AI-supported credit processes are no longer a utopia

Checking customer-related documents is a key task in the lending process. In order to make informed credit decisions and identify risks such as fraud at an early stage, banks and financial service providers must carefully analyze a wide range of documents – from salary statements and bank statements to land register extracts. This step is still often a manual process. This makes it both time-consuming and error-prone.

In recent years, the use of artificial intelligence (AI) has proven to be a promising lever for significantly increasing the efficiency and quality of these document checks. AI-supported applications have long been standard in construction financing, consumer loans, corporate financing and the granting of subsidies.

As part of projects, we work with financial institutions to test and implement specific solutions. Decision-makers in financial institutions should therefore know how they can use AI in a targeted manner to make document processes more intelligent. Our aim is to shorten processing times, identify fraud attempts at an early stage, ensure compliance with regulatory requirements and reduce costs.

The following article provides you with a compact overview of the technological potential and practical approaches for setting up smart document platforms.

Documents to be submitted vary greatly depending on the product type

Credit processes are very different depending on the product type. There are different documents and verification steps for each product type. Many customers start with the automated checking of documents in areas with a high number of inquiries. Construction financing, consumer loans, corporate loans and subsidy applications are predestined for this first step.

Construction financing

Real estate serves as collateral for construction financing, which is why extensive documents on the property and the borrower’s financial situation are checked. Typical documents and checks are

• Land register extract: Provides evidence of the ownership structure and any encumbrances (such as land charges). The bank checks whether the applicant is registered as the owner and what mortgages exist. AI can automatically read the relevant entries (owner, current loans, restrictions) from the usually scanned document using OCR (Optical Character Recognition), for example.
• Parcel map and building plans: These technical documents provide information about the location and condition of the property. Inspection services include checking whether the construction project matches the submitted plans and has been approved by the authorities. AI-supported image recognition can analyze plans and compare them with predefined specifications.
• Building permit and construction specifications: Ensure that the construction project is officially approved and meets certain standards. AI systems can search text passages for key terms (such as conditions or validity periods) to identify deviations or missing content.
• Proof of the borrower’s income: Despite real estate collateral, the bank checks the borrower’s creditworthiness (e. g. payslips, tax assessments) to determine their ability to repay. This verifies whether income and employment are plausible. Automated document analysis extracts salary data and compares it with reference values to identify anomalies.

Consumer loans, such as installment loans

The identity and creditworthiness of the applicant must be carefully checked for unsecured consumer loans such as installment loans. This is especially true if a bank’s clientele can demonstrate historically poor creditworthiness. Fewer documents are required here than for construction financing, but these are essential to rule out fraud and inability to pay:

• Identity documents: A valid identity card or passport is required for Know-Your-Customer (KYC) checks. The bank ensures that the document is genuine and belongs to the applicant. Modern AI-supported solutions use computer vision to check ID documents for security features and use face matching to compare the ID photo with a live photo of the customer. This automated ID check significantly speeds up the identity verification process.
• Proof of income: The most recent payslips and often bank statements are usually requested. They provide proof of income and financial situation. AI systems with OCR and NLP (Natural Language Processing) can read salary documents and differentiate between net and gross income, for example. A plausibility check by the AI compares the stated income with the bank statements (receipt of salary) and issues an alert in the event of deviations or conspicuous patterns (such as atypically high amounts).
• Schufa information/creditworthiness documents: Although credit reports are obtained electronically and not supplied by the customer, they are still part of the document check. AI can help with the analysis of Schufa data, for example by highlighting anomalies such as many recent credit inquiries or existing over-indebtedness signals through anomaly detection.

Corporate loans

In order to assess the solvency of the company and any collateral, lending to companies requires the review of extensive financial and legal documentation. Typical documents and checks includ:

• Annual financial statements (balance sheet and income statement): These provide an insight into the company’s financial position and earning power. The bank analyzes key figures such as equity ratio, cash flow and debt/equity ratio. AI systems can automatically extract these key financial figures from PDF annual financial statements and calculate trends over several years, for example. Anomalies (such as abrupt changes or unusual values compared to the industry) are highlighted by automated analyses.
• Business analyses (BWA) and financial planning: BWAs and planning calculations are used for an up-to-date view of financial development. AI can classify these documents (e.g. distinguish between BWA, balance sheet, planning) and read out the figures they contain. Machine learning can be used to validate forecasts by analysing plan/actual deviations from previous periods – this allows the bank to recognize whether current planning assumptions are realistic.
• Extract from the commercial register: Legal details about the company are checked here (company name, managing directors, capital, liability). An AI-supported text analysis can filter out the relevant facts from digital commercial register entries (such as company age, changes in management) and compare them with the application details.
• Collateral documentation: In the case of secured corporate loans, documents relating to collateral (land charges, assignment of claims, guarantees, etc.) are checked. Similar to construction financing, OCR and NLP are used here to automatically read contracts or expert opinions, for example. An AI tool could, for example, recognize whether a surety agreement contains all the required clauses.

Promotional loans and applications

Promotional loans (e.g. from KfW, EU programs or regional promotional banks) require compliance with special promotional conditions in addition to the normal credit check. Banks often act as intermediaries here and must check both the customer’s application documents and the approval notices from the funding agency:

• Funding application and decision: The applicant submits forms setting out the funding requirements (e.g. intended use, project description). Following a positive decision, the development bank issues an approval notice. KI can check whether all the required information is included in the application and whether it is consistent with the approval notice. For example, NLP can be used to read out the approved funding amount and conditions from the approval notice and automatically compare them with the loan application.
• Supporting documents: Certain supporting documents (implementation reports, invoices, proof of use) often have to be submitted in order to prove that the funds have been used as intended. AI-supported systems can automatically record submitted invoices and compare them with the approved budget. In addition, anomaly detection can provide indications of irregularities – for example, if a beneficiary repeatedly submits the same invoice documents (possible indicator of fraud).
• Communication with funding institutions: Correspondence or emails with the funding body (e.g. queries) can also be documented. AI can categorize the content of these letters (e.g. additional requests for documents, content-related queries) and thus help the processor to ensure that nothing is overlooked.

Documents may differ – but usable AI technology is the same for almost all content

One thing is clear: even for the most common banking services, the documents to be processed and the depth of the required check vary greatly. Basically, the checks can be divided into three types:

• Digitization: A data record is created from a PDF or a scan
• Validation of details: The content of information within the document is validated against previous information or legal requirements
• Fraud prevention: A machine learning model checks whether changes have been made to a document

Depending on the depth of the check, a mix of different AI methods promises a good check result, as the following graphic illustrates:

The added value of AI-supported credit processes and KYC activities is illustrated by four case studies, which only represent excerpts of innovative pilot approaches in the areas described above.

A coherent platform remains the be-all and end-all

Do you want to use AI to move step by step towards the dark processing of documents?

The truth is: many banks lack direct customer contact, for example in the form of a credit platform, in order to be able to collect and check documents directly from the customer. Savings banks, specialist institutions and cooperative banks in particular must therefore make a conscious decision as to whether they want to act independently in order to leverage efficiencies in the credit route.