AI发票核对：三家都算出940美元，后续验收指令却有差别
六行发票直接相加是2,152美元。去除重复，再核对单价、数量和算术后，可以进入付款审核队列的只有940美元。ChatGPT、Gemini和Claude都算对了这个数字。真正的差别出现在交给下一位同事的指令里。这次任务，WekeyLab AI选择ChatGPT的答复作为底稿。
我们会采用ChatGPT的交接稿
ChatGPT给出了五张独立发票的核对表、可送审的940美元、另列的572美元暂挂金额，以及每项异常的后续动作。尤其是East收货量未知这一项，它明确要求至少10件可用于本次开票。Lake的错误总额也没有被悄悄改成200美元后放行。
Gemini用更长的过程得出了同样的金额。Claude最初的判定表也正确，但后续却要求确认收货量是否为10件或以下。这把条件方向反了。最终合计正确，并不等于每一句操作指令都可以直接复制。
同一道问题，六行有意设置不同情况的数据
场景是一支虚构的美国办公用品团队，允许部分开票。通过条件包括：发票单价等于采购订单单价；标注总额等于数量乘单价；开票量既不超过订购量，也不超过可用于开票的已收货量。收货量未知必须暂挂。本题不含税费、运费、汇率换算或容差。
去重键是供应商加发票号。North N100与West N100不是同一张发票；完全相同的两行North N100只算一次。我们在提问前固定这些规则和检查项，并向三个网站各提交一次相同英文问题，不限制答复长度。中文与韩文版本是同一次运行的本地化，美元金额保持一致，不另造人民币测试。
采购订单按订购量／单价／可开票收货量排列：North P1为100／8美元／80；West P2为20／15／20；East P3为10／25／未知；South P4为12／10／12；Lake P5为5／40／5。发票则为North N100的80×8=640美元两行，West N100的20×15=300，East E200的10×25=250，South S300的12×11=132，以及Lake L400的5×40但标注总额190美元。完整输入保留在下方原文中。
核对基准是1,512美元，不是2,152美元
原始六行合计2,152美元，扣除重复的North 640美元，剩下五张发票，标注金额合计1,512美元。North 640加West 300通过，共940美元；East 250、South 132、Lake 190暂挂，共572美元。因此1,512=940+572。这是当前发票标注金额的核对，不是最终付款预测。
North订购100件、开票80件、已收80件，部分开票符合规则。South乘法正确，但11美元单价与采购订单10美元不符，12件合计差12美元；不能把132改为120就放行。Lake算术结果200美元，比标注190多10美元；也不能直接作为200美元通过。
我们另用Python Decimal独立计算各行与合计，再对三份原文逐一检查预先确定的8项标准，共24项。原始合计和去重合计属于额外的核对项，问题没有明确要求两项都写出，所以缺少它们不等于算错。测试没有连接财务系统，也没有付款。
ChatGPT：简洁，但关键条件没有省略
它删除North重复行，同时保留发票号相同的West。暂挂标注金额与可送审金额分别列明，Lake不得自动改数的说明也很直接。交接稿包含识别字段，接手人不用回头翻整道问题。
East的条件写成至少10件，方向正确，而且保留了可用于开票的收货量这一限定。在这个案例中，没有发现错误的逐单判定或凭空增加的付款日期。
我们的修改主要是把572美元暂挂合计和1,512美元控制合计也放进最终交接稿，方便下一位同事核对，不改变它原来的决定。
ChatGPT原文：“If it is at least 10 units, the invoice can then pass the stated checks.” 即至少10件时，才可能通过所述检查。
Gemini：详细推导，再附正式交接备忘录
它先逐行说明，再汇总成表，最后写备忘录。内容有所重复，但需要跟着核算的读者容易找到每一步。West未被误删，未知收货量与错误总额都被暂挂，940美元和572美元也与独立计算一致。
备忘录把第一组称为获准进入付款审核队列。结合上下文，这不是已经付款的声明。我们仍会改成可送审，让状态更直接。结尾承诺自己跟进异常，也应在实际发信人确定后才使用。
South的建议是要求贷项通知单或修订发票。这只是后续处理建议，并不代表已经抵扣或可以立刻解除暂挂。当前仍需解决单价差异，再核对对应单据。
Gemini原文：“Total Held for Checking: $572.00”。
Claude：表格正确，收货复核方向写反
Claude正确区分五张发票，给出940美元，并明确核对通过不等于最终批准。问题出在后续韩文指令：确认P3实际收货量是否为10件或以下。要支持10件开票，必须至少有10件可用于开票，同时开票量还要符合订购量上限。
例如只收9件，必须继续暂挂；收到10件，才可能通过收货条件。这不是文风问题。不过它当下仍把East标成Hold，我们没有据此声称它实际放行了错误付款。需要纠正的是后续判断条件。
备忘录还加入输入没有提供的2025年9月参考期间。这是无依据的报告期间，不是虚构的付款日。它把暂挂金额称为未定，又把East的250美元称为估计值。最终有效应付金额可能尚未确定，但当前标注的暂挂合计仍能算出572美元。
同一英文提问在该会话得到了韩文回答。我们保留原文，没有为得到更理想的结果重新提问。英文交接用途因此还需要翻译。这只是本次会话的表现，不能推成所有Claude会话的固定特征。
Claude原文：“입고팀에 P3 실입고 수량 확인 후 10개 이하인지 재검증”。其中“10개 이하”是10件或以下。
WekeyLab AI改写的付款审核交接稿
下面是我们基于同一虚构数据整理的可复用文案，不是第四个服务的回答，也不是付款凭证。
付款审核交接——五张独立发票。可送审：North/N100/P1 640美元、West/N100/P2 300美元，共940美元。North重复行已排除；West虽发票号相同，因供应商不同仍独立计算。按标注金额暂挂：East/E200/P3 250美元、South/S300/P4 132美元、Lake/L400/P5 190美元，共572美元。核对合计1,512=940+572美元。East须确认至少10件收货仍可用于本次开票，未知不能当作零或已确认。South须解决发票11美元与采购订单10美元的单价差异，再核对支持单据。Lake须确认或修订标注190美元与计算200美元的差异，不自行覆盖。各异常在相应检查通过前分别保持暂挂。本稿只记录核对结果，最终付款批准与付款日期尚未确定。
把规则写进可复用问题
可以这样问：“按我提供的匹配规则核对发票、采购订单和收货记录。使用指定键去重，分别保留未知值和异常，不要自行改数后就当成核实过。给出可进入下一审核阶段的金额、暂挂标注金额和去重控制合计。交接稿须列出异常识别字段、原因、下一项检查，不虚构批准、日期或负责人。”随后附上自己的规则和去除敏感信息的数据。
Oracle官方文档说明三方匹配涉及订单、收货和发票，并使用配置的容差。它支持工作流程背景，不代表本例零容差是所有公司的标准。此次比较的重点，是AI是否遵守实际给定的规则。
采购前比较条件可接着看供应商报价比较；提出采购数量前可看库存补货比较。本题搜索量尚未确认，选题依据是任务实用且可以独立验证。
来源与实际记录
实际网站采集：2026年9月23日09:02–09:06 KST。相同英文问题各提交一次，不限制长度。ChatGPT Free、Think关闭、未显示具体模型；Gemini Flash、套餐未确认；Claude Free、Sonnet 5 Medium。ChatGPT与Gemini用英文，Claude用韩文回答。输入和复制原文指纹一致；独立Decimal计算加24项逐项审核。未连接ERP或执行付款。一个虚构案例，不是模型综合排名。中文、韩文是同一次运行的本地化，保留美元以对应原测试。下方保留完整问题与回答。
Oracle: Match Approval Level Options
구매 견적 비교 / Vendor quote comparison
재고 보충 비교 / Inventory reorder comparison
ChatGPT · 2026-09-23 09:02–09:06 KST
I'm preparing a payment-review handoff for a fictional US office-supply team. Can you reconcile these invoice rows against our purchase orders and receipts, identify what can move to the payment-review queue and what needs checking, and write the handoff for my colleague? Do not actually approve or send payments. Our rules for this exercise: all amounts are USD, with no tax, freight or currency conversion. Each PO below has one line. To pass, an invoice's unit price must equal its PO price, its stated total must equal quantity times unit price, and its quantity must not exceed either the ordered quantity or the received quantity available for billing. A lower quantity is allowed: we accept partial invoices. There is no tolerance for differences. Unknown receipt quantity is not zero and cannot pass until checked. A wrong stated total is held; do not silently replace it with your calculation and pass it. Duplicate key is supplier plus invoice number, not invoice number alone. Identical repeated rows with the same key represent one invoice; count it once. All receipts and POs below are otherwise unbilled, and there are no other invoices. Keep holds separate from passed amounts; do not invent a payment date or treat a matching check as final approval. PO | Supplier | Ordered units | PO unit price | Received units available for billing P1 | North | 100 | 8.00 | 80 P2 | West | 20 | 15.00 | 20 P3 | East | 10 | 25.00 | unknown P4 | South | 12 | 10.00 | 12 P5 | Lake | 5 | 40.00 | 5 Row | Supplier | Invoice no. | PO | Invoiced units | Invoice unit price | Stated total R1 | North | N100 | P1 | 80 | 8.00 | 640.00 R2 | North | N100 | P1 | 80 | 8.00 | 640.00 R3 | West | N100 | P2 | 20 | 15.00 | 300.00 R4 | East | E200 | P3 | 10 | 25.00 | 250.00 R5 | South | S300 | P4 | 12 | 11.00 | 132.00 R6 | Lake | L400 | P5 | 5 | 40.00 | 190.00 Please show the reconciliation and the amount that can move forward under these rules, along with the follow-up needed for the other rows. Here’s the reconciliation, applying the rules exactly and counting the duplicate North rows only once. | Invoice key | PO | Qty | Unit price | Stated total | Result | Reason | | ------------ | -- | --: | ---------: | -----------: | -------- | ---------------------------------------------------------------------------------------- | | North + N100 | P1 | 80 | $8.00 | $640.00 | **Pass** | Price matches; $8 × 80 = $640; 80 ≤ 100 ordered and 80 received | | West + N100 | P2 | 20 | $15.00 | $300.00 | **Pass** | Price matches; $15 × 20 = $300; 20 ≤ 20 ordered and received | | East + E200 | P3 | 10 | $25.00 | $250.00 | **Hold** | Receipt quantity is unknown, so billing quantity cannot be confirmed | | South + S300 | P4 | 12 | $11.00 | $132.00 | **Hold** | Invoice unit price does not match PO price of $10.00 | | Lake + L400 | P5 | 5 | $40.00 | $190.00 | **Hold** | Stated total is incorrect: 5 × $40 = $200; do not replace the stated total automatically | ### Amount that can move forward * **North / N100:** $640.00 * **West / N100:** $300.00 * **Total eligible for payment-review queue: $940.00** * **Held amount as stated on invoices:** $250.00 + $132.00 + $190.00 = **$572.00** The two North rows (R1/R2) have the same supplier + invoice-number key and are identical, so they represent **one $640 invoice**, not two. ### Follow-up * **East / E200:** Verify the receipt quantity available for billing. If it is at least 10 units, the invoice can then pass the stated checks. * **South / S300:** Obtain a corrected invoice or otherwise resolve the $11.00 unit-price mismatch against the $10.00 PO price. * **Lake / L400:** Have the invoice corrected/confirmed because the stated $190.00 total does not equal 5 × $40.00. The matching calculation would be $200.00, but it must **not** be substituted automatically. ### Handoff to your colleague Payment-review handoff: **Move forward to payment review — $940.00 total** * North / invoice N100 / PO P1 — $640.00 * West / invoice N100 / PO P2 — $300.00 * North N100 appeared twice (R1/R2), but the duplicate key is supplier + invoice number, so count it once. **Hold / follow up** * East / E200 / P3 — receipt quantity is unknown; verify received quantity available for billing before passing. * South / S300 / P4 — invoice unit price is $11.00 vs PO price of $10.00; resolve the mismatch. * Lake / L400 / P5 — stated total is $190.00, but 5 × $40.00 = $200.00; obtain clarification/correction rather than substituting the calculated amount. The $940.00 figure is the amount that passes these reconciliation rules and can move to the payment-review queue. This is not final payment approval, and no payment has been approved or sent. If you want, I can also turn this into a compact spreadsheet-style exception report with columns for **pass/hold, reason, owner, and next action**.
Gemini · 2026-09-23 09:02–09:06 KST
I'm preparing a payment-review handoff for a fictional US office-supply team. Can you reconcile these invoice rows against our purchase orders and receipts, identify what can move to the payment-review queue and what needs checking, and write the handoff for my colleague? Do not actually approve or send payments.
Our rules for this exercise: all amounts are USD, with no tax, freight or currency conversion. Each PO below has one line. To pass, an invoice's unit price must equal its PO price, its stated total must equal quantity times unit price, and its quantity must not exceed either the ordered quantity or the received quantity available for billing. A lower quantity is allowed: we accept partial invoices. There is no tolerance for differences. Unknown receipt quantity is not zero and cannot pass until checked. A wrong stated total is held; do not silently replace it with your calculation and pass it.
Duplicate key is supplier plus invoice number, not invoice number alone. Identical repeated rows with the same key represent one invoice; count it once. All receipts and POs below are otherwise unbilled, and there are no other invoices. Keep holds separate from passed amounts; do not invent a payment date or treat a matching check as final approval.
PO | Supplier | Ordered units | PO unit price | Received units available for billing
P1 | North | 100 | 8.00 | 80
P2 | West | 20 | 15.00 | 20
P3 | East | 10 | 25.00 | unknown
P4 | South | 12 | 10.00 | 12
P5 | Lake | 5 | 40.00 | 5
Row | Supplier | Invoice no. | PO | Invoiced units | Invoice unit price | Stated total
R1 | North | N100 | P1 | 80 | 8.00 | 640.00
R2 | North | N100 | P1 | 80 | 8.00 | 640.00
R3 | West | N100 | P2 | 20 | 15.00 | 300.00
R4 | East | E200 | P3 | 10 | 25.00 | 250.00
R5 | South | S300 | P4 | 12 | 11.00 | 132.00
R6 | Lake | L400 | P5 | 5 | 40.00 | 190.00
Please show the reconciliation and the amount that can move forward under these rules, along with the follow-up needed for the other rows.
### Independent Row-by-Row Reconciliation
* **R1 & R2 (North | Inv: N100 | PO: P1):**
* *Analysis:* Identical repeated rows with the same key (North + N100). Evaluated once as a single invoice.
* *Unit Price:* Invoice $8.00 matches PO $8.00.
* *Stated Total:* $80 \times \$8.00 = \$640.00$. Stated total $640.00 is correct.
* *Quantity Check:* Invoiced 80 units $\le$ Ordered 100 units AND Invoiced 80 units $\le$ Received 80 units.
* *Status:* **PASS** ($640.00).
* **R3 (West | Inv: N100 | PO: P2):**
* *Analysis:* Supplier is West, making key (West + N100) distinct from North's N100.
* *Unit Price:* Invoice $15.00 matches PO $15.00.
* *Stated Total:* $20 \times \$15.00 = \$300.00$. Stated total $300.00 is correct.
* *Quantity Check:* Invoiced 20 units $\le$ Ordered 20 units AND Invoiced 20 units $\le$ Received 20 units.
* *Status:* **PASS** ($300.00).
* **R4 (East | Inv: E200 | PO: P3):**
* *Analysis:* Received quantity for P3 is **unknown**.
* *Rule Check:* Unknown receipt quantity cannot pass until verified.
* *Status:* **HOLD** ($250.00).
* **R5 (South | Inv: S300 | PO: P4):**
* *Analysis:* Invoice unit price is $11.00, but PO unit price is $10.00.
* *Rule Check:* Unit price does not match PO price.
* *Status:* **HOLD** ($132.00).
* **R6 (Lake | Inv: L400 | PO: P5):**
* *Analysis:* $5 \text{ units} \times \$40.00 = \$200.00$, but stated total on invoice is $190.00.
* *Rule Check:* Stated total is incorrect. Rules forbid silently correcting math to pass it.
* *Status:* **HOLD** ($190.00).
---
### Reconciliation Summary Table
| Invoice Key | PO | Invoiced Amount | Status | Action Required / Reason |
| --- | --- | --- | --- | --- |
| **North - N100** | P1 | $640.00 | **Pass** | Move to payment-review queue (R1/R2 deduplicated). |
| **West - N100** | P2 | $300.00 | **Pass** | Move to payment-review queue. |
| **East - E200** | P3 | $250.00 | **Hold** | Confirm receiving quantity for PO P3. |
| **South - S300** | P4 | $132.00 | **Hold** | Price variance: Invoiced at $11.00 vs PO price of $10.00. |
| **Lake - L400** | P5 | $190.00 | **Hold** | Stated total error: $5 \times \$40.00$ should be $200.00. |
* **Total Ready for Payment-Review Queue:** **$940.00**
* **Total Held for Checking:** **$572.00**
---
### Handoff Memo
**To:** Payment Review Team
**From:** Accounts Payable Reconciliation
**Subject:** Invoice Reconciliation Handoff & Exception Review
Hi Team,
I have completed the preliminary 3-way match reconciliation for the latest batch of office supply invoices against our open purchase orders and receiving logs. Please find the queue routing details below:
#### 1. Approved to Move to Payment-Review Queue ($940.00 Total)
The following invoices have successfully matched unit prices, line arithmetic, and receiving thresholds:
* **North (Inv #N100 / PO #P1):** **$640.00** (80 units @ $8.00/unit). *Note: R1 and R2 were duplicate submissions of the same invoice key and have been deduplicated.*
* **West (Inv #N100 / PO #P2):** **$300.00** (20 units @ $15.00/unit).
#### 2. Exceptions / Items On Hold ($572.00 Total)
The following items require further investigation before moving forward:
* **East (Inv #E200 / PO #P3) – $250.00:** On hold due to missing receipt confirmation. **Follow-up:** Check receiving log or contact warehouse receiving to confirm physical arrival of the 10 units.
* **South (Inv #S300 / PO #P4) – $132.00:** On hold due to a price discrepancy. Invoiced unit price is $11.00 versus agreed PO price of $10.00. **Follow-up:** Reach out to procurement/vendor to request a credit memo or revised invoice at the $10.00 rate.
* **Lake (Inv #L400 / PO #P5) – $190.00:** On hold due to a calculation error. Stated total is $190.00, but $5 \text{ units} \times \$40.00\text{/unit} = \$200.00$. **Follow-up:** Request a corrected invoice from Lake for the proper total amount.
Please proceed with processing the $940.00 approved queue. I will follow up on the missing receipt and price/math discrepancies.
Best regards,
AP SpecialistClaude · 2026-09-23 09:02–09:06 KST
I'm preparing a payment-review handoff for a fictional US office-supply team. Can you reconcile these invoice rows against our purchase orders and receipts, identify what can move to the payment-review queue and what needs checking, and write the handoff for my colleague? Do not actually approve or send payments. Our rules for this exercise: all amounts are USD, with no tax, freight or currency conversion. Each PO below has one line. To pass, an invoice's unit price must equal its PO price, its stated total must equal quantity times unit price, and its quantity must not exceed either the ordered quantity or the received quantity available for billing. A lower quantity is allowed: we accept partial invoices. There is no tolerance for differences. Unknown receipt quantity is not zero and cannot pass until checked. A wrong stated total is held; do not silently replace it with your calculation and pass it. Duplicate key is supplier plus invoice number, not invoice number alone. Identical repeated rows with the same key represent one invoice; count it once. All receipts and POs below are otherwise unbilled, and there are no other invoices. Keep holds separate from passed amounts; do not invent a payment date or treat a matching check as final approval. PO | Supplier | Ordered units | PO unit price | Received units available for billing P1 | North | 100 | 8.00 | 80 P2 | West | 20 | 15.00 | 20 P3 | East | 10 | 25.00 | unknown P4 | South | 12 | 10.00 | 12 P5 | Lake | 5 | 40.00 | 5 Row | Supplier | Invoice no. | PO | Invoiced units | Invoice unit price | Stated total R1 | North | N100 | P1 | 80 | 8.00 | 640.00 R2 | North | N100 | P1 | 80 | 8.00 | 640.00 R3 | West | N100 | P2 | 20 | 15.00 | 300.00 R4 | East | E200 | P3 | 10 | 25.00 | 250.00 R5 | South | S300 | P4 | 12 | 11.00 | 132.00 R6 | Lake | L400 | P5 | 5 | 40.00 | 190.00 Please show the reconciliation and the amount that can move forward under these rules, along with the follow-up needed for the other rows. ## 대사 결과 (Reconciliation) | Row | Supplier/Invoice | PO | 단가 일치 | Total 검산 | 수량 조건 | 판정 | |---|---|---|---|---|---|---| | R1+R2 (중복키, 1건 처리) | North/N100 | P1 | 8.00=8.00 ✓ | 80×8.00=640.00 ✓ | 80≤100(주문), 80≤80(입고) ✓ | **Pass** | | R3 | West/N100 | P2 | 15.00=15.00 ✓ | 20×15.00=300.00 ✓ | 20≤20(주문), 20≤20(입고) ✓ | **Pass** | | R4 | East/E200 | P3 | 25.00=25.00 ✓ | 10×25.00=250.00 ✓ | 입고수량 **unknown** → 확인 불가 | **Hold** | | R5 | South/S300 | P4 | 11.00≠10.00 ✗ | (참고: 12×11.00=132.00, 그러나 단가 자체 불일치) | — | **Hold** | | R6 | Lake/L400 | P5 | 40.00=40.00 ✓ | 5×40.00=**200.00**, 명세 총액 190.00 ✗ | 5≤5(주문), 5≤5(입고) | **Hold** | - North N100은 R1/R2가 동일 키(공급사+송장번호)로 완전히 동일한 행이므로 1건으로만 집계. ## Payment-Review Queue로 이관 가능 (Pass) | Supplier | Invoice | PO | 금액 | |---|---|---|---| | North | N100 | P1 | $640.00 | | West | N100 | P2 | $300.00 | | **합계** | | | **$940.00** | ※ 대사 통과는 최종 승인이 아니며, payment-review 대기 상태로만 이관. ## 보류 (Hold) 및 후속 조치 필요 | Supplier | Invoice | 보류 사유 | 필요 조치 | |---|---|---|---| | East / E200 | P3 | 입고수량(unknown) 미확인 — unknown은 0이 아니며 확인 전 통과 불가 | 입고팀에 P3 실입고 수량 확인 후 10개 이하인지 재검증 | | South / S300 | P4 | 송장 단가 $11.00 ≠ PO 단가 $10.00 | 공급사에 단가 상이 사유 문의 또는 PO/계약 조건 재확인. 임의로 $10 기준 재계산하여 통과 처리 금지 | | Lake / L400 | P5 | 명세 총액 $190.00 ≠ 계산값 $200.00 (5×40.00) | 공급사에 정정 송장 요청. 계산값으로 임의 대체 후 통과 처리 금지 | --- ## 인계 메모 **제목**: 사무용품 공급사 송장 대사 결과 — Payment-Review 이관 및 보류 건 (2025년 9월 기준, 예시) **상태**: 대사 완료, 승인/지급 미실행 **근거(Rationale)**: PO 5건, 인보이스 6행(중복 제거 시 5건) 대사. 단가 일치·총액 검산·수량(주문/입고 한도) 3개 기준 모두 충족한 건만 Pass 처리. 하나라도 미충족 시 자동 보정 없이 Hold. **금액(Cost)**: - 이관 가능(Pass): $940.00 (North $640.00 + West $300.00) - 보류(Hold): 금액 미확정 — East $250.00(추정, 입고 확인 전), South $132.00(단가 불일치로 보류), Lake $190.00 vs $200.00(총액 불일치) **예상 효과(Expected effect)**: $940.00을 payment-review 큐로 이관 시 정상 대사 완료 건 우선 처리 가능. 나머지 3건은 원인 확인 전까지 지급 보류로 리스크 차단. **대응(Response)**: 1. East(P3): 입고 담당자에 실입고 수량 확인 요청 2. South(P4): 공급사에 단가 상이 사유 확인 3. Lake(P5): 공급사에 총액 오류 정정 인보이스 요청 4. 위 3건은 확인 완료 전까지 payment-review 큐에 올리지 않음 궁금하신 부분이나 기준 조정이 필요하면 말씀해 주세요.