Accounting SOP — Bottleneck Map & AI Solutions

Document Collection & Data Entry

Invoices, statements, source documents → verify → enter into the system → classify → clean up

~40% of time15 problemsBottleneck #1

▸

1.1 Receive source documents from 5–10 channels (email, paper, scans, photos, e-invoices)

1.2 Validate source documents bottleneck

Problem

How AI solves it

Result

Problem 1

Manual checks: is the tax code correct, is the date valid, are the signatures complete, does the amount match the contract

🟡 Hybrid

AI cross-check: tax code → business-registration lookup, invoice date → contract deadline, amount → PO

Result

Flag anomalous invoices before posting to the ledger

Problem 2

VN: invoices >VND 5M must be paid by bank transfer to be CIT-deductible — very easy to miss

🟢 Plain code

Check amount >5M → flag if no accompanying bank-transfer document

Result

Auto-flag VN compliance breaches.

$0, <1 sec

Problem 3

Duplicate invoices: same vendor + amount + date → double-posted. >60% of invoice errors come from manual entry

🟢 Plain code

Detect duplicates: same vendor + amount ± date range

Result

Prevents duplicate payments.

$0, <1 sec

1.3 Manual data entry — keying in hundreds of transactions by hand worst bottleneck

Problem

How AI solves it

Result

Slow manual entry

Keying hundreds→thousands of transactions/month by hand. Fatigue causes errors.

🔴 AI (OCR+NLP)

PDF/image → Excel automatically. Claude Vision reads any layout.

10x faster

PJICO: 70% less time, 95% fewer errors

Transposition

Typing $91 instead of $19 — the most common error (Patriot Software)

🟡 Hybrid

AI checks entered figures vs OCR readings, flags when they differ >10%

Catches transposition errors

Reversal

Posting a debit as a credit — the ledger still balances, so it's VERY hard to catch

🟡 Hybrid

AI checks transaction type → expected debit/credit → flags if reversed

Catches errors a trial-balance check misses

Omission

Forgetting to post one transaction → the ledger still balances but is understated

🟢 Plain code

Compare the transaction count in the bank statement vs entries posted → flag missing

$0. Catches omitted transactions

1.4 Classify transactions into accounts / expense categories bottleneck

Problem

How AI solves it

Result

Misclassification

Wrong account/category → distorts expense reports and tax

🟡 Hybrid

AI classifies by vendor name + historical pattern. Sends only the vendor text, never the file.

~90% accuracy, 3 min

Inconsistency

Jan: "Office Supply", Feb: "Office supplies" — same category, different names

🟢 Plain code

Normalize category names, merge similar ones (fuzzy text match)

100% consistent. $0

New vendor

Never seen before → no pattern to learn from

🟡 Hybrid

AI suggests a category from the name + description, user confirms in 1 click

Cuts lookup time

1.5 Clean & standardize data time-consuming

Problem

How AI solves it

Result

Date formats

DD/MM vs MM/DD vs YYYY-MM-DD — a different style in every file

🟢 Plain code

Auto-detect the format → standardize

100% accurate. $0

Numbers stored as text

CSV import → numbers become text, hidden whitespace, stray characters

🟢 Plain code

Trim, convert type, remove special chars

100K rows: 8h → 12 min

Data silos

Data scattered across 5–10 systems (ERP, CRM, bank, each department's own Excel)

🟡 Hybrid

Code merges files + AI resolves conflicts when the same transaction shows different figures

One source of truth

1.6 Archive & organize the original source documents

1.7 VN: e-invoice formats differ across MISA, Viettel, FPT, VNPT VN-specific

Problem

How AI solves it

Result

VN-specific

XML/PDF e-invoice formats differ across service providers

🟡 Hybrid

AI reads multiple VN e-invoice formats → standardizes into one unified table

Batch-processes VN invoices

Bottleneck levelVery high — 15 problems

↓

Reconciliation & Verification

Ledger ↔ Bank ↔ AR/AP ↔ Intercompany ↔ Anomaly detection

~20% of time16 problemsBottleneck #2

▸

2.1 Bank Reconciliation heavy bottleneck

Problem

How AI solves it

Result

Manual matching

Thousands of transactions, 4–8 hours per run

🟢 Plain code

Fuzzy match: amount ± tolerance + date ± N days

87% matched, <1 sec, $0

Multi-bank formats

VCB, Techcombank, MB, ACB — each bank has a different format

🟡 Hybrid

AI detects columns, or a preset template per VN bank

Standardizes multi-bank data

Timing differences

Bank posts on T+1, the ledger on T+0 → a "false" discrepancy

🟢 Plain code

Match with a date tolerance of ±3 days, label as "timing difference"

Distinguishes real gaps from timing

Outstanding items

Checks/deposits in transit — must be tracked from one month to the next

🟢 Plain code

Track unmatched items from last month → auto carry forward

Nothing slips through

Explaining discrepancies

"It takes more time to explain than to fix" — Ledge.co

🟡 Hybrid

AI writes the explanation: "Likely timing" / "Missing from bank" / "Duplicate"

Automatic commentary

2.2 AR/AP Reconciliation & Aging time-consuming

Problem

How AI solves it

Result

Manual aging

Hundreds of customers, bucketed 0–30/31–60/61–90/>90 days

🟢 Plain code

DATEDIF → bucket → summary

$0, <1 sec, 100% accurate

Wrong provision

Miscalculating the TT48/VAS provision rate across buckets

🟢 Plain code

Apply the standard or custom rate → roll up

Compliant. $0

Late detection

You only learn a customer is a bad debt long after it's overdue

🟡 Hybrid

AI analyzes payment patterns → predicts which customers are at risk of paying late

Early warning before overdue

3-way AP matching

Match invoice + PO + receipt — time-consuming

🟢 Plain code

Code matches the 3 sources → flags discrepancies

Cuts AP processing 75%

2.3 Intercompany Reconciliation bottleneck

Problem

How AI solves it

Result

Amount/date mismatch

Entities record the same transaction with different figures

🟢 Plain code

Match cross-entity → flag differences

Catches intercompany gaps

Complex allocations

% of revenue vs headcount vs direct charge

🟡 Hybrid

Code computes → AI writes the methodology notes

4 entities, 3 currencies, automated

FX translation

Multi-currency → reporting currency

🟡 Hybrid

Code computes the rate → AI explains the FX gain/loss

No more manual FX math

Elimination entries

Intercompany elimination journal entries on consolidation — very error-prone

🟢 Plain code

Auto-generate elimination JEs from matched intercompany items

Fewer consolidation errors

2.4 Anomaly & fraud detection important

Problem

How AI solves it

Result

Outliers

Transactions >3 std dev from the mean

🟢 Code

Compute statistics → flag

Scans everything, not just a sample

Duplicate payment

Same vendor + amount + ±date

🟢 Code

Deduplication check

Prevents lost money

Round numbers

Exactly $10,000 — a fraud signal (Benford's Law)

🟢 Code

Detect round amounts > threshold

Benford analysis. $0

After hours

Weekend/late-night transactions — unusual for an SME

🟢 Code

Check the timestamp

Auto-flag

Unknown vendor

New vendor + large amount → possible shell company

🟡 Hybrid

AI cross-checks the database, flags new vendors with large amounts

Vendor-risk alerts

Bottleneck levelHigh — 16 problems

↓

Bookkeeping & Posting

Daily journal entries, depreciation, allocations, closing transfers, accruals

~15% of time6 problems

▸

3.1 Record daily journal entries

3.2 Wrong accounts, depreciation, allocations, closing transfers error-prone

Problem

How AI solves it

Result

Wrong debit/credit account

The ledger balances but the entry is wrong

🟡 Hybrid

AI suggests the account based on transaction type + history

Fewer misposted accounts

Recurring JEs

Monthly depreciation/allocation still keyed by hand

🟢 Plain code

Auto-generate recurring JEs from a template

100% automated. $0

Missing accruals

Expense incurred but no invoice yet → forgotten

🟡 Hybrid

AI compares this period's expenses vs the prior period, flags an unusual drop

Catches missed accruals

Late invoices

Invoices arrive after close → the books must be reopened. "#1 post-close disruption" — Numeric

🟡 Hybrid

AI tracks outstanding POs with no invoice → reminder before close

Fewer reopened books

3.3 Allocate costs across multiple departments / entities

3.4 Controller review & approval

Bottleneck levelMedium — 6 problems

↓

Financial Statements & Period-End Close

P&L, BS, CF, workpapers, month-end close, spreadsheet debugging

~15% of time8 problems

▸

4.1 Consolidate figures from multiple sources

4.2 Build the 3 linked financial statements hundreds of formulas

Problem

How AI solves it

Result

Error chain

P&L↔BS↔CF are linked — one error breaks the whole thing

🟢 Plain code

Auto-check: does P&L Net Income tie to the BS? Does the CF reconcile?

Instant balance check. $0

Footing errors

Subtotals ≠ the total line

🟢 Plain code

Auto-sum and compare → flag

100% accurate. $0

Non-standard workpapers

Every accountant does it their own way

🟡 Hybrid

AI builds a standard template: GL summary, detail, TB, adjustments

406 formulas / 1 prompt (Wharton)

4.3 Debug a complex Excel file — inherited, nobody understands it takes 1–2 days

Problem

How AI solves it

Result

Spreadsheet archaeology

Old file, the author has left, a tangled dependency chain

🔴 AI-heavy

"Walk me through it" → AI maps dependencies, explains the logic

2 days → 20 min

Post-close adjustments

Errors found after close → the books are reopened

🟡 Hybrid

AI scans the GL before close: misclassifications, missing entries

Fewer post-close entries

4.4 A drawn-out month-end close — 50% of teams >5 days (Ledge.co) deadline pressure

Problem

How AI solves it

Result

5–10 day close

Consolidating many sources, multi-step, under deadline pressure

🟡 Hybrid

AI builds a close checklist + tracking + flags items not yet done

Rently: 8 → 4 days

No visibility

Who's finished which step? Which step is blocked?

🟡 Hybrid

Real-time status dashboard for the controller

Fewer waiting-on bottlenecks

4.5 VN: Circular 99/2025 (effective 1 Jan 2026) changes many rules → templates need updating

Bottleneck levelMedium–High — 8 problems

↓

Analysis & Management Reporting

Variance, KPI dashboards, cash-flow forecasting, advisory for the CEO/CFO

~5% (often skipped)5 problems

▸

5.1 Variance Analysis + writing commentary often skipped

Problem

How AI solves it

Result

Out of time

Accountants are drained by Phases 1–4, with no energy left to analyze. Hand-writing commentary takes hours.

🟡 Hybrid

Code computes variances + color-codes them → AI writes a ~200-word explanatory narrative + a top-5 summary

Hours → a few minutes (Rillet)

5.2 KPI dashboard — the CEO wants one, but the accountant can't build charts

Problem

How AI solves it

Result

Don't know how

Building an Excel dashboard (charts, slicers, KPIs) is beyond many accountants

🟡 Hybrid

Code builds the charts + KPIs → AI writes the insights

30 min. Saved $400K (Jan Haugo)

5.3 Cash-flow forecasting — manual Excel, gut-feel, large errors

5.4 Nobody reads the financial statements → no insights for business decisions

5.5 Accountants want to move from "bookkeeping" → "advisory" but have no time

Bottleneck levelLow — but often skipped entirely

↓

Tax & Compliance

VAT, personal income tax, social insurance, finalization, audit, VN regulations

~5% (spikes at year-end)7 problems⚠ AI gets 2/3 of tax questions wrong

▸

6.1 Payroll, social insurance, personal income tax complex formulas

Problem

How AI solves it

Result

Progressive PIT

7 brackets + personal deductions of VND 11M + 4.4M/dependent. Very complex for 50+ employees.

🟢 Plain code

Fixed formula, computes gross→net automatically

80% less time. $0

SI/HI/UI

Employee vs employer rates, contribution caps change every year

🟢 Plain code

Apply the standard rates, update the cap annually

100% accurate. $0

6.2 Prepare & file monthly/quarterly VAT and PIT returns

6.3 VN regulations change constantly — new VAT Law 7/2025, Circular 99 1/2026, Decree 70 e-invoice VN-specific

Problem

How AI solves it

Result

Constant changes

VAT Law 48/2024, Circular 99/2025, Decree 70/2025. Reading the primary texts is very time-consuming.

🔴 AI-heavy

AI looks it up → summarizes the impact → proposes adjustments. ⚠ ALWAYS cross-check the primary text!

Faster than reading the source text

6.4 Expenses >VND 5M must be paid by bank transfer — easy to breach CIT rules VN-specific

Problem

How AI solves it

Result

CIT breach

Expenses >5M paid in cash → not deductible. Penalty of 20% tax + 0.03%/day interest.

🟢 Plain code

Check amount >5M + payment = cash? → flag before the audit

Auto-flag breaches. $0

6.5 Year-end finalization + preparing the audit file

6.6 VN: documents must be printed, signed, and stamped (hard copy) → ❌ AI can't solve this (legal requirement)

Bottleneck levelMedium (high at year-end) — 7 problems

Summary: 52 problems — how many can AI solve?

Total problems
across 6 phases

42%

🟢 Plain code
$0, <1 sec

42%

🟡 Hybrid
~$0.02–0.08/run

14%

🔴 AI-heavy
PDF, debugging, standards

60% of accounting time is stuck in Phases 1–2 (collection + reconciliation). 84% of problems can be solved with plain code or a hybrid approach — no heavy AI needed. 1 problem is unsolvable (VN mandates hard copies).

Monthly Accounting SOP — A Map of 52 Bottlenecks

Summary: 52 problems — how many can AI solve?