Way 1: without AI (the manual way)
Honest cost: 2-3 hours before every close.
- Read each payee string and remember how you coded it last time.
- Code row by row, checking past months for consistency.
- Flag the weird ones for the client or senior.
Coding transactions one by one means re-deciding the same vendor every month. The walkthrough below sets up a consistent categorization run: establish your category rules once, apply them in batches with the evidence noted per row, and park the ambiguous ones as questions for the client - never as guesses.
Pros of doing it manually
- Free except your time - no tools, no accounts.
- You learn the job deeply and control every detail.
- No data leaves your hands.
Cons of doing it manually
- 2-3 hours before every close - every single time the job runs.
- Quality depends on your energy that day.
- More volume just means more hours; it does not scale.
Way 2: with AI, but without AI Agent Skills (DIY prompting)
A chat assistant genuinely helps here. The catch is that you become the workflow designer. What you have to figure out yourself:
- You must paste your category list and past examples each run, or coding drifts.
- You must set a confidence bar in the prompt, or everything comes back equally certain.
- You must approve the batch yourself - nothing should post unreviewed.
Here is a starter prompt that works - copy it and bring the inputs listed:
Propose categories for these uncoded transactions: [paste date, payee, amount rows]. Follow these past coding examples: [paste]. Per row: proposed category, the evidence, confidence. Mark anything under 70% confident as REVIEW. Proposals only - do not post anything.More on this prompt: the free ai transaction categorization prompt page.
Pros of DIY prompting
- Free, or nearly - any chat assistant works.
- A first draft in minutes instead of hours.
- Fully flexible: tweak the prompt for any edge case.
Cons of DIY prompting
- You become the workflow designer: prompts, test data and output rules are yours to write, test and maintain.
- Output drifts between sessions unless you re-paste the rules every run.
- Nothing is saved for you - sheets, structure and context are rebuilt each time.
Way 3: with AI Agent Skills (free, Free)
The Transaction Categorization Agent is this job already designed: fixed evidence rules, output contract, test fixtures and stop points. You ask once:
Hello, categorize these uncoded transactions
And a finished, review-ready artifact lands in your own Drive:
| Date | Payee | Amount | Proposed category | Evidence | Review status |
|---|---|---|---|---|---|
| Sep 10 | AWS EMEA | $612.40 | Infra & hosting | Recurring monthly invoice matches AWS-2026-08 | Ready to post |
| Sep 11 | AMEX *TRAVELCO | $1,140.00 | Unclear - no trip logged | Card statement only; no expense report filed | Needs review |
Sample data shown. Your workspace fills with your real work.
Setup takes about 10 minutes: download, hand the Included files to the assistant you already use (Instinct, ChatGPT, Claude, Gemini or any other assistant that reads files), approve the fictional-data test. Then it runs on a standing cycle or a one-line ask, and stops at every human gate. Full method.
Pros of the ready-made agent
- About 10 minutes to a tested setup, not an afternoon of prompt tuning.
- The same output shape every run - evidence rules and stop points are already designed.
- Runs on a standing cycle or a one-line ask; files land in your own Drive.
Cons of the ready-made agent
- Costs free, where DIY prompting is free.
- The output format is opinionated - you adapt the template rather than rebuild it.
- It still stops and waits for your approval at decision points; it is not zero-touch.
The honest comparison
| Way 1: manual | Way 2: DIY AI | Way 3: free agent | |
|---|---|---|---|
| Time to first result | 2-3 hours before every close | An afternoon of prompt tuning | About 10 minutes, tested setup |
| Every run after | Same hours, every time | Re-paste the rules, sheet and context each run | One short ask, or a standing cycle |
| Consistency | Depends on the day | Drifts between sessions | Same output shape every run |
| Cost | Your hours | Free, plus your time | Free, files kept forever |
The manual way teaches you the job. DIY prompting is fine for a one-off. The agent wins when the job repeats every week.
More: the Transaction Categorization Agent product page · the AI transaction categorization guide · AI agents for bookkeepers
Questions people ask
What is the fastest way to categorize uncoded transactions?
The free ready-made agent: about 10 minutes of setup, then one short ask per run. Manual takes 2-3 hours before every close; DIY prompting takes an afternoon of tuning plus upkeep.
Can I categorize uncoded transactions without AI?
Yes - the manual steps are on this page. Expect 2-3 hours before every close each time the job runs.
Is the free agent worth it over DIY prompting?
If the job repeats weekly, yes: you stop rebuilding prompts and re-pasting context every run. For a one-off, DIY prompting is fine and free.
Last updated 13 September 2026 · Reviewed by the AI Agent Skills team.