Translating accounting jargon to Nepali for SME owners · AI for Accountants

A surprisingly large share of an accountant’s billable hours in Nepal is spent doing live translation. The Income Tax Act is in English. NFRS for SMEs is in English. The bank’s compliance form is in English. The client — the actual person paying you, often the founder of a hardware shop in Janakpur or a cooperative in Surkhet — speaks Nepali, sometimes Maithili or Tamang or Newari at home, and reads English only well enough to be politely confused by it. Most of what you are doing in those long phone calls is not advisory work in the technical sense; it is translation, and it is the bottleneck of every practice. AI changes this — but only if you understand why a faithful translation is the start of the work and not the end.

Translation alone is not enough

Suppose you take a clause like “Section 88 of the Income Tax Act allows the deductee to claim TDS credit only if the TDS certificate is filed with the income tax return for the year in which the income is taxable.” and ask any modern model to render it in Nepali. The output will be a fluent, grammatically correct Nepali paragraph — “आयकर ऐनको दफा ८८ अनुसार कट्टी गरिएको करको दाबी गर्न…” — and the client will read it, nod, and ask you the same question they asked before: “त्यसो भए मैले अहिले के गर्ने त?”

The translation was correct. It was also useless. The client did not have a language problem; they had a meaning problem. They could not extract a next action from the sentence even when the words made sense, because tax statutes are not written to be acted on — they are written to be unambiguous in court. The translation preserves the unambiguity. It preserves the uselessness too.

This is the core insight of this section. A practising Nepali accountant’s translation work is two jobs, not one. Job one: render the source faithfully in Nepali. Job two: write one plain sentence underneath that says, in the simplest possible form, what this means for this client, this week. AI is excellent at job one. It is competent at job two, but only if you ask for it explicitly — and you must always be the one who decides whether the consequence sentence is correct, because the consequence depends on facts the model does not have.

The translate-plus-consequence prompt

The prompt pattern is short and you will use it dozens of times a week:

Translate the following accounting / tax text into plain Nepali. Then, in one additional sentence, restate the practical consequence for a small business owner — the action they need to take, or the risk they face, in language a non-accountant can act on. Give me both the translation and the consequence sentence, clearly labelled. Do not add any number, date, or rule that is not present in the source.

Two things make this prompt work. First, it asks for both versions, labelled — the faithful Nepali for your records and the consequence sentence for the client. You will end up sending the consequence sentence and keeping the translation in your file. Second, the explicit ban on inventing numbers, dates, or rules guards against the model’s natural drift, which is to “helpfully” add a deadline or a section number that sounds right but is not in the source.

A worked example. You feed in: “Penalty under Section 119 is 0.05% per day on the unpaid TDS amount, calculated from the due date until the date of actual deposit.” The model returns the faithful Nepali. Then, in the consequence line, it produces something like: “यदि तपाईंको कम्पनीले समयमा TDS जम्मा गर्नुभएन भने, ढिलाइको हरेक दिनको ०.०५% जरिवाना तिर्नुपर्छ — अहिलेकै बाँकी रकममा हप्ताभरि ढिलाइले हजारौं रुपैयाँ थप्न सक्छ।” That is the sentence the client reads twice. That is what changes behaviour. The literal translation, by itself, would have changed nothing.

Vocabulary discipline: pick a word and keep it

Nepali accounting terminology is, frankly, still being negotiated. There is no settled equivalent for “accrual,” “deferred tax,” or “trial balance” that every practitioner uses. Different textbooks pick different words. The IRD’s own forms switch between Sanskrit-derived terms and direct English borrowings on the same page. Models, when left to themselves, will produce a slightly different Nepali term every time you ask, even within the same conversation. One paragraph will say विवरण, the next will say प्रतिवेदन, the third will say रिपोर्ट — and the client will reasonably think you are talking about three different documents.

The cure is a small per-client or per-firm glossary, fed into the prompt as context. Write it once and reuse it for years:

Trial balance → सन्तुलन परीक्षण
Profit and loss statement → नाफा-नोक्सान विवरण
Balance sheet → वासलात
TDS → अग्रिम कर कट्टी (TDS)
VAT input credit → VAT कर मिलान दाबी
Accrual basis → प्रोदभव आधार
Audit report → लेखापरीक्षण प्रतिवेदन

The terms themselves are negotiable; the discipline is not. Pick a Nepali term per concept and stick to it across every email, statement cover note, and conversation. Append the glossary to every translation prompt — “Use only the following Nepali terms for these concepts: …” — and the model will hold the line. The benefit compounds: clients gradually learn your vocabulary, and the cognitive load of every subsequent communication drops.

Cultural calibration: indirectness, honorifics, bad news

The model’s default Nepali register is what you might call Wikipedia Nepali — grammatically clean, lightly formal, and culturally flat. It works for a written report. It is wrong, in subtle ways, for the kinds of conversations a Nepali accountant actually has.

The clearest place this shows up is in the framing of bad news. A direct English sentence like “Your VAT return for Asar has errors that must be corrected within seven days” translates literally into a Nepali sentence that, in spoken or even written context with a senior family client, lands harshly. The owner reads it and, depending on temperament, either becomes defensive or feels lectured. The Nepali professional convention is more layered — acknowledge the situation, suggest the path, name the deadline last. “असार महिनाको VAT विवरणमा केही कुरा मिलाउन बाँकी देखिएको छ — हामीले सात दिनभित्र सच्याएर पुनः पठाउनुपर्ने हुन्छ।” Same factual content. Very different reception.

Tell the model this explicitly: “Translate for a Nepali small business owner. For sensitive content, use indirect framing — acknowledge the situation, propose the action, state the deadline last. Maintain appropriate honorifics throughout.” The model will adjust. Without this instruction, it will not. The default register is calibrated for a different audience.

Honorifics deserve their own paragraph. The model is improving but still occasionally addresses a sixty-year-old shopkeeper with the same verb forms it would use for a peer — तपाईं with informal verb endings, or worse, slipping into तिमी mid-paragraph. For older clients, repeat clients, and any client where the relationship matters, scan for honorific consistency before sending. This takes ten seconds; the cost of getting it wrong is a quiet erosion of the professional relationship that you may never hear named out loud.

Read both versions before sending

Even with a glossary, a tone instruction, and a clear prompt structure, the discipline that protects you is the same one you would apply to a junior translator’s draft: read it carefully before it leaves your machine. AI Nepali is dramatically better than it was two years ago, but the failure modes are subtle — a swapped honorific, a transliterated technical term where a real Nepali term exists, a sentence whose grammar is correct but whose register is wrong for the recipient.

The practical rule: read the English source, read the Nepali output, ask yourself whether the consequence sentence is actually true given the client’s situation (the model does not know that), and only then send. If you feel a sentence is awkward, it almost certainly is — fix it by hand rather than asking the model to “make it more natural”, which usually just produces a different kind of awkwardness. The model is a fast first-draft. You are the final editor. That division of labour is the entire game.

Check your understanding

Quick check

—

Why is a faithful, accurate Nepali translation of a tax-law clause often insufficient when communicating with a small business owner?

Nepali is too imprecise a language for tax content.
Modern AI cannot translate tax language accurately into Nepali.
A literal translation preserves the statutory register — clear in court but not action-oriented — so the client understands the words but still cannot extract a next step; the fix is to pair the translation with a one-sentence consequence statement in plain language.
Small business owners prefer to read in English regardless.

What comes next

Once you have a reliable way to draft fresh emails and translate technical content cleanly, the next leverage point is what you do with the same emails you send month after month. The next section is about identifying your recurring communications and turning them into reusable AI-assisted templates — and the specific signal that tells you a template has gone too far.