Let AI reconcile the day, but not post the books.

At closing time, a small hot-pot restaurant does not have one neat dataset. It has a daily report, a handwritten cash count, supplier invoices and a card settlement slip. AI becomes useful when it prepares these fragments for review without pretending that recognition is approval.

01 Reconciliation is not data entry

The visible task is copying figures, but the real task is deciding whether four records describe the same day. A total on a daily report, cash in the drawer, an expense document and card receipts each represent a different part of the close.

Manual entry used to hide that distinction. The owner would read the register total, type it into a spreadsheet and resolve the unexplained difference from memory, often after the context of the shift had already faded.

A better system preserves the differences between sources. Automation should shorten transcription while making the meaning and origin of every figure easier to inspect.

02 Identify the document before reading the number

Document type determines what a number can mean. A daily report contains sales and tender totals; a handwritten count records physical cash; a supplier invoice records an obligation or expense; a card slip records settlement through one channel.

Image quality is therefore part of the accounting workflow. Photos shared through a staff chat were sometimes compressed enough to erase faint print and table lines, leaving an otherwise capable model unable to distinguish a label from a total.

An unreadable source should create a question, not an estimate. The system asks for the original image or a closer photograph, marks the field as unresolved and leaves the draft incomplete until evidence arrives.

The safe response to an unclear figure is not a more confident guess. It is a visible gap.

03 Payment method is part of the event

An expense document proves that money is owed or spent, but not how it moved. Treating a supplier payment by bank transfer as cash leaving the drawer can produce a large difference even when every amount was read correctly.

The model must keep separate fields for the event and its settlement. Amount, date, source, category and payment method belong together, but none should be inferred from the presence of the others.

A missing payment method should remain missing. The draft can propose a classification from prior patterns, but it must label that proposal and ask the owner to confirm it before the cash position changes.

04 A day balances only when every channel exists

Reconciliation depends on a complete map of how customers paid. If the worksheet has cash and card but no field for a mobile-payment channel used that day, the numbers cannot balance by design.

Forcing an unfamiliar receipt into the nearest known category only conceals the omission. The system should preserve it as an unregistered channel, surface the unmatched amount and invite a one-time mapping that can be reused later.

Completeness matters more than apparent precision. A clean total built from an incomplete channel list is less trustworthy than a draft that openly shows one unresolved row.

05 The draft is the product boundary

AI output should stop at a reviewable draft. Each extracted figure needs its source image, document type, proposed field and confidence or exception status close enough for a person to compare without reconstructing the process.

The owner reviews relationships, not only digits. Do cash movements match the drawer count? Does a transfer appear in cash expenses? Is every sales channel represented? Has the same document been submitted twice?

Only the owner turns the draft into a posted record. That final action confirms both the figures and the business meaning assigned to them.

06 Exceptions deserve their own path

A robust workflow is designed around ordinary failures. Blurred images, contradictory totals, duplicate submissions, missing payment methods and new channels should enter a visible review queue rather than disappear inside a plausible answer.

Corrections should improve the next draft without rewriting history. The system can remember that a document layout belongs to a certain type or that a channel maps to a field, while preserving who changed the current record and why.

Success is not the highest possible automation rate. It is fewer numbers typed by hand, fewer unexplained differences at closing and no uncertain item quietly becoming fact.

07 Keep the final click human

The last click is not leftover friction. It is the moment when someone with context accepts responsibility for a record that will affect cash, purchasing and later decisions.

AI earns its place before that moment. It sorts the documents, transcribes legible fields, checks relationships and frames specific questions so the owner spends attention on exceptions rather than repetitive entry.

08 Our approach

First, separate evidence from interpretation. Preserve the source, identify the document and keep amount, payment method and channel as distinct facts.

Second, make uncertainty operational. Ask for a better image, a missing method or a new channel mapping instead of guessing a number into place.

Third, reserve posting for the accountable person. AI prepares and checks the draft; a human confirms what the record means and decides when it becomes final.