Chapter 04 · Section I · 17 min read
VAT returns and TDS — what AI can and cannot help with
A working line between what an AI assistant can safely do inside a Nepali VAT or TDS workflow, and the parts of the job that must never leave a human hand.
Tax work in Nepal is unforgiving in a particular way. The IRD does not care whether your TDS deduction was off by one percentage point because the rate changed in last year’s Finance Act or because a chatbot invented a citation; the assessment, the interest, and the penalty land the same. So the question for an accountant using AI in VAT and TDS work is not “is this clever?” — it is “if this is wrong, who pays, and how fast can I catch it before they do?” Everything in this section is built around that question.
Where AI genuinely helps in a VAT return
Nepal’s VAT regime — 13% standard rate as of 2026, with zero-rated exports, a list of exempt supplies under Schedule 1 of the VAT Act 2052, and the partially-exempt input-credit mess that anyone with mixed supplies has lived through — is procedural work. Most of the month-end VAT exercise is reconciliation, not interpretation. That is precisely the shape of task AI handles well.
A good first use is a one-page reconciliation memo that ties your sales register and purchase register to the figures going on the D-03 / D-04 return. Hand the model the totals from your accounting system, the totals from your VAT working file, and the differences. Ask for a structured narrative: what reconciles cleanly, where the timing differences sit, which invoices crossed the period, which credit notes are still in dispute. The numbers are yours; the model arranges the explanation. A senior who used to spend forty-five minutes writing this from scratch can review it in eight.
A second good use is summarising treatment — a paragraph for each unusual line on the return. Zero-rated exports with which destination, exempt supplies under which schedule entry, reverse-charge on imported services, partial input credit ratio for the month. The model writes the prose; you supply the rule and check the citation. Done well, this memo becomes the workpaper that survives the next IRD inspection without you having to reconstruct what you were thinking eleven months ago.
Where AI must not be allowed near a VAT return
There are three places where you have to draw a hard line, because the cost of an error is asymmetric.
1. Filling the IRD portal directly. No agent, no copilot, no “I’ll just have it click through D-03 for me.” The portal is the official act of filing. A human signs that act. If the model misreads a field — sales versus taxable sales, total versus net of credit notes — the return is wrong from the moment it is submitted, and the correction process is painful. Use AI to prepare the values; type or paste them yourself.
2. Deciding borderline supplies. Is a particular consulting fee an exempt financial service or a taxable professional service? Is a training delivered partly online and partly in Kathmandu a single supply or two? These are interpretive calls that depend on the contract, the substance, and sometimes a circular the model has never seen. A model will give you an answer in either direction with equal confidence. Use it to lay out the arguments; do not let it pick the side.
3. Computing the final amount without a re-check. The model can add, but it can also drift — silently picking up a number from the wrong cell, treating a credit as a debit, including an invoice twice because you pasted it twice. The final VAT payable that goes on the return is checked by a human against the source workings every single time. No exceptions for “this client is small” or “the model has done it right for six months.”
TDS — where the hallucination risk is highest
TDS is where I have seen the most serious AI mistakes in Nepali practice, and the pattern is always the same: the model produces a confident, well-structured answer that cites a specific section of the Income Tax Act 2058 — and the section either does not say what the model claims, or does not exist at all.
This is not a quirk; it is how language models behave. Asked for a specific legal citation in a small jurisdiction, the model interpolates from patterns it has seen in larger ones. The output looks correct because it is structured correctly. The citation is wrong because the model has never actually read the underlying section in a way you can rely on.
The safe pattern for a TDS question — say, the rate on a payment to a non-resident consultant — is to use the model for a structured first cut, then verify everything before quoting. A useful prompt skeleton: “I am paying a non-resident professional in [country] for [service] performed [where]. Lay out, as a structured analysis: (a) the likely section of the Income Tax Act 2058, (b) whether a DTA with [country] applies and what its relevant article would be, (c) the default TDS rate and any exemption, (d) what evidence I need on file, (e) what could change this answer. Cite the source for each point.”
You then take that structured cut and verify each citation against the actual Act, the actual DTA, and the current year’s Finance Act. The model has done the scaffolding; you have done the law. Quote nothing the model produced until you have seen it in the source.
The “tax rates change” reality
The single most common AI failure in Nepali tax work is rate drift. Models are trained on text from a particular cut-off, and rates in Nepal are revised every year by the Finance Act — sometimes materially, sometimes only in scope or exemption. A model that confidently states “the TDS rate on rent in Nepal is 10%” is making a statement about a snapshot of the world that may already be a year out of date by the time you read it.
Treat any specific rate the model produces as a hypothesis until you have checked it against the current Finance Act and any IRD notification issued since. This is true even for rates that have been stable for years; the year you stop checking is the year they change. Build the habit into the prompt itself — ask the model to flag every rate it states with a note that it must be verified against the current year — and into the workpaper, where every quoted rate has a date and a source next to it.
A concrete prompt pattern for both
The pattern that works for VAT and TDS is the same, and it is worth memorising.
1. Give the model your actual data — the invoice, the contract clause, the totals — not a hypothetical. 2. State the rule you are applying and the year. (“VAT Act 2052, reverse charge on imported services, FY 2082/83.”) 3. Ask for a structured analysis with named citations, not a conclusion. 4. Ask explicitly what could change the answer. 5. Verify every citation in the source before acting.
Used this way, the model compresses the time you spend organising your thinking. It does not compress the time you spend reading the law, because that time is the work.
Check your understanding
Quick check
—Which of these is the safest way to use AI inside a monthly VAT return workflow in Nepal?
Quick check
—A client is paying a non-resident consultant in Singapore for advisory work delivered remotely. You ask a chatbot for the applicable TDS rate, and it cites a specific section of the Income Tax Act 2058 along with a rate. What is the right next step?
What comes next
VAT and TDS are the most procedural piece of the compliance year, which is why AI helps so visibly there. The next section moves to a harder kind of text — IRD circulars and NRB directives, where the rules themselves have to be read carefully — and looks at how AI helps you read, summarise, and draft submissions without quietly substituting its summary for the rule itself.