20 September 2026 · २०८३ असोज ४ · 8 minutes
AI is already in the law office
Every law practice runs on one resource, and it is not brilliance. It is hours. Reading takes hours, checking a citation takes hours, and going through three thousand pages of discovery takes a week — after which a partner asks you to go through it again with a different question in mind.
That is the part AI is eating. Not the lawyering. The hours.
The legal press has got this wrong in both directions. First came the panic that chatbots would replace associates, which was silly. Then came the backlash: the fabricated-citation sanctions, the columns about how AI clearly cannot handle real legal work. Both readings missed what was actually being built, because the interesting products stopped trying to write your brief and started trying to understand everything that surrounds the brief.
The first wave failed for a specific reason
Early legal AI was judged on whether it could produce text. Could it draft a contract? Summarise a judgment? Write a research memo? Yes, mostly — and that turned out to be the least useful thing it could do.
The problem was not quality. It was that you could not check the work. A general-purpose model hands you a confident paragraph with a citation attached and no way to trace either one back to a source. Lawyers cannot use that. The whole profession is built on being able to point at something and say: here, this is where it comes from, this is the authority, this is the page in the record. A convincing answer and a defensible answer are different objects, and only one of them survives opposing counsel.
The tools worth looking at are built around that gap rather than around the writing. They sit inside the word processor, connect factual assertions to the documents that support them, build timelines out of a record, check citations, assemble tables of authorities.
The design philosophy matters more than any single feature. The goal is not to produce legal text. It is to let a lawyer process far more material without ever losing the thread back to the evidence. If the next decade of legal technology has a defining idea, that is probably it.
Adoption is messier than the headlines suggest
The American Bar Association's 2025 Legal Industry Report surveyed more than 2,800 legal professionals and found 31% personally using generative AI for work, up from 27% the year before. Firm-level adoption lagged, held back by worries about accuracy, ethics and policy.
Read that carefully, because it is a few points of movement, not a stampede. And the more revealing number is the gap between what individuals do and what their firms have approved. That gap is where the sanctions come from. An associate under deadline pressure pastes a passage into a consumer chatbot at eleven at night, gets a case name that does not exist, and nobody at the firm ever wrote a rule about it because the managing partner was still waiting for the technology committee to report back.
The risk is not that firms adopt AI recklessly. It is that they refuse to adopt it officially while their people use it anyway.
Law punishes error differently from other industries. A bad marketing email costs you a customer. A fabricated citation costs you your credibility with a judge who will remember your name, and an incorrect legal proposition can cost a client something they cannot get back.
The bar associations are doing the most important work
Here is the actual position. The bar guidance published over the past two years will shape legal AI more than any product launch, any funding round, and certainly more than any award.
The New York City Bar's Formal Opinion 2024-5 works through what generative AI means for confidentiality, conflicts, competence, diligence, supervision, client communication and candour toward tribunals. The Washington State Bar Association issued Advisory Opinion 2025-05 on AI-enabled tools and the professional conduct rules. The ABA has published on confidentiality, competence, verification and supervision of AI output.
None of that is endorsement. It is something more durable. Professional bodies are writing the conditions under which a lawyer can use these tools and still be defensible when something goes wrong, which is the only question that ever really mattered.
The State of Nevada's AI Workgroup went further and evaluated specific products, looking at what each could and could not do in practice. That is unusual, and slightly uncomfortable — bar bodies naming commercial products invites obvious problems. But it is also genuinely useful to a solo practitioner in Reno who has no way to evaluate eight vendors on her own.
Institutions are participating rather than observing. In January 2025 the American Arbitration Association's International Centre for Dispute Resolution announced a partnership covering roughly 5,500 arbitrators and mediators, following a six-month pilot. When the people who decide disputes start testing the tools, the conversation has moved past whether any of this is legitimate.
The awards matter less. They are a reasonable signal about who is doing the judging and what criteria the industry has settled on. They are not evidence that the software works.
The economics will not fix themselves
Now the uncomfortable part. Almost every article about legal AI ends up at access to justice, and almost none of them explain how we get there.
Traditional legal work ties revenue to time. More hours, more money. If a task that took ten hours now takes four, the firm has a choice, and we should be honest about which choice most firms will make. They will take on more matters and keep the margin. Some will invest the saved time in better analysis, which is the good outcome. A few will move to fixed fees. Very few will voluntarily cut what they charge people who currently cannot afford a lawyer at all.
Cheaper production does not become cheaper service without somebody deciding it should. That is a policy question and a professional-culture question, not a technology question, and pretending otherwise lets everyone off the hook.
There is also the comfortable idea that AI closes the gap between big firms and small ones. It might widen it. Large firms have the budgets, the data governance, the training programmes, the ability to negotiate enterprise terms, and enough matters to make a tool pay for itself. A three-person practice has a monthly subscription and no time to learn the software. The gap closes only where small firms get access to good tooling and someone teaches them to use it well, and that does not happen on its own either.
What actually changes about the job
Give a lawyer a three-thousand-page file and a machine can now surface the relevant documents, build a first chronology, link factual claims to their sources, flag gaps in the record, organise the authorities, summarise the depositions and produce a rough draft.
Then the lawyer asks the questions that still have no automated answer. Is this legally correct? Is this fact genuinely supported, or does the citation just look plausible? Is this authority still good law? Does this argument serve the client? Should we make it at all, in front of this judge, in this posture?
Those are judgment calls and they stay judgment calls. What changes is the ratio. Less time searching, formatting, and reading things that turn out to be irrelevant. More time thinking about strategy, which is what clients thought they were paying for all along.
The skill set shifts accordingly. Verification, data security, knowing where a tool's competence ends, writing prompts that produce something checkable, supervising juniors who are using these systems whether you approved it or not. The ABA has been explicit that lawyers need to understand both the capabilities and the limits of what they are using. Understanding law and understanding the technology used to practise it are becoming the same professional qualification.
And this is Nepal's problem
The global picture is clear enough. Bars are writing guidance, courts are dealing with AI-generated filings, arbitration bodies are running pilots, large firms are deploying at scale.
So the question here is not whether Nepali lawyers will use AI. They already do, quietly, mostly with consumer chatbots that know nothing about Nepali law and will invent a Supreme Court decision if you ask nicely. The real question is when Nepal's legal ecosystem builds something of its own.
And most people underestimate how much of that work is not AI work at all.
You cannot import an American legal AI product and expect it to handle Nepali law. A useful system would need our legislation, our judgments, our procedural rules, our citation conventions, our drafting practice in both Nepali and English, our forms, our institutions. That corpus exists. Thousands of judgments, statutes, gazette notifications, government decisions, land records, company filings, and decades of accumulated practice knowledge sitting in books, law reports and the heads of senior advocates.
Most of it is not machine-readable. Some of it is not reliably available in any form. Nobody has settled who owns it or who is allowed to build on it.
That is the bottleneck, and it is a digitisation problem, a data-rights problem and an institutional-will problem long before it is a modelling problem. Whoever solves the boring part first will own the interesting part later. Every month we wait is a month of judgments going into filing cabinets instead of into a searchable record.
This is the part we decided to work on. Likhat AI, made by Hardflip Technologies, is a drafting tool for Nepali advocates that turns a short brief into a first draft in Nepali with statute citations resolved against a real corpus. A lawyer reviews, a lawyer signs.
The profession spent decades digitising libraries, then documents, then research. Connecting all of it intelligently is the current chapter, and countries that build trusted legal AI early will not just get better software. They will change who can afford a lawyer.
AI is not knocking on the door of the legal profession. It has been in the office for a while now, using the good chair. The only open question is how carefully we decide to work with it.
Hardflip Technologies
Founder, Hardflip Technologies
20 September 2026 · २०८३ असोज ४
Product facts in this piece are as of 14 September 2026.
This tool does not give legal advice. Every draft is usable only after review by a licensed advocate.