The legal workflow, mapped to where AI helps

A week, in stages

Take a concrete example. Picture a six-partner commercial firm in Kathmandu that handles banking and finance disputes for two domestic banks and a handful of corporate clients. The associate we are following — call her Anisha — is in her third year, working on three live matters: a loan recovery suit at the High Court, a contract dispute heading into arbitration, and a regulatory advisory for a client navigating an NRB unified directive. A normal week for Anisha contains roughly the following stages, in roughly these proportions.

Monday morning, client intake (three hours). A new client arrives — the managing director of a trading company facing a guarantee call. Notes are taken in English and Nepali, documents are scanned, the facts are messy. Today: Anisha types up the intake memo by hand after the meeting and prepares an issue list for the partner. With AI: she dictates a rough narrative into the phone, pastes the transcript and the scanned documents into the model, and asks for a structured intake memo and an issue list in a format the firm uses. Time falls from three hours to about forty-five minutes, including her edit. New risk: the client’s name and financial details have now passed through a foreign tool; the firm’s confidentiality policy needs to have addressed this before, not after.

Monday afternoon, contract review (four hours). The corporate client has sent a draft master services agreement for review against the NRB directive. Today: Anisha reads the contract front to back, marks clauses against the directive, drafts a comment letter. With AI: she uploads both documents, asks for a clause-by-clause comparison highlighting potential conflicts, then reads the flagged clauses against the directive herself. Time falls from four hours to about ninety minutes. New risk: clauses the model did not flag may contain the real problem; the model is a screen, not a substitute for one careful read.

Tuesday, legal research (six hours, across the day). The arbitration matter turns on whether a particular liquidated-damages clause is a penalty under section 504 of the Civil Code 2074. Today: Anisha searches Nepal Kanoon Patrika, reads four to six decisions, prepares a research note. With AI: she uses the model to generate candidate arguments for and against, to draft the structure of the note, and to translate two Indian Supreme Court decisions she will rely on by analogy. She does the actual NKP research by hand, opening every cited case in the Supreme Court repository. Time falls from six hours to perhaps four — and importantly, the time falls in the structuring and translation, not in the citation work. The verification load does not move.

Wednesday, drafting (five hours). The High Court matter requires a written submission. Today: Anisha drafts from scratch using the firm’s templates, then circulates for the partner’s comments. With AI: she gives the model the case facts, the issues, the firm’s template, and the cases she wants relied on, and asks for a first draft. She rewrites the argument sections heavily, leaves the procedural sections largely as drafted, and verifies every citation. Time falls from five hours to about two and a half. New risk: confident-sounding paragraphs that misstate the holding of a cited case slip through if she edits too lightly.

Thursday morning, court appearance (four hours, including travel). Today: Anisha goes to the High Court, waits her turn, argues the matter, returns. With AI: same. The bench does not accept oral submissions generated in chambers by a chatbot, and even if it did, the act of standing up and responding to questions from the bench is the work the client is paying her firm to do. AI does not touch this stage.

Thursday afternoon, advisory work (three hours). The NRB directive needs to be explained to the client in writing, with practical implications. Today: Anisha reads the directive, drafts an opinion in English, has it translated into Nepali for the board. With AI: she uses the model to produce a structured summary of the directive (which she verifies against the source), drafts the opinion with the model’s help, and uses the model to produce a first Nepali translation of the client-facing portions. Time falls from three hours to about an hour and a quarter. New risk: the model paraphrases the directive in a way that softens an important obligation; she must read the bare directive herself before signing.

Friday, billing, file management, miscellaneous (three hours). Time entries, invoices, file notes, internal updates. Today: mostly manual data entry. With AI: the model drafts time entry narratives from her diary, drafts internal status emails from her notes, and produces a weekly summary for the partner. Time falls modestly. The risk here is low — the source is her own week and she is the only person who can verify it.

What the map shows

Add it up and the picture becomes clear. In a roughly twenty-eight-hour week of substantive legal work, AI compresses perhaps ten to twelve hours — the intake structuring, the contract comparison, the research drafting, the first-pass drafting, the advisory summarising, and the administrative typing. It does not compress the court appearance, the bench’s questions, the partner’s judgement on strategy, the client’s hand-holding through a hard decision, or — crucially — the verification of every cited authority and every quoted provision.

The honest reframing is this: AI gives Anisha back roughly a day of her week. The interesting question is what she does with that day. The wrong answer is “the same volume of work in less time” — that path leads, within a year or two, to a firm that has quietly lowered its standard of verification. The right answer is some mix of: more careful verification, deeper research, more client time, more time on the harder questions that previously got the short end of the week, and — for the firm as a whole — the capacity to take on the matters that previously fell through.

The same map for the sole practitioner

The picture is similar for a sole practitioner in Pokhara handling family-law matters — divorces under the Civil Code 2074, partition suits, guardianship petitions — even though the substance differs. Intake is heavy on emotional content the model handles awkwardly (use it only for the structural notes, not the empathy). Drafting petitions follows well-worn templates the model produces competently. Court appearance, mediation, and the careful conversations with clients about realistic outcomes are entirely unchanged. The compression sits in the same places: structuring, drafting, summarising. The protected hours sit in the same places: judgement, advocacy, and verification of the bare Act and the cited decisions.

What does not move, and why

Three categories of legal work are essentially AI-resistant, not because the technology is technically incapable but because the professional structure does not permit substitution.

Court appearance and oral advocacy. The bench addresses you, not your tool. Responding well to a judge’s question is judgement under pressure, and that judgement is not delegable.

The formation of professional opinion. Whether the client should fight or settle, whether the clause is enforceable, whether the petition is maintainable — these are the questions the client is paying you, by name, to answer. The model can list considerations. It cannot weigh them with your professional responsibility behind the weighing.

Verification of authority. Every section of an Act, every paragraph of a circular, every paragraph of an NKP case that goes into a submission with your signature on it must be opened in the source. The model can help you find what to read. It cannot do the reading on your behalf, because if it is wrong, the consequence falls on you.

The week, then, is not shorter. It is reshaped. The lawyer who understands the reshape — who deliberately moves the saved hours into the protected categories — is the lawyer whose work will be both faster and better than it was a year ago. The lawyer who treats the saved hours as pure productivity gain is the lawyer who will, sometime in the next year, sign a submission containing a Supreme Court citation that does not exist.

Check your understanding

Quick check

—

In a typical week for a Nepali commercial associate, which stage is most fairly described as HIGH AI leverage — a stage where AI can realistically compress hours of work, given proper verification?

1. AProducing a clause-by-clause comparison of a draft contract against an NRB unified directive that the associate then reviews against the source.
2. BStanding in the High Court and responding to questions from the bench about a written submission.
3. CDeciding, after meeting the client, whether to recommend settlement or to litigate the matter to judgement.
4. DVerifying that each Supreme Court decision cited in a written submission actually exists and stands for the proposition claimed.

Check answer→

Quick check

—

Which of the following best captures what does NOT change when a Nepali legal practice introduces AI into its workflow?

1. AThe time spent on first-draft contract review.
2. BThe time spent on routine letter drafting.
3. CThe time spent translating English clauses into plain Nepali for client briefings.
4. DThe time spent on oral advocacy in court and on verifying that every cited authority actually exists.

Check answer→

What comes next

The workflow map shows where AI compresses time and where it does not. It also shows, by implication, where the danger sits: the stages where the model produces confident, well-formatted text that someone has to verify against an authority the model has not actually read. The next section looks directly at that problem — hallucination — and at the duty of competence that places the consequence of any hallucinated authority squarely on the lawyer, not the tool.