. With more time, CounselCheck would grow into infrastructure the firm can actually run on, not just a tool a few curious Partners and Associates ty once.

How it benefits the law firm

• Secure environment. Verification happens inside a Trusted Execution Environment — a setup where the firm's data stays confidential even while the AI is processing it.
• Lower-cost alternative. Many firms already pay for expensive tools like Harvey to draft and refine work. CounselCheck is built to do the verification half of that job at a fraction of the cost.
• Private environment per client. Each firm — or even each client matter — gets its own isolated space, so nothing crosses between them.
• A dedicated data bank of Indian case law and statutes, so verification is checked against a source built for Indian practice, not general web knowledge.

I'm also realistic about where the profession is: most lawyers are AI-curious, a smaller group is AI-literate, and only a small fraction are AI-native. A tool built for the profession has to work for the first group, not just the last.

How will CounselCheck benefits the lawyer using it

-** A built-in prompt library**. Most lawyers aren't trained to prompt well, and learning by trial and error burns tokens and money while they figure it out. I give you a working starting point instead - a prompt bank for several use cases.

• Minimal steps to verify. The check should take less effort than the doubt it removes. Whether you're an associate, senior associate, principal associate, counsel, or partner, whatever your age or comfort level with AI, you should be able to use it without a learning curve.
• An honest question about billing. Should a lawyer be able to bill for time spent reviewing AI-generated work? I think that conversation needs to happen and a tool that makes verification fast and visible makes it easier to answer.
• A better way for junior lawyers to actually learn. If juniors keep drafting themselves and use CounselCheck only to verify rather than letting AI draft for them they build real judgment and confidence, instead of skipping the step where that skill is formed.
• More time for the parts of the job AI can't do. The goal isn't to automate lawyering, it's to free up time for judgment, strategy, and the arguments only a human would think to make. CounselCheck is deliberately simple: it's a balance between human and AI, not a replacement of one by the other.
• Human accountability, by design. Wherever CounselCheck falls short, that gap is a feature, not a bug — it keeps the lawyer, not the AI, accountable for the final work.
• Grounded in how AI can check itself, not just answer. Andrej Karpathy's "LLM Council" project showed that when multiple AI models review and rank each other's answers, and a designated "chairman" model synthesizes the result, you get outputs with far fewer blind spots than trusting any single model's first answer. CounselCheck is built on that same instinct, applied to legal work instead of general answers: instead of trusting one AI's draft, you put a second, independent AI in the room whose only job is to check the first one's work. It doesn't matter how confident the drafting AI sounded - CounselCheck is the chairman that reviews the claim against the actual source before anyone signs off on it.

As a 3-day prototype, CounselCheck currently runs as a single verification pass, not the full council architecture. With more time, the natural next step is a true council-style setup, multiple models independently verifying the same clause, cross-checking each other's findings, before a final check is issued. That's next!