This does not mean that every document created using generative tools is fake, nor that a fake document has necessarily been produced using AI.<\/p>\n\n\n\n
For those working in the credit sector, the question remains a very practical one: does what we read correspond to a real person, a real business and a real transaction? A neat letterhead and a VAT number entered in the correct place are not enough to answer this question.<\/p>\n\n\n\n
CASE STUDY:<\/strong><\/h5>\n\n\n\nTwo quotes, two suppliers who didn\u2019t add up<\/h3>\n\n\n\n
During an anti-fraud investigation into an application for investment funding, we examined two quotations for renovation work. Preliminary checks on the applicant company had not revealed any particular issues. The documents also appeared to be complete: descriptions of the work, amounts, supplier details and contact information.<\/p>\n\n\n\n
The problem arose when we stopped just looking at the quotes and started checking who had issued them<\/strong>.<\/p>\n\n\n\nThe VAT registration number did not exist<\/h3>\n\n\n\n
The VAT number shown on the first quotation turned out not to exist. <\/p>\n\n\n\n
It is therefore not a case of a business that subsequently ceased trading, but rather a tax identification number that does not appear to have been assigned to the person in question. This is a significant inconsistency which seriously calls into question the document\u2019s reliability.<\/strong> Whilst this does not, in itself, prove that the document was deliberately falsified, it does make it necessary to ascertain who actually issued the quotation and why that number was used. This can be verified by means of the Revenue Agency service<\/a>.<\/p>\n\n\n\nA company name, a different legal form and conflicting contact details<\/h3>\n\n\n\n
In the second quotation, the VAT number was linked to a company with a different legal form from that stated in the header. The declared business activity did not appear to be consistent with the services described. The contact details also raised doubts: the email address referred to a different business, and the certified email address did not match that of the identified company.<\/p>\n\n\n\n
A discrepancy in the ATECO code or contact details, taken on its own, may have a legitimate explanation. In this case, however, there was more than one inconsistency. By cross-referencing Company Register<\/a>, INI-PEC<\/a> and with the tax figures, the picture became hard to ignore.<\/p>\n\n\n\nThe crucial discrepancies lay in the identity of the person who claimed to have issued it.<\/strong><\/p>\n\n\n\nWe therefore asked the artificial intelligence to check the quotes<\/h3>\n\n\n\n
At that point, we tried running the documents through an artificial intelligence tool. The request was specific: to check whether they were forged and to produce a report setting out the findings.<\/p>\n\n\n\n
The system focused primarily on the technical and financial aspects. It analysed the amounts, the descriptions of the work and some missing information. He identified issues that required clarification and did not certify the estimates as accurate.<\/strong><\/p>\n\n\n\n![\"\"](\"https:\/\/senapa.eu\/wp-content\/uploads\/2026\/10\/Screenshot-2026-10-20-160341.png\")
<\/figure>\n\n\n\nIn the report, he advised against approving or paying advances without further clarification. And, importantly, he stated that he had carried out a documentary and arithmetic analysis of the PDF, not a company or tax audit<\/strong>.<\/p>\n\n\n\n![\"\"](\"https:\/\/senapa.eu\/wp-content\/uploads\/2026\/10\/Screenshot-2026-10-20-160250-1.png\")
<\/figure>\n\n\n\nIn this regard, the system has been corrected: it did not conceal the limitations of its own work. However, it also failed to identify the anomalies which, in our view, were crucial: the non-existent VAT number, the conflicting company details, and the contact details that needed clarifying.<\/p>\n\n\n\n
The report wasn\u2019t necessarily wrong. It was simply insufficient for the verification we wanted to carry out.<\/strong><\/p>\n\n\n\nAnd that is an important distinction. An answer may be accurate within the scope that the system has examined, but it can become dangerous if someone interprets it as a comprehensive judgement on the reliability of the practice. Our experiment does not prove that AI is incapable of carrying out these checks. It demonstrates that we must ensure it has carried them out, using the appropriate sources and tools \u2013 in fact, taking longer than a human would in many cases. <\/p>\n\n\n\n
An incomplete conclusion can lead to a wrong decision<\/h3>\n\n\n\n
Let us imagine that this report is included in the preliminary investigation without any further checks. The loan could be assessed on the basis of documents whose origin has not been clarified. The problem might only come to light after the loan has been disbursed, perhaps when the loan becomes non-performing.<\/p>\n\n\n\n
If a guarantor is involved, the consequences will depend on the terms of the guarantee and the applicable regulations: a discrepancy in the documentation does not automatically result in the loss of cover. The fact remains, however, that the bank, the guarantor and the applicant may find themselves having to deal with a dispute that arose much earlier, at the point when someone deemed the checks to be sufficient, when in fact they were not.<\/p>\n\n\n\n
There is also the opposite risk: classifying a legitimate transaction as suspicious, thereby blocking a customer without sufficient evidence. Automating a check does not remove the responsibility for the decision taken on the basis of that check.<\/strong><\/p>\n\n\n\nDoes a fraud examiner now have to do the work of an IT specialist as well?<\/h2>\n\n\n\n
Until a few years ago, in many organisations, those responsible for fraud examination could turn to the IT department for technical support. The analyst would reconstruct transactions, check documents and link individuals and companies; IT specialists would work on the tools and infrastructure. It wasn\u2019t a complete separation \u2014 computer forensics has been around for some time \u2014 but it worked well for many investigations.<\/p>\n\n\n\n
\nToday, we are asking technology to step in precisely where analysis is concerned. It analyses financial statements, interprets documents, collates information and suggests conclusions. And so a question arises: Do those who designed the control really understand the type of fraud that the control is supposed to detect?<\/strong><\/p>\n<\/blockquote>\n\n\n\nA skilled developer can create a technically flawless system without understanding the investigative implications of a deactivated VAT number (there are even more complex examples than this). An experienced fraud examiner may spot that very same anomaly at first glance, but may not know how an AI agent linked to external databases works, or what authorisations it holds (or which of these details are actually sourced from open and verifiable sources).<\/p>\n\n\n\n
Neither of them is incompetent. They have different areas of expertise.<\/strong> The danger is in thinking that, by incorporating them into a single instrument, one can also eliminate the people who possess them.<\/p>\n\n\n\nA lightning-fast junior staff member<\/h3>\n\n\n\n
For certain tasks, AI is a huge help. It extracts data, organises information and prepares comparisons that would otherwise take hours. But when we ask it to make an investigative judgement, to me it sometimes feels like a very fast junior colleague: it produces a lot, but I have to check what it has understood, what it has verified and what it has left out.<\/p>\n\n\n\n
Furthermore, a generative system does not necessarily learn permanently just because we correct one of its responses. Next time, we need to know how it has been configured and which checks it will actually carry out.<\/p>\n\n\n\n
\nThe time saved in preparing the analysis may end up being spent on reviewing it. This does not always happen, but it is naive to assume that every hour saved through automation is an hour of work eliminated.<\/p>\n<\/blockquote>\n\n\n\n
The butterfly effect of an incorrect check<\/h3>\n\n\n\n
With AI agents, the problem can become more serious. One system misreads a piece of company data; another uses that data to prepare a report; a third considers it reliable and includes it in a risk assessment. If no one checks the first step, the error is passed on along with the information.<\/p>\n\n\n\n
That\u2019s what worries me about the knock-on effect: not just the initial mistake, but how many decisions it might influence before anyone realises it. OWASP<\/a> <\/mark>It highlights, amongst the risks associated with AI systems, excessive autonomy, overly broad authorisations and the need for human oversight of the most sensitive operations.<\/p>\n\n\n\nGreater productivity, fewer staff. And coffee in bed.<\/h3>\n\n\n\n
Then there is the issue of work, which, in my view, is still not discussed enough. Artificial intelligence is often presented as a promise: faster results, more streamlined processes, and perhaps fewer staff.<\/p>\n\n\n\n
\nMeanwhile, fraud examiners should keep up to date with investigative techniques, be familiar with IT systems, have a basic understanding of cybersecurity, oversee the use of AI and, of course, work even faster.<\/p>\n\n\n\n
The only thing left is for him to be asked to bring coffee to bed in the morning as well.<\/strong><\/p>\n<\/blockquote>\n\n\n\nThat was a deliberate joke, but the issue is a serious one. We cannot keep adding new skills and responsibilities to a professional role whilst, at the same time, claiming that their job has become easier.<\/p>\n\n\n\n
Article 4 of the AI Act<\/a> It calls on suppliers and professional users to take steps to ensure that the staff involved have a sufficient level of AI literacy, taking into account their knowledge, experience and working context. It does not require everyone to become a programmer. However, it points out that using these tools effectively requires training.<\/p>\n\n\n\nThis item and a loan cannot have the same margin of error<\/h2>\n\n\n\n
AI could certainly write an article like this, provided it is guided and reviewed by someone who knows the subject. But what harm can a mistake in an article do? Not always just a little, let\u2019s be clear: even information published incorrectly can harm people and businesses. However, we have the opportunity to check it before publication, take action and correct it.<\/p>\n\n\n\n
If, on the other hand, the error relates to an anti-fraud check, we may only discover it once the loan has already been paid out, a legitimate customer has been turned down, or a report has led to consequences. Correcting the report does not always allow us to undo what has happened.<\/p>\n\n\n\n
That\u2019s why it\u2019s not enough for me to know how much time AI saves us. I also want to know how much a faulty check might cost.<\/strong><\/p>\n\n\n\nWho monitors the monitor?<\/h2>\n\n\n\n
In our test, the artificial intelligence had correctly acknowledged its own limitations. The risk would have been to ignore them.<\/p>\n\n\n\n
We are investing heavily in making checks faster. Are we investing just as much in the skills of those responsible for carrying them out?<\/p>\n\n\n\n
Because a report can be corrected. Funding disbursed on the basis of incorrect checks, however, cannot always be rectified.<\/p>\n\n\n\n
\nIf artificial intelligence is used to detect fraud, we must ensure that those who oversee it are still able to recognise it.<\/strong><\/p>\n<\/blockquote>\n\n\n\n
\n\n\n\nLearn more about AI and fraud prevention with SENAPA Academy<\/h2>\n\n\n\n
In SENAPA Academy 2026 Training Catalogue<\/a> We have dedicated a specific section to Data Analysis, AI and Technology, featuring modules such as AI & Anti-Fraud, Prompting for Anti-Fraud Activities, Deepfakes and Synthetic Content, and Anti-Fraud Data Analytics. These courses are designed to explore the use of technological tools in auditing in greater depth and, above all, to understand their limitations and risks.<\/p>\n\n\n\nSources and references<\/h2>\n\n\n\n\n- EBA \u2013 Opinion on the misuse of innovative products and AML\/CFT risks.<\/a><\/li>\n\n\n\n
- Revenue Agency \u2013 VAT number verification<\/a>; Company Register<\/a>; INI-PEC<\/a>.<\/li>\n\n\n\n
- OWASP \u2013 AI Agent Security Cheat Sheet.<\/a><\/li>\n\n\n\n
- European Commission \u2013 AI Act, Article 4.<\/a><\/li>\n\n\n\n
- SENAPA Academy \u2013 Training Catalogue 2026.<\/a><\/li>\n<\/ul>\n\n\n\n
<\/p>","protected":false},"excerpt":{"rendered":"
Due preventivi sottoposti all’intelligenza artificiale: un report dettagliato, verifiche mancanti e rischi reali per chi deve prevenire le frodi creditizie.<\/p>","protected":false},"author":37,"featured_media":9507,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_uf_show_specific_survey":0,"_uf_disable_surveys":false,"footnotes":""},"categories":[109],"tags":[],"class_list":["post-9488","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-senapa-news"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t
The VAT registration number did not exist<\/h3>\n\n\n\n
The VAT number shown on the first quotation turned out not to exist. <\/p>\n\n\n\n
It is therefore not a case of a business that subsequently ceased trading, but rather a tax identification number that does not appear to have been assigned to the person in question. This is a significant inconsistency which seriously calls into question the document\u2019s reliability.<\/strong> Whilst this does not, in itself, prove that the document was deliberately falsified, it does make it necessary to ascertain who actually issued the quotation and why that number was used. This can be verified by means of the Revenue Agency service<\/a>.<\/p>\n\n\n\n In the second quotation, the VAT number was linked to a company with a different legal form from that stated in the header. The declared business activity did not appear to be consistent with the services described. The contact details also raised doubts: the email address referred to a different business, and the certified email address did not match that of the identified company.<\/p>\n\n\n\n A discrepancy in the ATECO code or contact details, taken on its own, may have a legitimate explanation. In this case, however, there was more than one inconsistency. By cross-referencing Company Register<\/a>, INI-PEC<\/a> and with the tax figures, the picture became hard to ignore.<\/p>\n\n\n\n The crucial discrepancies lay in the identity of the person who claimed to have issued it.<\/strong><\/p>\n\n\n\n At that point, we tried running the documents through an artificial intelligence tool. The request was specific: to check whether they were forged and to produce a report setting out the findings.<\/p>\n\n\n\n The system focused primarily on the technical and financial aspects. It analysed the amounts, the descriptions of the work and some missing information. He identified issues that required clarification and did not certify the estimates as accurate.<\/strong><\/p>\n\n\n\n In the report, he advised against approving or paying advances without further clarification. And, importantly, he stated that he had carried out a documentary and arithmetic analysis of the PDF, not a company or tax audit<\/strong>.<\/p>\n\n\n\n In this regard, the system has been corrected: it did not conceal the limitations of its own work. However, it also failed to identify the anomalies which, in our view, were crucial: the non-existent VAT number, the conflicting company details, and the contact details that needed clarifying.<\/p>\n\n\n\n The report wasn\u2019t necessarily wrong. It was simply insufficient for the verification we wanted to carry out.<\/strong><\/p>\n\n\n\n And that is an important distinction. An answer may be accurate within the scope that the system has examined, but it can become dangerous if someone interprets it as a comprehensive judgement on the reliability of the practice. Our experiment does not prove that AI is incapable of carrying out these checks. It demonstrates that we must ensure it has carried them out, using the appropriate sources and tools \u2013 in fact, taking longer than a human would in many cases. <\/p>\n\n\n\n Let us imagine that this report is included in the preliminary investigation without any further checks. The loan could be assessed on the basis of documents whose origin has not been clarified. The problem might only come to light after the loan has been disbursed, perhaps when the loan becomes non-performing.<\/p>\n\n\n\n If a guarantor is involved, the consequences will depend on the terms of the guarantee and the applicable regulations: a discrepancy in the documentation does not automatically result in the loss of cover. The fact remains, however, that the bank, the guarantor and the applicant may find themselves having to deal with a dispute that arose much earlier, at the point when someone deemed the checks to be sufficient, when in fact they were not.<\/p>\n\n\n\n There is also the opposite risk: classifying a legitimate transaction as suspicious, thereby blocking a customer without sufficient evidence. Automating a check does not remove the responsibility for the decision taken on the basis of that check.<\/strong><\/p>\n\n\n\n Until a few years ago, in many organisations, those responsible for fraud examination could turn to the IT department for technical support. The analyst would reconstruct transactions, check documents and link individuals and companies; IT specialists would work on the tools and infrastructure. It wasn\u2019t a complete separation \u2014 computer forensics has been around for some time \u2014 but it worked well for many investigations.<\/p>\n\n\n\n Today, we are asking technology to step in precisely where analysis is concerned. It analyses financial statements, interprets documents, collates information and suggests conclusions. And so a question arises: Do those who designed the control really understand the type of fraud that the control is supposed to detect?<\/strong><\/p>\n<\/blockquote>\n\n\n\n A skilled developer can create a technically flawless system without understanding the investigative implications of a deactivated VAT number (there are even more complex examples than this). An experienced fraud examiner may spot that very same anomaly at first glance, but may not know how an AI agent linked to external databases works, or what authorisations it holds (or which of these details are actually sourced from open and verifiable sources).<\/p>\n\n\n\n Neither of them is incompetent. They have different areas of expertise.<\/strong> The danger is in thinking that, by incorporating them into a single instrument, one can also eliminate the people who possess them.<\/p>\n\n\n\n For certain tasks, AI is a huge help. It extracts data, organises information and prepares comparisons that would otherwise take hours. But when we ask it to make an investigative judgement, to me it sometimes feels like a very fast junior colleague: it produces a lot, but I have to check what it has understood, what it has verified and what it has left out.<\/p>\n\n\n\n Furthermore, a generative system does not necessarily learn permanently just because we correct one of its responses. Next time, we need to know how it has been configured and which checks it will actually carry out.<\/p>\n\n\n\n The time saved in preparing the analysis may end up being spent on reviewing it. This does not always happen, but it is naive to assume that every hour saved through automation is an hour of work eliminated.<\/p>\n<\/blockquote>\n\n\n\n With AI agents, the problem can become more serious. One system misreads a piece of company data; another uses that data to prepare a report; a third considers it reliable and includes it in a risk assessment. If no one checks the first step, the error is passed on along with the information.<\/p>\n\n\n\n That\u2019s what worries me about the knock-on effect: not just the initial mistake, but how many decisions it might influence before anyone realises it. OWASP<\/a> <\/mark>It highlights, amongst the risks associated with AI systems, excessive autonomy, overly broad authorisations and the need for human oversight of the most sensitive operations.<\/p>\n\n\n\n Then there is the issue of work, which, in my view, is still not discussed enough. Artificial intelligence is often presented as a promise: faster results, more streamlined processes, and perhaps fewer staff.<\/p>\n\n\n\n Meanwhile, fraud examiners should keep up to date with investigative techniques, be familiar with IT systems, have a basic understanding of cybersecurity, oversee the use of AI and, of course, work even faster.<\/p>\n\n\n\n The only thing left is for him to be asked to bring coffee to bed in the morning as well.<\/strong><\/p>\n<\/blockquote>\n\n\n\n That was a deliberate joke, but the issue is a serious one. We cannot keep adding new skills and responsibilities to a professional role whilst, at the same time, claiming that their job has become easier.<\/p>\n\n\n\n Article 4 of the AI Act<\/a> It calls on suppliers and professional users to take steps to ensure that the staff involved have a sufficient level of AI literacy, taking into account their knowledge, experience and working context. It does not require everyone to become a programmer. However, it points out that using these tools effectively requires training.<\/p>\n\n\n\n AI could certainly write an article like this, provided it is guided and reviewed by someone who knows the subject. But what harm can a mistake in an article do? Not always just a little, let\u2019s be clear: even information published incorrectly can harm people and businesses. However, we have the opportunity to check it before publication, take action and correct it.<\/p>\n\n\n\n If, on the other hand, the error relates to an anti-fraud check, we may only discover it once the loan has already been paid out, a legitimate customer has been turned down, or a report has led to consequences. Correcting the report does not always allow us to undo what has happened.<\/p>\n\n\n\n That\u2019s why it\u2019s not enough for me to know how much time AI saves us. I also want to know how much a faulty check might cost.<\/strong><\/p>\n\n\n\n In our test, the artificial intelligence had correctly acknowledged its own limitations. The risk would have been to ignore them.<\/p>\n\n\n\n We are investing heavily in making checks faster. Are we investing just as much in the skills of those responsible for carrying them out?<\/p>\n\n\n\n Because a report can be corrected. Funding disbursed on the basis of incorrect checks, however, cannot always be rectified.<\/p>\n\n\n\n If artificial intelligence is used to detect fraud, we must ensure that those who oversee it are still able to recognise it.<\/strong><\/p>\n<\/blockquote>\n\n\n\n In SENAPA Academy 2026 Training Catalogue<\/a> We have dedicated a specific section to Data Analysis, AI and Technology, featuring modules such as AI & Anti-Fraud, Prompting for Anti-Fraud Activities, Deepfakes and Synthetic Content, and Anti-Fraud Data Analytics. These courses are designed to explore the use of technological tools in auditing in greater depth and, above all, to understand their limitations and risks.<\/p>\n\n\n\n <\/p>","protected":false},"excerpt":{"rendered":" Due preventivi sottoposti all’intelligenza artificiale: un report dettagliato, verifiche mancanti e rischi reali per chi deve prevenire le frodi creditizie.<\/p>","protected":false},"author":37,"featured_media":9507,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_uf_show_specific_survey":0,"_uf_disable_surveys":false,"footnotes":""},"categories":[109],"tags":[],"class_list":["post-9488","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-senapa-news"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\tA company name, a different legal form and conflicting contact details<\/h3>\n\n\n\n
We therefore asked the artificial intelligence to check the quotes<\/h3>\n\n\n\n
<\/figure>\n\n\n\n<\/figure>\n\n\n\nAn incomplete conclusion can lead to a wrong decision<\/h3>\n\n\n\n
Does a fraud examiner now have to do the work of an IT specialist as well?<\/h2>\n\n\n\n
\n
A lightning-fast junior staff member<\/h3>\n\n\n\n
\n
The butterfly effect of an incorrect check<\/h3>\n\n\n\n
Greater productivity, fewer staff. And coffee in bed.<\/h3>\n\n\n\n
\n
This item and a loan cannot have the same margin of error<\/h2>\n\n\n\n
Who monitors the monitor?<\/h2>\n\n\n\n
\n
\n\n\n\nLearn more about AI and fraud prevention with SENAPA Academy<\/h2>\n\n\n\n
Sources and references<\/h2>\n\n\n\n
\n