[Scenario]Finance Scenarios: AI for Bookkeeping, Reconciliation & Reports

Summary: Hand off the most time-consuming finance tasks — bookkeeping, reconciliation, reports — to AI as your assistant.

1. The Pain Points

If you work in finance, you know there are a few nights every month that feel like a trial by fire — those late evenings at month-end.

The moment you open Excel, three things pile on at once:

• Sorting invoices: a pile of VAT special invoices, e-invoices, and taxi receipts mixed in your email attachments. You open each one, check the amount, check the title, and decide whether it goes to "Administrative Expenses – Office Supplies" or "Travel Expenses – Transportation." Your hands ache and your eyes burn.
• Bank reconciliation: the Excel export from your bank and the Excel export from your accounting system don't share the same columns. The amount matches but the date is off by two days; sometimes everything is identical except the memo. Those outstanding items (amounts already booked by one side but not yet by the other because the voucher is still in transit — e.g., the bank has debited a payment the company hasn't recorded yet) force you to eyeball row after row.
• Generating reports: your boss wants the income statement, the balance sheet, and the cash flow statement (in accounting circles, these three are called the three financial statements). You pull the numbers one by one out of the trial balance, chain formulas together, and one misplaced decimal point means a full recalculation.

Across China's small and mid-sized businesses, these three tasks eat up roughly 40%–60% of finance-team working hours — but at least half of that is "grunt work" (moving numbers around, lining up columns), not real accounting judgment.

This article helps you outsource that grunt-work half to AI: AI drafts, you sign off. Notice the wording — drafts, not finals. AI cannot replace the statutory duties of an accountant (we'll hammer this point repeatedly in the Section 5 red-line section).

2. What You'll Walk Away With

After working through this article, you'll have three things in hand:

1. An "AI bookkeeping assistant" workflow: invoice photos are OCR'd (OCR = turning the text in an image into editable, structured data) into structured fields, then auto-classified and posted to 金蝶 / 用友 / 管家婆 — what used to take a full day of invoice-taping now takes one hour.
2. An "AI bank reconciliation" workflow: hand your bank-statement Excel and your company-ledger Excel to AI, and it automatically flags outstanding items with notes on the cause of each discrepancy — what used to take half a day now takes 10 minutes.
3. An "AI three-statements draft": tell AI in plain English "generate the November 2024 income statement," and it returns an editable Excel draft — building a draft from scratch drops from scratch to 5 minutes.

Prerequisites:

• Your company uses 金蝶 KIS / 云星辰, 用友 U8 / NC Cloud, or 管家婆 (any version) — all three expose APIs (think of an API as a small window a piece of software opens so other programs can talk to it) or RPA hooks (so software can click and type on your behalf), which AI can call to do real work.
• Your computer can reach 千问 / Kimi / ChatGPT — internet access is all you need.
• At least one account on 金数据 or 简道云 (for building the invoice-collection entry point).
• Your chart of accounts (the "buckets" every transaction falls into — "Cash," "Bank Deposits," "Administrative Expenses," and so on) is stable, and the business has run for at least one full accounting period.

Critical boundary (read this before any case study): finance work touches regulations such as the Accounting Law (PRC), the Invoice Management Regulations, and the Accounting Archives Management Regulations — AI can only produce "drafts / first cuts," and it cannot replace the accountant's duties of review, signing, booking, and archiving. Specifically:

• "AI-generated bookkeeping entries" ≠ bookkeeping done — AI gives you "suggested journal entries," and a human accountant must review them and post them manually (or via 金蝶 / 用友's "voucher review" workflow).
• "AI-recognized invoices" ≠ invoice-compliance verification — the authenticity of VAT special invoices, the eligibility of input-tax deductions, and the consistency of buyer name and tax ID must be verified through the tax-control system or the e-invoice service platform; AI cannot put its name on that.
• "AI-generated statements" ≠ statements filed externally — the three financial statements submitted to authorities (industry & commerce, tax, banks) must be manually reviewed, signed, and stamped by an accountant; AI drafts are only for internal discussion and cross-footing checks (verifying that the figures in different statements back each other up rather than contradict each other).

This boundary will be re-flagged every time a case study hits a compliance point. Don't use AI as your accountant. Treat it as the accountant's intern.

3. Case Studies

The three cases below are ordered progressively — Case 1 (bookkeeping) is the most-used, Case 2 (reconciliation) shows the biggest payoff, Case 3 (reporting) is the toughest on prompt engineering. Every step comes with concrete prompts and tool buttons — copy, paste, run.

Case 1: AI Bookkeeping Assistant — Auto-Classify Invoice Photos into the Ledger

Background: at a trading company, finance clerk Xiao Li handles 200+ invoices every month (VAT special invoices, e-invoices, taxi receipts, Didi e-receipts). What used to take a full day of sorting and taping should now take one hour.

Step 1: Turn Invoice Photos into Structured Data with Qwen OCR (5 minutes)

Qwen OCR is Alibaba Cloud's OCR service, with optimization tuned for Chinese receipts.

Steps:

1. Open the Qwen OCR online demo (no coding required);
2. Upload your invoice photos / scans in batch (up to 20 at a time);
3. Under "Recognition Type," pick "Receipts & Vouchers – VAT Invoice" or "General Text" (taxi / train tickets use General);
4. Click "Start Recognition" and download the result as .xlsx. Fields are pre-filled: invoice code, invoice number, issue date, buyer info, seller info, amount, tax amount, total (price + tax), remarks.

Advanced variant (for high-volume companies): write a VBA macro (a macro is a recorded script that can be replayed over and over) to batch-call Alibaba Cloud's OCR API and process 1,000+ images at once, with each recognition result landing on its own row in Excel. A code snippet is provided at the end of this section.

Step 2: Feed OCR Results + Your Chart of Accounts to AI for Categorization (10 minutes)

Open any chat box — 千问 / ChatGPT / Kimi — paste the OCR output, and attach your company's chart of accounts plus the prompt.

Prompt template (copy-paste, then replace the [ ] placeholders):

You are my bookkeeping assistant. My company is a general VAT taxpayer, and our main business is [apparel retail].
Below is an excerpt of my Chart of Accounts:

| Account Code | Account Name          |
|--------------|-----------------------|
| 6601         | Selling Expenses       |
| 6601.01      | Selling Expenses - Advertising |
| 6602         | Administrative Expenses|
| 6602.01      | Administrative Expenses - Office Supplies |
| 6602.02      | Administrative Expenses - Travel |
| 6602.02.01   | Administrative Expenses - Travel - Transportation |
| 6602.02.02   | Administrative Expenses - Travel - Lodging |
| ... (replace with your company's full chart of accounts) |

Below is this month's invoice list from OCR (one invoice per row):

| Invoice No. | Issue Date  | Seller              | Amount  | Tax   | Memo / Remarks              |
|-------------|-------------|---------------------|---------|-------|-----------------------------|
| 12345678    | 2024-11-03  | Beijing Didi Mobility | 127.00  | 7.62  | Ride on Nov 2 (6% rate example) |
| 23456789    | 2024-11-05  | Jinjiang Inn         | 484.00  | 29.04 | Lodging on Shanghai trip (6% rate example) |
| ... (convert your OCR .xlsx to this table format and paste it in) |

Please do two things for me:
1. Map each invoice to the most appropriate account ("Didi" → Administrative Expenses - Travel - Transportation; "hotel" → Administrative Expenses - Travel - Lodging; "office supplies" → Administrative Expenses - Office Supplies; "ad placement" → Selling Expenses - Advertising; and so on);
2. Fill in the memo according to the "memo standard": format "YYYY-MM-DD + seller short name + business content," e.g., "2024-11-03 Beijing Didi ride."

Output three columns: Invoice No. / Debit account (down to the leaf level) / Complete memo.
Do not output the credit account — credit is uniformly "Bank Deposits / Cash on Hand," and I'll fill that in myself.

AI returns those three columns — that is your "suggested journal entry."

Step 3: Review + Post (20 minutes)

This step cannot be skipped. As we said earlier, AI only gives a first cut:

1. Spot-check 10–20%: pick 20 invoices at random and compare them against the originals to see whether the categorization is right. Common mistakes: "Advertising" filed under "Office Supplies," or "client entertainment meals" filed under "Travel – Meals" (in reality it's "Business Entertainment," which is treated very differently for tax purposes).
2. Post the vouchers manually in 金蝶 / 用友 / 管家婆: fill in both debit and credit, splitting the amount into "tax-exclusive amount + tax amount" on two lines (special VAT invoice) or a single "price + tax total" line (general VAT invoice);
3. Voucher review: after entry, run the "Reviewer" workflow inside 金蝶 / 用友 (usually you or your supervisor), a real signature is mandatory at this step — AI cannot sign.
4. Post to the ledger: once approved, click "Post," and the voucher is formally booked into the general ledger.

OCR + VBA batch-call code snippet (optional, advanced):

vba
Sub BatchOCR_Aliyun()
    Dim filePath As String, apiKey As String
    apiKey = "Your Alibaba Cloud AccessKey"
    filePath = "C:\Invoices\"
    Dim f As String, row As Long
    row = 2
    f = Dir(filePath & "*.jpg")
    Do While f <> ""
        ' Call Alibaba Cloud OCR API and write the result into ActiveSheet
        ' (the HTTP request part uses MSXML2.XMLHTTP — omitted here for brevity)
        row = row + 1
        f = Dir()
    Loop
End Sub

Never written VBA? No problem — first run Steps 1–2 manually; automate the batch in Step 3 only after that. Get it working before you automate it is the safe path for newcomers.

Case 2: AI Bank Reconciliation — Let AI Find the Outstanding Items Across Two Sheets

Background: at a manufacturing company, finance lead Sister Wang reconciles last month's bank statement (exported from China Merchants Bank as .xlsx) with the company ledger (exported from 金蝶 as .xlsx) on the 5th of every month, flagging outstanding items. What used to take half a day now takes 10 minutes.

Step 1: Unify the Field Formats of Both Sheets (3 minutes)

The two sheets' original columns are different, and AI gets confused by inconsistent fields, so first add a few helper columns in Excel.

Bank statement sheet (China Merchants Bank export), append three columns:

Original Date	Original Memo	Original Inflow	Original Outflow	Original Balance	Unified Date	Unified Amount	Unified Memo
20241103	Goods payment	50000		150000	2024-11-03	50000.00	20241103 Goods payment

Company ledger sheet (金蝶 export), append the same three columns using the same logic.

Step 2: Feed Both Sheets + Reconciliation Rules to AI (5 minutes)

Prompt template:

You are my reconciliation assistant. Below are two tables:

[Bank Statement] (excerpt)
| Unified Date  | Unified Amount | Unified Memo            |
|---------------|----------------|-------------------------|
| 2024-11-03    | 50000.00       | 20241103 Goods payment  |
| 2024-11-05    | 1280.00        | Didi Mobility batch     |
| 2024-11-08    | -30000.00      | Pay Supplier A          |
| ...           |                |                         |

[Company Ledger] (excerpt)
| Unified Date  | Unified Amount | Unified Memo              |
|---------------|----------------|---------------------------|
| 2024-11-03    | 50000.00       | Receipt from Customer A   |
| 2024-11-05    | 1280.00        | Reimbursement - Didi      |
| 2024-11-09    | -30000.00      | Pay Supplier A goods      |
| ...           |                |                           |

[Reconciliation Rules]
1. Same transaction: equal amount (tolerance ±0.01) + date difference ≤ 3 days → considered matched, do not output;
2. Equal amount but date difference > 3 days → flag as "Date Lag" — could be a bank-side posting delay or a timing difference in booking;
3. Unequal amounts → flag as "Amount Diff" and keep both sides' amounts for human review;
4. Appearing on only one side → flag as "One-Side Outstanding".

Please output an [Outstanding Items List] categorized per rules 2/3/4, with each entry showing both sides' date, amount, and memo.
Also give me the totals: how many matched, how many likely date-lag, how many amount-diff, how many one-side outstanding.

AI returns a categorized list directly — more than 10× faster than a hand-rolled VLOOKUP (VLOOKUP is the Excel function that looks up a value in another table by a key column; in reconciliation scenarios the formula gets unwieldy).

Step 3: Manual Review + Bank Reconciliation Statement (10 minutes)

AI gives you "suggested matches," and the final call on outstanding items sits with the accountant:

1. Look at "Date Lag": common causes are a bank's interest-accrual day falling on the last day of the month while the booking date is the 1st of the next month — in these cases, just adjust the booking period per the rule.
2. Look at "Amount Diff": usually a service fee (e.g., the bank deducts 5 yuan as a fee while the company books the full amount), or a red-letter reversal on one side (red-letter means a negative entry used to cancel a prior error).
3. Look at "One-Side Outstanding": in the bank but not the books = the company missed a posting; in the books but not the bank = an in-transit item (e.g., a check issued in late November that the payee cashes in early December).
4. Issue the bank reconciliation statement (bank-statement balance + company-received-but-bank-hasn't – company-paid-but-bank-hasn't = book balance) — this is the accountant's statutory step, and AI cannot substitute for it.

Advanced variant: use Power Query (Power Query is Excel's built-in data-wrangling tool that can pull from multiple Excels / databases, merge, and transform) for "fully automated reconciliation" — set the two sheets as queries, refresh automatically on every open, and only involve AI prompts for "items with differences." For colleagues already comfortable with Power Query, this step halves the time again.

Case 3: AI Three Statements — Generate Report Drafts in Plain English

Background: at a services firm, finance director Mr. Zhang delivers the three statements (income statement, balance sheet, cash flow statement) to the boss on the 8th of every month. Pulling the numbers from the trial balance used to take 2 hours; now 5 minutes for a draft.

Step 1: Export This Month's General Ledger from 金蝶 / 用友 / 管家婆 (5 minutes)

Export the general ledger as .xlsx. The columns are usually: Date / Voucher No. / Memo / Debit Account / Debit Amount / Credit Account / Credit Amount.

Step 2: Feed the Ledger + Report Template + Prompts to AI (5 minutes)

Prompt template (income statement example):

You are my reporting assistant. My company is a general VAT taxpayer, subject to the Accounting Standards for Business Enterprises.

Below is my general-ledger data for November 2024 (excerpt):
| Date       | Voucher No. | Memo                | Debit Account       | Debit Amount | Credit Account  | Credit Amount |
|------------|-------------|---------------------|--------------------|--------------|-----------------|---------------|
| 2024-11-30 | J-1101      | Carry forward revenue| Main Operating Revenue| 500000      | Retained Earnings| 500000       |
| ... (paste your general ledger) |

Please generate an [Income Statement Draft] for November 2024, formatted per the standard template in the Application Guide to the Accounting Standards for Business Enterprises issued by the Ministry of Finance:
- I. Operating Revenue
  - Less: Operating Costs
  - Less: Taxes and Surcharges
  - Less: Selling Expenses
  - Less: Administrative Expenses
  - Less: R&D Expenses
  - Less: Finance Expenses
  - Add: Other Income
  - Add: Investment Income
  - Add: Gains from Changes in Fair Value
  - Add: Credit Impairment Losses (losses entered with "-")
  - Add: Asset Impairment Losses
  - Add: Gains from Asset Disposal
- II. Operating Profit
  - Add: Non-operating Income
  - Less: Non-operating Expenses
- III. Total Profit
  - Less: Income Tax Expense
- IV. Net Profit

For each line, give me the number + the calculation source (e.g., "Main Operating Revenue 500,000 = total carry-forward from revenue accounts to Retained Earnings").
Where the number is questionable, flag it with "⚠️ Needs Review."

AI returns a draft report with annotations — that's what a "draft" should look like: every number has a source, every doubt is flagged.

Step 3: Manual Cross-footing + Adjustment (30 minutes)

"Cross-footing" is accounting jargon for verifying that the figures in different statements back each other up instead of contradicting each other. This step is one AI cannot replace:

1. Income Statement Net Profit = Balance Sheet's change in Retained Earnings (ending – beginning) — if they don't match, go back to the general ledger.
2. Balance Sheet Total Assets = Total Liabilities + Total Owner's Equity — this is the accounting identity, and it always holds.
3. Cash Flow Statement ending cash = Balance Sheet ending Cash & Cash Equivalents — usually there's a small difference, and the gap equals "cash equivalents" (e.g., short-term wealth-management products maturing within three months).
4. Items with unusual swings: for example, Administrative Expenses doubling in one month — AI may not flag it, but the accountant must look manually — it could be a missed posting, or it could be a genuinely large expenditure. Once cross-footing passes, paste the statements back into 金蝶 / 用友 / 管家婆's reporting module and have the accountant manually "post" or "issue" them.

Advanced variant: if you're already on Power BI (PivotTable is the basic version; Power BI is the advanced one, capable of interactive dashboards), pipe 金蝶 / 用友 data into a Power BI model and write DAX measures to auto-compute the three statements (DAX is the formula language specific to Power BI and Power Pivot, similar to Excel functions but more powerful). Refresh the data once a month and the statements come out by themselves — AI only handles "interpreting variances" and "explaining anomalies." Once this setup is in place, the monthly report drops from 2 hours to 15 minutes, but the upfront investment is heavy, so it suits companies with a large recurring reporting load.

4. Principle Recap

The three cases look different on the surface, but the playbook behind them is the same — the three-piece combo of OCR + LLM + business-system API:

Stage	What AI Does	Why It Can Do It	Where the Boundary Is
OCR (Qwen, Baidu OCR, PaddleOCR)	Pulls numbers, fields, and tables out of images / PDFs	Deep learning hits 99%+ accuracy on printed Chinese characters; VAT invoices are standardized, so accuracy is even higher	Handwriting, blurry scans, and overlapping seals reduce accuracy — critical documents must be human-verified
LLM (千问, ChatGPT, Kimi, Claude)	Understands unstructured text / tables, classifies them, generates new text	LLMs excel at pattern matching and language understanding; mapping "Didi Mobility" to "Travel – Transportation" is a correlation learned from massive data	LLMs "hallucinate" with a straight face (industry term), so numeric fields must be human-checked, and AI-suggested accounts must be human spot-checked
Business-system API / macros (金蝶 / 用友 APIs, VBA macros, Power Query)	Writes AI-processed results into real finance systems	An API is the software's little outward-facing window; VBA can record scripts; Power Query can pull data automatically	Write operations must carry authorization and audit logs (i.e., a record of who changed what) — finance systems' write permissions must be tightly controlled

The principle in one line: AI handles "moving + classifying + drafting"; accountants handle "review + judgment + signing." These two jobs cannot be merged — merging them is where compliance risk enters.

5. Pitfall Guide

Finance scenarios have more "traps" than other domains, and each one below deserves a careful read.

Pitfall 1: Posting AI's "Suggested Entries" Directly — This Is a Compliance Red Line

Consequence: violates Article 14 of the Accounting Law (PRC) — "Accounting institutions and accounting personnel must verify the original vouchers in accordance with the unified national accounting system." Light consequences: messy books; heavy consequences: deemed false bookkeeping.

Correct approach:

• AI output → accountant spot-checks each one (≥20%) → manual entry into the system → voucher review workflow → post;
• Never let AI auto-post. Whether you write VBA or use RPA, the "Post" button must be clicked by a human;
• In 金蝶 / 用友, give AI tools "read-only" permissions rather than "modify" — AI reads but does not write, and the final write is a human action.

Pitfall 2: Uploading Invoice Photos to Public AI — This Is a Data-Security Red Line

Consequence: customer invoices and bank statements are sensitive data; sending them to third-party AI platforms risks data leakage and may violate the Data Security Law (PRC) and the Personal Information Protection Law (PRC).

Correct approach:

• Prefer privately deployed OCR / LLM (e.g., Alibaba Cloud Proprietary Cloud, Huawei Cloud Stack, Azure Private Deployment) so data never leaves the corporate network;
• When public AI is unavoidable: ① Desensitize first (mask the company name, tax ID, bank account); ② Pick versions that support "data not used for training" (ChatGPT Enterprise, Claude for Work, paid 千问 API all support this); ③ Sign a Data Processing Agreement (DPA) with the AI vendor;
• Prefer the AI modules built into 金蝶 / 用友 — the vendors have already handled the data on a compliance basis.

Pitfall 3: AI Misreads the Buyer Name or Tax ID on a Special Invoice — This Is a Compliance Problem at Entry

Consequence: VAT input tax cannot be deducted, the invoice is deemed non-compliant, and the company's tax-credit rating may take a hit.

Correct approach:

• The buyer information (company name, tax ID, address, phone, bank account) in the OCR output must be checked against the tax-control system or the e-invoice service platform;
• Run a fixed monthly "invoice-compliance check": verify every booked invoice through the "Comprehensive VAT Invoice Service Platform," and let AI only do the "first-pass screening."

Pitfall 4: AI Files "Business Entertainment" Under "Travel – Meals" — This Is a Tax-Handling Minefield

Consequence: Business Entertainment is capped at "60% of actual spend and not exceeding 5‰ of operating revenue," while Travel Expense has no such cap; mis-categorizing leads to a large tax adjustment at the annual reconciliation.

Correct approach:

• List an "easily confused accounts comparison" in the prompt (Entertainment vs. Travel, Advertising vs. Promotion, R&D vs. Administrative);
• Large-amount expenses (single transaction > 10,000) must be human-verified on account;
• Add a line to the prompt: "If the invoice information is not enough to determine the account, output '⚠️ Needs Human Judgment,' don't guess." — better to flag than to guess wrong.

Pitfall 5: AI-Generated Statements "Look Right" but Don't Cross-foot — This Is a Gap in Accounting Common Sense

Consequence: figures in different statements contradict each other, the boss questions them, and you redo the work.

Correct approach:

• Always do the three-step cross-foot: ① Assets = Liabilities + Equity; ② Net Profit = change in Profit Distribution; ③ Cash = Balance Sheet Cash & Cash Equivalents;
• AI-generated statements do not come with cross-foot checks — AI doesn't know the "accounting identity," it just generates structure;
• Setting up "cross-foot formulas" inside 金蝶 / 用友's reporting module for automatic checks is a safer safety net.

Pitfall 6: Blindly Trusting the "Generic Chart of Accounts" in AI's Prompt — This Is a Business-Knowledge Misalignment

Consequence: every company's account structure differs; AI's generic knowledge may produce accounts that don't line up with your actual chart.

Correct approach:

• Always attach your real chart of accounts before feeding data (a screenshot of the table is fine — let AI see it);
• Add a line to the prompt: "If the invoice information is not enough to determine the account, output 'Needs supplementary info: xxxx,' don't guess."

6. Advanced Extensions

Once the three cases are running smoothly, you can extend them in the following three directions — and these are also what later articles in the vba.net series will cover in depth:

1. VBA Macros for Auto-Reconciliation (Pairs with Case 2)

If you reconcile daily or weekly, bake the reconciliation rules into a VBA macro and let Excel run them:

• Use Workbooks.Open to open the bank statement and the company ledger;
• Use Range.Find or Scripting.Dictionary for high-speed matching (Scripting.Dictionary is VBA's built-in high-speed key-value data structure, more than 50× faster than VLOOKUP);
• Output the differences to a third sheet, and humans only look at the differences;
• Once this is in place, reconciling 10,000 transactions drops from 2 hours to 5 minutes.

For reference, see the B-track articles in this series (B14 covers SQL, B18 covers batch scripting, B26 covers M queries).

2. Power Query for Automated Data Pull (Pairs with Cases 1 and 3)

Power Query is Excel's built-in ETL tool (ETL = "Extract-Transform-Load," moving data from one place to another while tidying it up):

• Set the .xlsx exports from 金蝶 / 用友 as data sources;
• Auto-refresh on every open, with AI only involved at the "categorize / generate" step;
• Combine with PivotTable (a PivotTable summarizes numbers quickly across dimensions) for dynamic analysis and use DAX to write year-on-year / month-on-month measures;
• This setup suits companies whose reporting needs are stable and repeat every month.

3. AI Agent + 金数据 / 简道云 as Invoice-Collection Front-End (Pairs with Case 1)

Have employees fill out a 金数据 / 简道云 form and upload invoice photos when they file expenses, triggering an automated flow:

• Employee submits → 金数据 / 简道云 fires a webhook (a webhook is an "event notification" — it shouts when something happens);
• The webhook calls the Qwen OCR API → calls the LLM to categorize the account → writes the entry into 金蝶 / 用友's "Pending Review" voucher pool;
• The accountant sees the vouchers tagged "AI Pending Review" in 金蝶 / 用友, and a click of "Approve" posts them.

This is end-to-end automation — from employee taping the invoice to booking the entry, all running automatically, with the accountant only stepping in at the final review. This is a textbook scenario for the AI Agent series — see C6_AI Agent in this series.

Closing note: the finance domain is where AI deployment is both most valuable and most risky. The value — AI really can lift finance out of the "invoice-taping machine" and let accountants focus on higher-value work (tax planning, budget analysis, decision support). The risk — blurred compliance boundaries, AI mis-judgments, unclear signature responsibility. Hold the line of "AI drafts, the accountant signs off," and you're a finance professional who uses AI well; cross that line, and you're risking both your company and your own career. That line matters more than any prompt-engineering trick.

---

Recap of key takeaways:

1. AI bookkeeping: OCR invoices → AI categorizes accounts → accountant reviews and posts;
2. AI reconciliation: unify fields across two sheets → AI finds outstanding items → accountant issues the bank reconciliation statement;
3. AI reporting: feed the general ledger to AI → generate three-statement drafts → accountant cross-foots and issues;
4. Iron rule: AI drafts, the accountant has the final say — the right to sign is never delegated.