AI Scanning an Invoice

This new issue, dedicated to ChatGPT and Sermo (my programming gateway to ChatGPT, Perplexity, and Claude 3.5 Sonnet), will showcase the ability of LLM-based AIs (Large Language Models) to extract details from an invoice.

To demonstrate this, I'll be using a real invoice—there's really no secret to reveal or hide here.

I'll be employing a new feature of Sermo: the ability to analyze an invoice.

First, here is the scan of the invoice:

Here is the call to Sermo for detecting the information contained in the invoice:

$sermo = new Sermo();

$t1 = microtime(true);
$a = $sermo->analyzeInvoice( 'https://www.trql.fm/mazout.jpg' );
$t2 = microtime(true);

if ( preg_match( '/```xml(?P<code>.*)```/si',$a,$matches ) )
{
    $xml = $matches['code'];
    echo "<pre>",htmlentities( $xml ),"</pre>";
}

echo "<p>Performance: ",number_format( $t2 - $t1,5,',','.' )," secs to examine the invoice</p>";        

As you can see, a single line of code is enough to ask Sermo to scan the invoice. Here is the result:

<invoice>
    <recipient>
        <name>LATO SENSU MANAGEMENT</name>
        <vatNumber>BE 0878127142</vatNumber>
    </recipient>
    <issuer>
        <name>BERTRAND Combustibles</name>
        <address>
            <street>Rue du Chêne</street>
            <number>1</number>
            <postalCode>5070</postalCode>
            <locality>Fosses-la-Ville</locality>
        </address>
        <phone>0483/037209</phone>
        <email>[email protected]</email>
        <website>www.bertrandcombustibles.be</website>
        <vatNumber>BE 0742.406.430</vatNumber>
        <bankAccount>BE02 7326 6822 2240</bankAccount>
    </issuer>
    <invoiceDetails>
        <lineOfDetail>Gas oil Diesel (EXTRA) - 10PPM - 1.000 L - 748,35 EUR - 21,00 EUR - 157,15 EUR</lineOfDetail>
    </invoiceDetails>
    <invoiceNo>V1259084</invoiceNo>
    <invoiceDate>20240726</invoiceDate>
    <totalHTVA>748,35</totalHTVA>
    <totalTVA>157,15</totalTVA>
    <totalTVAC>905,50</totalTVAC>
    <prepayment></prepayment>
    <totalPaymentDue>905,50</totalPaymentDue>
    <paymentDueDate>20240726</paymentDueDate>
    <currency>EUR</currency>
    <expenseType>Heating oil</expenseType>
    <expenseAccount>61232</expenseAccount>
</invoice>
Performance: 9,68427 secs to examine the invoice        

Compare the invoice scan and the rendered result for yourself!

Since I have tailored Sermo's detection to align with Belgian accounting, I was also able to ask it to categorize the expense (currently, the categories are not predetermined) and identify the expense account that should be affected in the accounting. However, I have not yet succeeded in accurately determining this expense account or in allocating the total amount between the expense account and the VAT account. Nor have I been able to automatically determine that the credit should be made to the accounts payable (account 401). Finally, if the invoice has a total balance of 0, indicating it has been paid, I could also determine that there should be a debit to account 401 (accounts payable) and a credit to account 512 (bank). These are future developments planned for Sermo.

Conclusions

Originally just a simple algorithm that determines the next word, LLMs now offer astonishing possibilities that are chilling in their implications for the workforce and, consequently, our society. What I have just demonstrated here is that the average person might soon be able to do without an accountant. This is not science fiction; it is reality! It's already here.

In an upcoming article, I will share the new Sermo code with you.

Previous article - Next article