Document Intelligence and its re-checks

How Achekam's Document Intelligence reads a receipt, invoice or statement with AI, which figures it re-checks afterwards, and where a person takes over.

By Achekam · 11 October 2026 · 5 min read · For Accountants & audit firms in South Africa, Nigeria, Kenya

AI reading tools are impressively fluent. Give one a crumpled till slip and it returns a neat list of items, a total and the VAT, in confident tidy columns. The trouble is that fluency is not accuracy. A model can misread a seven as a one, invent a VAT figure that was never printed, or quietly add up its own list differently from the total at the bottom of the slip.

For a firm processing clients' documents, those errors matter twice: once because the client's books are wrong, and again because a reviewer who trusts the tidy output may never look at the original. The right question about any AI reading tool is not "how good is the AI?" but "what checks happen after it?"

This article explains what Achekam's Document Intelligence does with a receipt, invoice or statement, which figures are re-checked once the AI has finished, and where a person takes over.

Uploading a document

Staff can upload a receipt, an invoice or a statement to Document Intelligence for AI to read. The reading goes through Achekam's single AI engine, which can use Anthropic, OpenAI, Google Gemini, DeepSeek or xAI, using the firm's own API key first and the platform's key only where allowed.

Two practical points before you rely on it. First, it needs an AI key: with no key at all, every AI feature says so and drafts nothing. Second, the output is a reading of the document, not a decision about how to account for it. The person who uploaded it still owns what happens next.

Totals, VAT and dates, re-checked

Once the AI has read the document, Achekam re-checks the totals, the VAT and the dates. These are the three places where reading errors do the most damage:

• Totals, because a line misread or skipped changes the amount posted;
• VAT, because a wrongly read or invented VAT figure changes a VAT return, not just an expense;
• Dates, because a document placed in the wrong month can fall into a closed period or the wrong tax period.

A re-check that fails is a signal to look at the original document, not to trust the AI's version and override the check.

Figures that are not in the transcription

Achekam also flags any figure that does not appear in its own transcription of the document. If the AI's answer contains a number that cannot be found in the text read from the page, the figure is flagged for a person to check.

That is the check aimed at the most dangerous AI failure: inventing something plausible. A VAT amount on a receipt that never showed VAT, a total "corrected" by the model, or a date filled in from context are all examples. A flagged figure needs a person to compare it with the document itself.

The same principle runs through Achekam's other document reading. When supplier invoices and receipts are read through Financial Intake, VAT that is not printed on the document stays empty rather than being worked out. An empty field is honest; a calculated one dressed up as read is not.

Documents that are refused or need extra set-up

Some documents should not go to an AI vendor at all without a deliberate decision. Restricted data classes, such as payroll and bank data, are refused by Document Intelligence unless they are allowed for a particular purpose.

So if a staff member uploads a bank statement or a payslip, expect a refusal unless the firm has allowed that purpose. If it has, the decision and the reason should be in the firm's AI policy. Bank statements have their own import route in Achekam, which checks that the opening balance plus the movements equals the closing balance, and that is usually the better path for them anyway.

Scans and photos in Financial Intake

Document Intelligence is one way to have AI read a document. Financial Intake is the other main route for supplier paperwork. There, Achekam can read supplier invoices and receipts into their fields and propose a journal for each, but it reads scans and photos only when an outside OCR service or an image-reading AI key is set up. Without one, a review item called OCR_ENGINE_NOT_CONFIGURED is raised, so nobody mistakes an unread scan for a processed one. Setting that up is covered in reading invoices with OCR.

From reading to a posting candidate

A reading, however well checked, is not an entry in the books. Journals proposed from documents, bank lines or adjustments are only candidates. Only a person can approve and post one, and AI is refused both steps. Low-confidence fields and possible duplicates wait in the Intake Review Queue for a person to resolve; working the intake review queue explains how to clear it.

That leaves a clear division of labour:

What reviewers should still do

The re-checks narrow the risk; they do not remove the need to look. ACCA's article on examining evidence explains the difference between audit procedures and audit evidence and how to judge which evidence is relevant, a useful frame for deciding how much reliance an AI reading deserves.

For bookkeeping and review work, a short routine:

1. Open the original document for every flagged figure and every failed re-check.
2. Sample unflagged readings each month and compare them with the originals.
3. Watch for patterns, such as one client's receipts, or one supplier's layout, that keep producing flags.
4. Never approve a posting candidate from a reading you have not looked at when it carries a flag.

Pick ten documents your team processed last month, compare each AI reading with the original, and use what you find to decide how much sampling your firm needs.

General guidance only, not advice on your own circumstances. Tax and payroll rules change: check the current position with the authority or your adviser.