A stack of receipts photographed and then forgotten in your phone until month end. This routine breaks bookkeeping into four steps — upload, extract, tabulate, summarise — each with a prompt you can paste as-is. Then a more advanced move: stop directing each step, and hand the AI a goal and a set of rules to work through on its own.
🧾 The four steps
① Upload — Photos or PDFs, all of them at once.
② Extract — Date, merchant, category, amount — only what is actually on the receipt.
③ Tabulate — One row per entry, your columns, as CSV straight into a spreadsheet.
④ Summarise — Total, category shares, the five largest — ready to paste into a monthly report.
① Upload the receipts
Photos, PDFs and screenshots all work. No tidying required, but a few things decide how accurate everything after this will be:
- Shoot the whole receipt, all four corners — crop off the total and every number after it is wrong
- Even light, with no shadow or glare sitting across the figures
- Flatten folds and creases; thermal paper is the worst offender
- Upload a whole month together rather than in batches — anomalies only show up when the month is seen at once
② Extract the fields
This step is not about speed, it is about not guessing. A blurry photo, a merchant name in tiny print, a total sitting under a crease — the model is perfectly capable of smoothing all of that over for you, and that is exactly the danger. Every rule in the prompt exists to make it leave a gap and flag it, instead of handing you a number that merely looks right.
I've uploaded receipts (photos or PDFs). Extract these fields from each one, one block per receipt: - date (YYYY-MM-DD) - merchant - category (meals / transport / office supplies / software / marketing / professional services / travel / other) - amount and currency - payment method (only if it's on the receipt) - receipt number (only if it's on the receipt) Rules: 1. Read only what is actually visible on the receipt. If you can't read it, write "unclear" — do not guess. 2. Take the total printed on the receipt as the amount. If the line items don't add up to that total, give me both numbers and flag the mismatch. 3. List service charge, tip and tax separately. Don't fold them in as ordinary spending. 4. When you're done, list the fields you were unsure about, so I can check them myself.
③ Turn it into a table
The format is yours to choose. The prompt below uses one that works out of the box: date | merchant | category | amount | currency | payment method | receipt number | notes. Want a project or client column? Change one line.
Turn those receipts into a table, one row per entry, with these columns in this order: date | merchant | category | amount | currency | payment method | receipt number | notes Rules: 1. Dates always YYYY-MM-DD, so they sort properly. 2. Amounts as digits only — no currency symbols, no thousands separators (e.g. 1280.50). Currency goes in its own column. 3. Use only the categories I listed. Don't invent new ones. If something genuinely fits none of them, write "other" and explain in the notes column. 4. Leave a field blank if you couldn't read it. Don't write "N/A", and don't guess. 5. Output as CSV so I can paste it straight into Google Sheets or Excel. 6. No totals row at the bottom — I'll calculate that in the spreadsheet.
Two things catch people out: keep currency symbols out of the amount column (a spreadsheet reads them as text and refuses to add up), and don’t let it append a totals row — the first time you sort, it ends up in the middle.
④ The expense summary
A summary is only worth having once the table exists, because every figure in it can be recomputed. This step is not asking the AI for its impressions — it is asking it to calculate the total, the category shares and the five largest expenses from the table, in a form you can paste into a monthly report.
Using the table above, write an expense summary I can paste straight into a spreadsheet or a monthly report: 1. Total spend for the period, by currency 2. Each category's amount and share, largest first 3. The five largest individual expenses, with dates and merchants 4. If I've given you last period's figures, compare: which categories rose, and by what percentage 5. One closing line: the single thing most worth noticing this period Rules: every number must be calculated from the table. Don't add anything that isn't in it, and don't state a conclusion stronger than the data supports.
The bookkeeping floor: What the AI produces is a copy, not evidence — keep the original receipts. Before anything goes into the books, spot-check the totals and the dates: get those wrong and the whole table has to be redone.
🤖 Going further: an AI agent for financial planning
In the first four steps you direct each move. The advanced version inverts that: you state the goal and the rules clearly, and the AI works through the whole routine before it replies. With several dozen receipts a month, discussing them one at a time is unbearable — this is where the time actually comes back.
- Read — Date, merchant, amount and currency from each receipt.
- Categorise — From a fixed list, with no inventing new ones.
- Find anomalies — Repeat charges, above-average amounts, new merchants, totals that don’t reconcile, miscategorised merchants.
- Analyse — Which categories dominate? What recurs monthly, what was a one-off?
- Summarise — One page: total, category breakdown, top 5, the anomaly list.
- Recommend — Three specific savings, each pointing at the expenses behind it.
If you don’t define "anomaly", the model will define it for you. So the prompt pins it down: what percentage counts as high, what window counts as a repeat, what counts as the wrong category.
Goal: from this month's spending, identify the major expenses, the anomalies, and anything worth my attention. I'm giving you every receipt for the month at once. Work through this whole routine yourself — don't check in at each step: 1. Read: extract date, merchant, category, amount and currency from each receipt. 2. Categorise: use only these fixed categories — meals / transport / office supplies / software / marketing / professional services / travel / other. 3. Find anomalies: flag anything matching one of these — - the same merchant charging more than once in a short window - an amount more than 50% above the average for its category - a merchant appearing for the first time this month - a receipt total that doesn't match its line items - a merchant that doesn't belong in the category it landed in 4. Analyse: which categories dominate? Which costs recur every month, and which were one-offs? 5. Summarise: one page or less — total, category breakdown, top 5 expenses, the anomaly list. 6. Recommend: three specific ways to spend less, each pointing at the actual expenses that support it. Rules: - Every conclusion must name the receipt or the line it came from. Don't tell me anything I can't trace back. - Anything you couldn't read, or that's missing data, goes in a "needs input from me" list. Don't fill it in yourself. - No investment or tax advice. Only what the numbers themselves show. - Work through all six steps, then reply once.
The load-bearing rule is the one requiring every conclusion to name the receipt behind it. With it, you can pull a few and check them. Without it, you get a page that reads convincingly and cannot be verified.
One last thing: This saves you the data entry, not the accounting judgement. How to file, and what is deductible, is a question for your accountant — what the AI is genuinely good at is having the numbers in order before they reach anyone.
