What AI does well for accountants — and what it does not

The honest taxonomy

Strip out the hype and current generative AI — the GPT-class chatbots, the Claude-style assistants, the embedded copilots inside Tally, Zoho, and Excel — is reliably useful for a narrower band of work than the brochures suggest. The band is still very wide, and it covers a lot of what a Nepali bookkeeper or audit junior actually does on a Tuesday morning. But it has edges, and walking past those edges is how firms get into trouble.

The clearest way to think about it is to sort tasks by one question: how cheaply can you check the output? If verification is fast and the cost of a quiet error is small, AI is a force-multiplier. If verification is slow and the cost of a quiet error is large — penalties, restated accounts, ICAN scrutiny, a client losing a tender — then AI is a liability dressed as a shortcut.

What AI does well

Six categories, in roughly descending order of how confident you can be.

1. Drafting prose from your own figures. Hand the model a clean trial balance and a few notes about the business, ask it for a management commentary or a board summary, and it will produce a competent first draft in under a minute. The numbers come from you; the model only arranges words around them. A senior reviewer can fix tone in five minutes, where writing from scratch would have taken forty.

2. Summarising long regulatory documents. IRD circulars run to thirty pages and repeat themselves. NRB monetary policy statements are dense. ICAN technical pronouncements assume you have read the previous twelve. A model can compress these into a structured brief — “what is new, what changed, what stays the same, what you must do by when.” You still read the source for anything you act on. But the model gets you to the relevant paragraph in two minutes instead of forty.

3. Translating accounting language into Nepali for SME clients. A garments wholesaler in Birgunj does not want to hear about “provisions for doubtful debts.” She wants to hear, in plain Nepali, that some customers may not pay and we are setting aside money in case. The model handles this translation well, especially if you give it the English source and ask for client-facing Nepali. Read it before you send it — Nepali nuance in the model’s output is uneven — but the productivity gain is real.

4. First-draft journal narrations and working-paper notes. “Being purchase of office equipment from XYZ Suppliers, VAT extracted, capitalised under furniture and fixtures” is not a sentence anyone enjoys writing forty times a day. Give the model the transaction details, ask for narrations in your firm’s house style, and it will produce them faster than you can type. Review for accuracy of account heads and VAT treatment.

5. Extracting structured data from invoices and bank statements. Combine an OCR step with a language model and you can pull date, vendor, PAN, sub-total, VAT amount, and gross from a stack of scanned invoices into a spreadsheet. Accuracy is high on clean prints, poor on faded thermal receipts, and middling on handwritten ones. The new workflow is: machine extracts, human spot-checks the totals against the source. The bookkeeper’s hour becomes fifteen minutes.

6. Candidate explanations for variances. Your gross margin dropped four points this quarter. Hand the model the numbers and the operational context — “we changed suppliers in Falgun, electricity tariff rose, we ran a discount campaign” — and it will list plausible drivers, ranked, with the arithmetic to test each one. You still do the testing. But the hypothesis generation, which is the slow part of variance analysis, takes minutes.

What AI does badly, or unsafely

The same technology fails — and fails most dangerously when it looks like it is succeeding — on a different cluster of tasks.

1. Arithmetic over long lists. A language model is not a calculator. It can perform short arithmetic in its head, often correctly, but it has no internal ledger. Ask it to sum a 200-line expense schedule and it will produce a number that is roughly right and occasionally catastrophically wrong, with no warning either way. Always re-compute totals in the spreadsheet or in the accounting software. Use the model to explain the totals, never to produce them.

2. Citing specific sections of the Income Tax Act 2058. The model has read about the Income Tax Act. It has not necessarily read the current Finance Act amendments. It will confidently quote a section number that does not exist, or attribute a rate to the wrong sub-section. For any client advice that turns on a specific provision — section 88 withholding rates, section 57 depreciation pools, schedule-1 personal slabs — you verify against the IRD website or the bare act. Treat any citation the model gives you as a lead, not a source.

3. Borderline tax positions and judgement calls. Whether a payment to a related party is at arm’s length, whether a particular receipt is capital or revenue, whether a TDS exemption applies — these are the questions a client pays a chartered accountant to answer. The model can list the considerations. It cannot weigh them in your client’s facts, and it does not carry the professional responsibility if the weighing is wrong.

4. Computing final tax liability on anything non-trivial. A salaried individual’s annual tax is straightforward enough that even there a spreadsheet is more trustworthy. A trading firm with depreciable assets, carried-forward losses, withholding credits, and a partial VAT refund claim is not. Use the model to draft the structure of the computation; do the computation itself in software designed for it.

5. Signing off on audit conclusions. Auditing is not document production; it is the formation and recording of professional opinion. The model can help draft a management letter once you have decided what the opinion is. It cannot form the opinion. ICAN’s standards, and your name on the report, require that a human did.

The verification test, restated

Put the two lists together and one rule falls out. Before you use AI on any accounting task, ask: if the output is wrong in a way I do not immediately spot, what does it cost — and how long would it take me to catch the error from the source? If the cost is small and the catch is fast, use the model freely. If the cost is large and the catch is slow, the model is the wrong tool, no matter how confident it sounds.

This is not a counsel of timidity. Most accounting work — drafting, summarising, translating, extracting, narrating, hypothesising — clears the test. A surprising amount of the day, by hours, is exactly the work AI is good at. The professional act is knowing which hour is which.

Check your understanding

Quick check

—

Which of the following is the most reliably strong use of current generative AI in a Nepali accounting practice?

1. AComputing the final tax payable for a VAT-registered trading firm with carried-forward losses.
2. BProducing a first-draft management commentary from a trial balance you have already prepared.
3. CCiting the exact section and sub-section of the Income Tax Act 2058 that governs a non-resident withholding question.
4. DSumming a 200-line expense schedule and producing the total.

Check answer→

Quick check

—

What is the single best question to ask before deciding whether to use AI on a specific accounting task?

1. AIs the task interesting enough to justify learning a new tool?
2. BDoes my firm have a paid subscription, or only the free version?
3. CIf the output is silently wrong, how costly is the error and how quickly can I catch it against a trusted source?
4. DHas anyone else in the office tried this with AI before?

Check answer→

What comes next

Knowing what AI does well in the abstract is half the picture. The other half is knowing where in your actual month-to-month workflow those strengths land — which stages of bookkeeping, close, reporting, and filing change, and which do not. The next section walks through a typical small-firm month and marks the map.