Use case #13 · CPA
CPA finishes two years of write-up work in about four hours with Claude Cowork and confidence scores
- Close/bookkeeping
- Intermediate
- Claude Cowork, Xero
- Source checked · podcast recap
The short version
What they did
Pointed Claude at folders of bank statements (PDFs and CSVs) to extract every transaction into Xero import files, then had it code the transactions against the prior year’s general ledger with a confidence score on each one, and reviewed only the low-confidence items.
Result (as claimed)
“Using Claude Cowork, he finished everything in about four hours, including gathering documents.” “Claude missed just six transactions out of 2,200, and two of those were due to credit card statement date issues, not AI error.”
Prompt
I’m an accountant using Claude Cowork to draft transaction coding for [period]. I’ll provide [last year’s general ledger], [this period’s bank transactions] and [our accounting software’s import layout]. Follow the prior ledger’s account names and coding patterns.
1. Assign each transaction to a proposed account, citing a similar prior-year entry where available.
2. Add a confidence score from 0 to 100% and a short reason for each choice.
3. Return a CSV I can sort by confidence, keeping the original date, description and amount alongside the proposed coding.Add a short summary of uncertain or unusual items and any rows you couldn’t process. Don’t invent accounts or change amounts. Mark unmatched items for review instead of forcing a choice. Confidence is your estimate, not proof the coding is correct. Don’t import or post anything.
Do it yourself
These steps are my suggested way in, not Blake Oliver’s setup.
Level: Intermediate for a full year of statements with an agent like Claude Cowork. Coding one month in a chat is Beginner.
What you need
- Last year’s general ledger detail and this year’s bank and card statements, as CSV or PDF.
- A chat assistant on a business plan, or an agent that can work through folders of files, such as Claude Cowork.
- Your accounting software’s import template.
Start simple
- Start with one month for a client whose prior year you already coded.
- Paste the case’s Prompt with the prior-year ledger and that month’s transactions.
- Sort the result by confidence. Review every item below a cutoff you set, such as 80%.
- Spot-check ten high-confidence items too.
- Compare its coding with what you would have done. Note the patterns it misses.
Take it further
- Let an agent extract transactions from PDF statements into your import format.
- Add standing rules, such as how owner draws or loan payments are coded, to the instructions.
- Run a full year, then review only the low-confidence and unusual items.
Check your work
- Tie each month’s extracted totals to the statement’s beginning and ending balances. That catches missed transactions.
- Watch statement-date cutoffs on credit cards. The source found errors there.
- Import nothing until you have reviewed it. You sign off on the books, not the AI.
Before you rely on it
- This is general information about how accountants use AI, not professional guidance for your engagements.
- You stay responsible for the work under your professional standards, including review and sign-off. An AI tool’s output is a draft, not evidence.
- Before you send client data to an AI tool, read the tool’s data and training terms, and get any client consent you need. For tax return information, that includes IRC section 7216.
Source checked. The link, author, date and quote were confirmed against the source. The result itself has not been reproduced. Prompt adapted for this library, inspired by the case. Business size: Client: a service business, 2,200 transactions across nine accounts. Results are quoted as the source claimed them and never improved. Spot an error or a broken link? Email charlie@cbarmorecpa.com.