IRD and NRB filings — reading rules and drafting submissions

What the text actually looks like

Working with Nepali regulators is a reading job before it is a drafting job. IRD circulars are dense, repetitive, and often refer back to prior circulars from years you have to look up. NRB directives — the Unified Directives in particular — are organised by sector, revised annually, and not always available in English. Inspection reports and query letters come in their own house style, sometimes in Nepali, sometimes in a mix.

For a partner who has read this material for twenty years, the structure is familiar. For a junior who is encountering a directive for the first time at 8 p.m., it is forty minutes of figuring out what the document is even asking before any thinking can begin. That gap is where AI earns its keep — and where it most often misleads.

Where AI genuinely helps

There are three workflows where AI is reliably useful with regulator text, and the common feature is that the original document remains in the loop.

1. First-read summaries. Drop a circular into a model and ask for a structured brief: what is the document, what is new compared with the prior position, who must do what, by when, and what are the explicit exceptions. The output is a scaffolding that lets you decide which sections of the original you need to read carefully and which you can scan. The summary is not the rule; it is a map of the rule.

2. Cover notes and explanatory letters. Most filings to the IRD or NRB benefit from a one-page cover note that explains what is being submitted and why. The model is excellent at this — give it the underlying numbers, the relevant rule, and the firm’s house tone, and it produces a polished draft in under a minute. The senior edits for accuracy and the partner signs.

3. Structured responses to query letters. When the IRD writes asking why a particular deduction was claimed, or NRB asks for clarification on an FX transaction, the response has a predictable shape: acknowledge the query, restate the facts, set out the legal basis, attach the evidence, conclude. The model is good at writing into that shape. Your value-add is the legal basis and the evidence — the model assembles the prose around them.

The non-negotiable rule

There is one rule for all three workflows and it does not bend.

The failure mode is specific and worth recognising. A model summarising a forty-page directive will produce a clean six-paragraph brief. Most of the brief will be correct. One paragraph may quietly omit the carve-out that exempts the very transaction you were asking about, or invert a “shall not” into a “may.” Because the rest of the summary is accurate, the error is camouflaged. The only way to catch it is to read the original — and the only person who can catch it is someone who knows what to look for.

Concrete examples from practice

An NRB FX reporting cover letter. A trading client has remitted advance payment for an import and needs to file the supporting documentation under the Foreign Exchange Regulation Act and the relevant NRB directive. The underlying facts are simple; the cover letter has to map them precisely to the directive’s documentary requirements. The model produces a competent draft from the facts and the directive text. The senior verifies that every requirement listed in the directive is addressed in the letter — including the one that the model may have skipped because it appeared as a sub-clause two pages later.

Summarising an IRD inspection report. The report runs eighteen pages, half in Nepali, with line-item findings on disallowed expenses, TDS shortfalls, and a VAT timing issue. The model produces a structured summary by finding, with the IRD’s reasoning beside each one. The partner uses the summary to decide which findings to contest and which to accept, then reads the original passages for the contested ones in full before responding.

A written submission in a tax dispute. The client has been assessed for additional VAT on a supply the firm believes is exempt. The submission needs to set out the facts, the relevant provisions of the VAT Act 2052 and any IRD circulars, the case for exemption, and the precedent. The model drafts the structure; the senior writes the legal argument; the partner reviews and signs. The model is not authoring the legal position — it is composing the prose around a position the firm has taken.

Bilingual drafting

A practical reality of NRB and IRD work is that many filings, especially the formal ones, are in Nepali. The Nepali on these documents is a particular register — formal, technical, sometimes archaic — that is hard for a junior to produce well from scratch.

AI is genuinely useful for a polished bilingual draft. Write the English version with the senior’s input, then ask the model to produce a parallel Nepali version in the appropriate register. The output is almost always usable as a starting point and saves an hour of fumbling with vocabulary. But two cautions are non-negotiable.

First, a native review is essential. The model will get the broad shape right and miss the specific term. कर निर्धारण and कर निर्धारण आदेश are not interchangeable, and the model does not always know which one belongs in a given sentence. Someone fluent in formal Nepali — and ideally familiar with the regulator’s preferred terms — reads the draft before it goes out.

Second, the legal weight is in the Nepali version, where the filing is in Nepali. The English you drafted in your head is a working translation. If there is a discrepancy, it is the Nepali wording that the regulator will read and act on. Review accordingly.

Check your understanding

Quick check

—

An AI model produces a clean six-paragraph summary of a thirty-page NRB directive. You need to decide whether a specific client transaction is permitted under the directive. What is the correct way to use this summary?

1. ATreat the summary as the firm’s legal position and proceed with the transaction, since the model has condensed the relevant rules.
2. BUse the summary as a map of the directive — to decide which sections to read carefully — and then read the original directive in full for the sections that govern your client’s transaction before forming a position.
3. CAsk a second model for a competing summary and act on whichever one is shorter.
4. DForward the summary to the client as the firm’s advice, since it captures the essence of the directive.

Check answer→

What comes next

Reading and drafting are most of the visible work, but the part that determines whether the work survives an ICAN review or an IRD reassessment is the audit trail. The next section sets out a short, opinionated working standard for documenting AI use in tax and compliance work — what to keep, how to keep it, and the small-firm policy that makes it actually happen.