A small Bangkok accounting firm goes digital, Part 6: AI joins the workflow — OCR extraction, accountant review
Once submissions and the intake list are stable, bring AI into the daily workflow: extract data from voucher images with OCR, let AI arrange it into structured drafts, and have bookkeepers review before posting. AI saves typing and sorting hours; it does not replace judgment — account assignment, posting and review stay with the bookkeeper, and final responsibility stays with the licensed CPA.
Last chapter, clients saw their tiers and upgraded on their own in the PWA. This chapter brings AI inside the firm: let machines do the typing and sorting, and free the bookkeepers’ hands for judgment.
Where a bookkeeper’s hours actually go
The reality of this trade: a large share of working hours goes into typing and sorting, not judgment. When an invoice arrives, the bookkeeper’s job is to key in the date, the party, the amount and the tax number, and assign the right account — the first half is mechanical, the second half is professional.
In the busy season, manual keying is the biggest bottleneck. Images come in and the numbers on paper have to be typed line by line into the software; one error, one omission, and everything downstream stops reconciling.
What AI does: extraction and sorting
The AI introduced in this step does exactly two things:
- OCR extraction: read text and numbers from voucher images — invoice number, date, supplier, amount, tax number;
- Structured drafts: arrange what was recognized into structured records — this is a purchase invoice, this much, from this supplier — lined up as a to-do list.
The bookkeeper receives a draft, not “already posted entries.” The draft’s job is to save the typing: the keys that should be pressed have been pressed by the machine; the accounts to assign and the judgments to make are still the bookkeeper’s.
Judgment always stays with the bookkeeper, responsibility with the licensed CPA
This is the most important boundary of the whole series, and an industry common sense:
- Account assignment: which account this voucher belongs to is the bookkeeper’s call; AI only suggests or leaves it blank;
- Posting and review: a draft must be reviewed by a bookkeeper before posting; anything unusual goes back to a human;
- Final responsibility: the accuracy of the books and filings always rests with the licensed CPA, not the system.
Thai vouchers routinely mix Thai, English and numerals, with plenty of handwriting and inconsistent receipt formats; OCR accuracy is limited. The flow is therefore deliberately designed as “AI first, human second”: the parts recognized correctly save time, the parts it gets wrong a human fixes — across the whole batch, keying hours actually come down.
What to watch in this step
- Is accuracy really good enough? How many vouchers out of ten need the numbers re-keyed by hand? If almost everything needs correcting, the step is premature;
- Does the review flow save time? How much faster is reviewing a draft than keying from zero? Does checking where AI went wrong sometimes take longer than doing it yourself?
- Which vouchers still need a human? Handwriting, damage, weird formats — if AI cannot handle them, mark them honestly and route them to a human.
Only when the data proves the time saved should AI stay in the workflow; if no time is saved, fall back to manual keying and wait for better image quality or better recognition.
Boundaries matter: the OCR/AI voucher extraction engine is an independent system project, not part of the standard website package. AI only extracts and sorts; it does not replace accounting software, the bookkeeper’s judgment, or the licensed CPA’s responsibility.