Opinion: using ChatGPT for invoice tallies often yields close but inaccurate results

"I tried tossing my invoice tallies into ChatGPT, but..."

The business owner started the conversation that way from the other side of the screen share.

"The numbers weren't completely off, but they didn't match up, either."

This 'not completely off, but doesn't match' situation—I think quite a few people can relate to that.

Right now, while working as a company employee, I'm also sitting next to business owners and helping them think through AI utilization for free. This is a story from the second Zoom session.

At the end, I'll leave the entire prompt I created from the bottleneck we hit that day. Even if it's not invoices, you can use it for any task you perform using the same steps every month.

Getting numbers that are 'close but no cigar'

At this person's company, performance reports arrive every month from partner companies in Excel format. The format hasn't changed in about six or seven years. Since it's a partner they find it difficult to make requests to, they accept the files as they are and tally them by hand.

There are two types of contracts: half-day and full-day, with different rates. The table has an AM column and a PM column; rows with a number in only one of them are half-days, and rows with numbers in both are full-days. Therefore, you need to count the 'rows with only one' and the 'rows with both' separately.

Last month, there were 176 half-day entries.

"I want to get this result in one go."

So, they showed me the actual instructions they gave to ChatGPT.

"Calculate the total for column G and column H individually, and the total for rows that have both."

...Ah. I thought, this is it.

It wasn't the AI's fault

At a glance, these instructions look fine. They specify the columns and state what they want to get out of it.

But something is missing.

"Only add up rows where one is blank and the other has a number"

This isn't written down.

Because it's not written, the AI has no choice but to interpret 'total for individual' in its own way. That's why the numbers aren't completely off, but they don't match either. You get numbers that are close, but not quite right.

Now, I imagine some of you reading this might be thinking, 'Well, obviously, I know that without you telling me.'

But you don't. The AI has no way of knowing how your company distinguishes between a half-day and a full day; it's not written anywhere.

'Add up only the rows where one is blank and the other has a number'—putting conditions into words like this is exactly what requirements definition is.

It sounds like a difficult term, but all you're doing is 'saying out loud the decisions you make by hand every month.'

Also, I told them one more thing on the spot.

'If you have the AI calculate directly, it can be slightly off. It's more accurate to have it come up with a formula or calculate it using code.'

Counting numbers isn't the AI's strong suit. It's better suited to building the tools that do the counting.

And, I didn't quite get there either.

So, I suggested we build a conditional counting formula (SUMIFS) together, and we started working on it while screen sharing.

Well, that didn't go smoothly at all.

The range to count and the range to check for conditions were misaligned, leading to errors everywhere. Merged header rows like 'August 2026' were getting mixed into the range, producing strange numbers. We had to remove those and narrow down the range again.

'It's 141 now.'

The goal is 176.

'Maybe some row is off. We're getting closer, though. 176 and 141.'

I had an idea of the cause. With the current condition, we were only picking up one side of the half-days. We needed to add both the rows where AM is blank and PM has a number, and the rows where it's the other way around.

But we ran out of time before we could add the formula for the other side.

Stuck at 141, we carried it over to the next session.

'So, just put it into words yourself' is a dead end.

Writing it this way might make it sound like I spotted the cause and solved it.

In reality, that's not the case. The time we spent together in front of the screen, both of us fiddling with cells and saying, 'Huh, it's still not 176,' was much longer.

And, the thing that bothered me most that day wasn't that.

I can easily say, "Let's put the conditions into words," but the person being asked doesn't know where to start writing. The more routine the monthly task is, the more your hands move on their own, so you don't see it as a process yourself.

I've built about 45 tools within my company, but only a few remained in use. I wrote about this before in The conditions for "tools that get used," learned from 110 version updates in 2 weeks. I've always felt keenly that handing something over is harder than building it.

Saying "Please verbalize it yourself" is like handing something over without actually handing anything over. I previously wrote about a mistake I made with a president who kept saying for half a year, "I wish there was someone to teach me AI", and I stumbled in the same place back then, too.

So, I made it homework.

Instead of me asking questions, I have the AI ask one question at a time. Once they answer, a single sheet of requirements definition is completed at the end. I created and sent a prompt like that. I'll leave it right at the end of this article.

Looking at the resulting sheet is still to come.

The goal of this prompt is not to write the requirements definition for them. If I write it for them, they'll just call me again when the work changes next time. That's not self-sufficiency.

What I'm doing is the same as when I stopped "teaching" AI to my colleagues and they started using it like crazy. I stopped teaching and switched to being the one who listens.

As for me, I haven't decided yet how to hand over the next part, "allocating the total expenses to departments based on the ratio of the number of items." When we meet next, we'll start by looking at the sheet that was produced together.

It hasn't reached 176 yet.

If you read this article and thought, "Oh, that's us too," or "We're stuck here," please write it in the comments. Let's think about it together.

If you try the prompt below and get a sheet, just write "Got it" in the comments. I'll contact you, and we'll decide together in 30 minutes where to start. It's not me who will be doing the building, but you yourself.

I'll leave that prompt here.

Just copy the prompt below and paste it into a new chat in ChatGPT (or Copilot). It will ask you "what kind of work it is" in 1-3 lines at the beginning, so give a rough answer. After that, it will ask you one question at a time.

I've incorporated the sticking point from that day directly into the design.

• Conditions to add were not verbalized → Ask back for judgment rules using concrete examples
• AI was made to calculate directly and it was off → Always make it calculate using formulas or code
• The tools you can use depend on the person → Check your environment first

あなたは、中小企業の業務改善を支援する「業務ヒアリング担当」です。
これから私が普段やっている業務について、あなたが質問しながら中身を聞き出し、
最後に「要件定義シート」と「仕組み化の選択肢」をまとめてください。
# 進め方（必ずこの順番）
## ステップ1：ざっくり入力してもらう
最初に、次の質問だけをしてください。ほかのことはまだ聞かないでください。
「改善したい業務を、1〜3行でざっくり教えてください。
（例：毎月届くExcelを集計して、部署ごとに費用を割り振っている）」
## ステップ2：工程の仮説を作って、確認する
私の答えを読んで、この業務の工程を「たぶんこういう流れですよね」と
3〜8工程の番号つきリストで想像して出してください。
- 各工程は「何を受け取って → 何をして → 何を出す」の形で1行にする
- 想像で書いた部分には必ず（仮）と付ける
- 最後に「合っているところ・違うところ・抜けている工程を教えてください」と聞く
私が直したら、その内容で工程リストを更新してもう一度だけ見せてください。
## ステップ3：工程ごとにインタビューする
確定した工程を1つずつ順番に掘り下げます。
**質問は1回に1つだけ**。私が答えてから次の質問に進んでください。
1つの工程で聞くことは、次の中から「その工程に必要なもの」を選んでください。
- 何を：受け取るもの（ファイル名・形式・どこから届くか）／作るもの（誰に渡すか）
- 手順：実際に手を動かす順番。「クリック」「コピー」「入力」「目で見て確認」まで細かく
- 道具：使っているソフトやサービス（Excel・メール・社内システム・ChatGPTなど）
- 判断ルール：どういう条件のときに、どうするか
  （例：「AMとPMのどちらか片方だけ数字がある行は半日として数える」）
  私の答えがあいまいなら、具体例を1つ出して「この場合はどうなりますか？」と聞き返す
- 例外：いつもと違うパターン、ミスが起きやすいところ
- 手間：1回あたりの時間、頻度（毎日／毎週／毎月）、担当者は誰か
- 困りごと：面倒なところ、時間がかかるところ、間違えやすいところ
聞き方のルール：
- 専門用語を使わない。使うときは一言で説明する
- 私の答えを勝手に言い換えたり、足りない所を想像で埋めたりしない。分からない所は質問する
- 5問ごとに「ここまでの理解」を3行でまとめて、合っているか確認する
## ステップ4：環境を確認する
工程のインタビューが終わったら、次をまとめて確認してください。
- パソコン：Windows か Mac か
- Excel：どのバージョン／Microsoft 365 か。Excelの中で Copilot が使えるか
- メール：Outlook か Gmail か など
- 使えるAI：ChatGPT（無料／有料）、Copilot、Gemini など
- 社内のルールで使ってはいけないもの（外部サービスにデータを入れてよいか など）
## ステップ5：要件定義シートを出す
次の形でまとめてください。私が話していないことは書かず、分からない所は「未確認」と書くこと。
【業務名】
【目的】この業務は何のためにやっているか（1行）
【頻度・時間・担当】
【インプット】受け取るもの（ファイル・形式・届き方）
【アウトプット】作るもの（誰に・どの形で）
【工程】番号つきで「受け取るもの → やること → 出すもの」
【判断ルール】「〜のときは〜する」の形で、条件をすべて
【例外・ミスが起きやすい所】
【使っている道具・環境】
【困りごと】時間がかかる所・面倒な所を、手間の大きい順に
【未確認のこと】
## ステップ6：仕組み化の選択肢を出す
要件定義シートをもとに、改善の方法を次の4段階で提案してください。
各案に「やること」「必要な道具」「準備にかかる時間の目安」「向いている条件」「注意点」を書くこと。
1. Excelの数式や機能だけで解決する
2. ChatGPT などのAIに、ファイルと指示文を渡して処理させる（そのまま使える指示文も書く）
3. 小さな集計ツールを作る
4. メールで届いたファイルを自動で読み込んで処理する仕組みにする
提案のルール：
- **合計や件数などの計算は、AIが頭の中で直接計算しないこと。**
  必ずExcelの数式か、Pythonなどのコードで計算させる方法にする（AIの暗算は微妙にずれることがあるため）
- 私の環境（ステップ4）で使えない方法は、使えない理由を書いたうえで後ろに回す
- 最後に「まず最初に試すならこれ」を1つだけ、理由つきで勧める
では、ステップ1の質問から始めてください。