The accounting workflow, mapped to where AI helps

The month, before AI

Before any tooling, a month at Bagmati Traders looks roughly like this. Days 1-20: the bookkeeper enters sales invoices, purchase bills, bank transactions, and petty cash into Tally as documents arrive — most of it by typing from paper or PDF. Days 21-25: month-end close. Bank reconciliations, supplier and customer ledger tie-outs, inventory adjustment for goods in transit, accrual of unpaid utilities and salaries. Days 26-28: VAT return preparation, sales and purchase register cross-checks, filing on the IRD portal by the 25th of the next Nepali month. Day 29-30: management reports — a P&L, a cash position, a simple aging — and the partner’s review meeting with the owner.

Hours, roughly: 60-80 for data entry, 15-20 for close, 8-12 for VAT, 4-6 for reports and the review. The bookkeeper does the first three; the partner does the fourth and signs off everything. Nothing here is unusual — most VAT-registered SMEs in Kathmandu Valley run on something close to this rhythm.

Stage by stage, with AI in the loop

Walk through the same month now, but assume the bookkeeper has a decent paid chatbot, an OCR-plus-extraction tool for invoices, and a copilot inside the spreadsheet she already uses.

1. Source document intake — high leverage. This is where the biggest hour savings live. Scanned and PDF invoices flow through an extraction tool that pulls vendor, PAN, date, sub-total, VAT, and gross into a staging spreadsheet. The bookkeeper spot-checks rows where confidence is low and corrects them. Clean prints — vendor invoices on letterhead — run at ninety-percent-plus accuracy. Faded thermal receipts and handwritten bills still need manual entry. Net effect: data entry collapses from sixty hours to fifteen or twenty, with the saved hours going to spot-checking and exception handling, not to going home early.

2. Bookkeeping and posting — medium leverage. Once data is staged, the model can suggest the correct account head and a draft narration for each line. Most of the time it is right; the bookkeeper accepts or corrects. This is faster than typing narrations from scratch, but slower than people imagine — every line still needs a human eye, and you cannot batch-accept blindly without inviting silent misclassifications. Expect maybe a third of the posting time saved, not all of it.

3. Month-end close — low to medium leverage. Bank reconciliations are mostly arithmetic and matching, which is exactly what AI is bad at and Excel is good at. The model helps with the narrative parts of close — drafting accrual memos, summarising why a particular ledger swung — and with surfacing unusual entries for the bookkeeper to investigate. The arithmetic of the close itself is unchanged. Plan for the same fifteen hours, but with the partner reviewer getting a cleaner package on day twenty-five.

4. VAT return preparation — low leverage, high risk. The temptation is to ask the model to “prepare the VAT return.” Resist it. VAT computation is structured, rule-driven, and unforgiving — exactly the kind of work that belongs in Tally, in the IRD portal’s built-in checks, and in a spreadsheet you trust. What AI does help with is explanatory: writing the cover note, drafting the explanation for why input credit on a particular invoice was deferred, summarising any departmental correspondence. The return itself remains a human-and-software job. Hours saved: minimal. Hours risked, if you let the model produce the numbers: substantial.

5. Management reports and commentary — high leverage. The partner’s monthly report to the owner is largely prose around a few core figures. Hand the model the P&L, the cash position, the aging, and a few notes on the month’s operations, and ask for a one-page commentary in the firm’s voice. First draft in ninety seconds, partner editing time perhaps ten minutes, total time down from three or four hours to under one. This is the single most visible win in a typical small-firm month.

6. Client communications — medium to high leverage. The chasing letters to overdue customers, the polite reminders to suppliers, the short notes to clients explaining a quarter’s results — all of these are pattern-based prose where the model does well. Especially valuable for switching register between English internal notes and Nepali client-facing letters. Save real time, but always read before sending; the model occasionally hallucinates a relationship detail or a wrong amount.

7. Audit preparation and confirmation letters — medium leverage. When statutory audit begins, much of the early work is producing schedules, drafting confirmation letters to banks and major debtors, and summarising loan agreements and lease contracts. The model accelerates all of the drafting and most of the summarisation. It does not perform the audit; it does not form the opinion; it does not test controls. Treat it as a faster typist for the junior, supervised by the senior.

What the new month looks like

Add it up across Bagmati Traders. The bookkeeper’s data entry shrinks from a dominant chore to a managed exception queue. The partner’s report-writing afternoon shrinks to a morning. The VAT cycle is essentially unchanged — and that is correct. The close is moderately faster, mostly because the close package arrives at the partner cleaner, not because the arithmetic moves faster. Audit prep, when its season comes, runs perhaps thirty percent faster.

The bookkeeper has not lost her job; her job has changed shape. She spends less time typing and more time judging — which lines the extractor got wrong, which posting suggestion is dubious, which exception needs partner input. That is a more skilled job than before, not a less skilled one.

What does not change

This is the part most articles about AI in accounting skip, and it matters most.

The partner’s final review does not change. The signature on the audit report does not change. The professional responsibility for the VAT return filed in the firm’s name does not change. The duty to retain workpapers that let someone reconstruct what was done, by whom, and on what basis does not change. ICAN’s standards do not contain an AI carve-out; they continue to assume a human accountant made the call.

This is the right frame for everything that follows in this course. When we talk in later chapters about specific tools for bookkeeping, audit, tax, or client work, we are talking about reshaping where the hours go. We are never talking about removing the human from the loop. In Nepali professional practice the human is the loop; AI is the loom.

Check your understanding

Quick check

—

In a typical Nepali small-firm month, which stage gives AI the highest leverage in saved hours?

1. AVAT return computation and filing on the IRD portal.
2. BBank reconciliation arithmetic during month-end close.
3. CSource document intake — extracting fields from scanned invoices and bills into a staging spreadsheet.
4. DFinal partner review of the monthly accounts before sign-off.

Check answer→

Quick check

—

When AI enters a small Nepali accounting firm's workflow, what does NOT change?

1. AThe time spent on source-document data entry.
2. BThe professional responsibility for filings and the partner’s final review and sign-off.
3. CThe speed of drafting management commentary.
4. DThe hours spent typing journal narrations.

Check answer→

What comes next

You now have a sense of where AI helps and where it does not, both in the abstract and across a real firm’s month. The next question is operational. Before you actually introduce any of this into the office, you need a way to think about cost, risk, and the audit-trail problem — what gets documented when a chatbot is in the loop. That is the next section.