How to Edit an AI Draft of a Historic Account Book

Historic account books are hard.

The handwriting is only the beginning. Entries may run across multiple lines. Dates and amounts occupy narrow columns. Dollars and cents—or pounds, shillings, and pence—may be separated in ways that aren’t immediately obvious. Text from the facing page can creep across the spine.

AI can give you a useful first draft. But how do you turn that draft into a transcription you can trust?

In our recent webinar on editing historic account books, Ben demonstrated the process he uses while working with the nineteenth-century account books of Virginia planter Jeremiah White Graves. His approach breaks the work into three passes.

First, fix the structure.

Before reading individual entries, look at the page as a whole. Is this one account or several? Does it cover a single year? Are there entries written across other entries? Do most transactions occupy one line or several?

Then load the AI draft and clean up its structure:

• Remove transcribed watermarks, call numbers, and other material that isn’t part of the document.
• Delete false columns created from photographs of the facing page across the spine.
• Make sure amount columns are divided correctly.
• Consolidate entries that continue across multiple lines.

At this stage, don’t worry about whether every word is right. The goal is to make the transcription reflect the organization of the page.

Next, verify the contents.

Once the table is structured correctly, work through the page systematically:

• Verify every date.
• Verify every currency amount.
• Correct the wording of each entry.
• Mark insertions, deletions, superscripts, and unclear text as appropriate.
• Save the page and review it again. 

Ben spends about an hour on a full page of dense accounts. Detailed markup may feel excessive if your research question doesn’t depend on every superscript or deletion. But it serves another purpose: it forces you to look closely instead of accepting a plausible-looking draft.

Finally, use AI to help find what you missed.

This is where AI becomes useful in a different way—not as the original transcriber, but as another set of eyes.

Ben indexes the people named in the account, then compares his corrected transcription with the original AI draft to spot anomalies. He also reads the model’s reasoning, paying particular attention to passages it struggled with.

For account books with debit and credit sides, he takes the analysis one step further. He gives ChatGPT both corrected transcriptions and asks it to look for inconsistencies or possible transcription errors.

The important distinction is that he focuses on the problems ChatGPT identifies, not necessarily the solutions it proposes.

If the arithmetic doesn’t work, an amount may deserve another look. If one page carries forward $11.99 but the previous page appears to end with $11.00, one of those readings may be wrong. If an account from 1836 appears to have been settled in 1834, it’s worth reexamining the date.

The AI doesn’t need to know the right answer to point you toward the right question.

That may be the most useful role for AI in historical research: not replacing close reading, but helping us decide where to look more closely.

Watch the webinar or try AI Drafts with your own account books.