Automated Credit Assessment: How It Works in Practice
Automated credit assessment in leasing and finance: what AI can do, what people must decide, and what GDPR, the Danish AML Act and the AI Act require.
Table of contents
How can AI shorten the processing time in a credit assessment?
AI shortens processing time by taking the preparatory work: reading accounts and documents, extracting key figures, gathering information from the registers and systems you already use, and laying out a clear basis for the person who is to make the decision. The decisive decision to grant or refuse credit should stay with a human, especially when the customer is a private individual. How much time you save depends on how much is manual today, and you can measure that before you build anything.
TL;DR:
- Time is a competitive parameter. A leasing customer who waits days for an answer has often already asked someone else.
- Separate preparation from decision. AI is strong at reading, summarising and sorting. The credit decision itself about a natural person is a high-risk area under the EU AI Act and requires human control.
- Start with document reading and case preparation. It gives time back without moving responsibility for the decision.
- Work out your own numbers. The number of cases, the time per case and the share lost along the way decide whether it is worth building.
In the finance world, response time is a competitive parameter. When a business customer is at the machine dealer and wants to lease an excavator, or a private customer needs a car, they now expect an answer. The Danish Financial Supervisory Authority's mapping from 2024 shows that a large share of Danish financial companies already use AI somewhere in their operations, and that almost all expect it to become more important. So the question is not whether you should look at it, but where to begin and how to do it responsibly.
What is AI-driven credit assessment?
The term covers two very different things, and it is a good idea to keep them apart:
1. Document reading and case preparation. AI (OCR and language models) reads accounts, payslips, bank statements and applications, extracts key figures and information, cross-checks against the sources you have access to, and writes a short summary of the case with any gaps marked. It is work that a credit employee today spends a large part of their time on.
2. Scoring and decision support. A machine-learning model calculates a risk score from many data points and gives a recommendation. That is a different and more demanding discipline, with model validation, explainability and regulation. Many financial companies have used statistical scoring models for years. What is new is that more data and more advanced models can be used.
The first category is almost always the right place to start. It removes the tedious work without changing who makes the decision.
Why is the financial sector under pressure?
Customers' expectations have changed. Customers who have got answers in an app within minutes do not accept waiting several days for a leasing offer.
The rules have been tightened. The Danish AML Act, GDPR and the EU AI Act set requirements for documentation, explainability and control. That makes manual processes expensive, and automated processes require proper documentation.
Competition comes from several directions. Fintech providers and digital banks have set a new standard for speed, while the established houses have the advantage of trust, experience and a broad customer base. The question is whether you can combine the two.
What does manual credit assessment really cost?
There is no single figure that applies to everyone. But credit employees typically spend a large part of their time on something other than the assessment itself: collecting documents, data entry, cross-checking against the CVR and registers, chasing missing attachments and internal approval rounds. This is how you can measure it:
| Measurement | What it tells you |
|---|---|
| Time from application to answer | How long does the customer wait? |
| Employee time per case | How many hours go to preparation, and how many to the assessment itself? |
| Share of cases missing information at first review | How much time goes to chasing attachments? |
| Share of applicants who drop out along the way | What does waiting cost in lost contracts? |
| Time on KYC and document checks | What could a machine take first? |
| Errors and reopened cases | How often does something have to be redone? |
Measure them for a month. Then you have a basis for calculating what a solution must save for it to be worth building. That is worth more than any average figure from a report.
Use case 1: Document reading and case preparation
This is where most financial companies can recover the most time with the least risk. The flow can look like this:
- Documents are received. The customer or the seller uploads accounts, a payslip, a bank statement or an application. The system confirms receipt and checks that the documents are legible and complete.
- Information is extracted. AI reads the documents and finds key figures, ownership, amounts and dates, and looks them up in the sources you already use, e.g. the CVR.
- Gaps are marked. If an attachment is missing or a figure does not match, it is flagged instead of being discovered later.
- The case is summarised. The credit employee gets an overview with the sources next to each figure, so it is quick to check.
- A person assesses. The credit employee approves, rejects or asks for more information.
Case handling of this type is one of the tasks we can build: documents read, information gathered and gaps marked for your assessment. The system may read the document and fill in the fields, but it asks first about everything that decides a case. See the example on the use-cases page.
It is worth building when a large part of the work is reading something through and moving information from it. It is not worth building if every case is entirely its own, and what takes the time is the assessment rather than the reading.
Use case 2: KYC and document collection
The AML Act requires you to know your customers: identity, beneficial owners and in some cases source-of-funds information. For the credit department it is time-consuming, and for the customer it is frustrating to have to send the same things several times.
AI can help with the practical part of the process:
- Ask the customer for the missing documents in a conversation, and explain why.
- Check that the document is legible and of the right type.
- Register it in the case and keep a log of what has been received and when.
- Assemble the case for the compliance employee.
What AI should not do alone is decide whether a customer is in order. Screening against lists of politically exposed persons and sanctions is carried out with the providers and processes your compliance department has approved, and the decision lies with an employee. The customer must always know that they are talking to an AI system, and that a person can be brought in.
Use case 3: Lease expiry and retention
Leasing lives on re-leasing. When a contract approaches expiry, there is a window in which the customer is still open to a conversation. Most companies send a standard email, which the customer ignores.
A better solution is to let a system find the contracts that expire within the window you choose, sort them by a rule you can read, and write a personal draft to each customer based on what you know: model, mileage, service history and payment record. The salesperson reviews the list and sends. Nothing goes out without a yes.
This is what we call resuming a conversation you have already had. It is not built to call bought lists. Read more about that example here, and about how it can look in the car trade in our post on AI for car sales and leasing.
Use case 4: Residual value and pricing
In leasing, the residual value is what decides whether the business makes money. If you guess too high on what a car will be worth in three years, you lose on the residual value. If you guess too low, you lose the contract to a competitor.
AI-based models can use data from the used-car market, auctions, changes in taxes, interest rates and model development to suggest a range for the expected value. Electric cars are an area of particular uncertainty, because technology and prices change quickly.
Remember two things: a model is only as good as the data it is trained on, and a range is not a guarantee. Use the model as decision support, and let an experienced employee have the final word, especially for new models and unusual conditions.
What should AI not do alone?
- Refuse an application. A refusal of a natural person should be made by a human, with a justification the customer can understand.
- Make a final credit decision about a private individual without human control. See the section on rules below.
- Screen for money laundering without an approved process.
- Call customers you have no relationship with.
- Change terms or prices without approval.
Real-time fraud detection is a specialist discipline of its own, typically held by banks and payment providers with special data and models. It is not something we build.
Which rules apply?
This is not legal advice, but these are the rules you need to have under control before you build. Bring your compliance department and possibly a lawyer in from the start.
GDPR Article 22. As a starting point, you have the right not to be subject to a decision based solely on automated processing, when it has legal effects or similarly significantly affects the person. There are exceptions, and requirements apply including, among others, human intervention, the opportunity to express your view, and the opportunity to contest the decision.
The EU AI Act. Systems used to assess the creditworthiness of natural persons or establish their credit score are classified as high-risk, with the exception of systems used to detect financial fraud. For high-risk systems, requirements apply to risk management, data quality, documentation, logging, transparency and human oversight. The timing of when the requirements apply has been postponed, so check the current timetable before you plan to a specific date. The regulation also contains transparency requirements when people communicate directly with an AI system.
The AML Act. Customer due diligence, risk assessment, screening and record-keeping. Retention requirements can take precedence over the customer's wish for deletion.
GDPR more generally. A clear basis for processing, data minimisation, information to the data subjects and a data processing agreement with every provider that processes data for you. Always ask where data is processed, and on what basis it is transferred outside the EU/EEA if it is.
The Danish FSA. The Danish Financial Supervisory Authority has published mappings and guidance on good practice when using AI in the financial sector, with a focus on governance, responsibility and documentation. Read them before you start (see the sources at the bottom).
The practical consequence: always build the solution so that it is clear what the system has done, what it based that on, and who has decided. That is both good practice and the precondition for being able to explain a decision.
How to get started
- Measure the current state. Use the measurements above for a month.
- Choose one type of case. For example, standard corporate leasing with few and well-known documents.
- Look at the data. Which documents come in, in what format, and how uniform are they?
- Define what the system may do. Read, fill in, mark gaps, suggest. Not decide.
- Run in parallel. Let the system prepare the cases while the employees work as they do today, and compare. Where does the system get it wrong?
- Go live on the one case type, measure, and then build on.
How long this takes depends on your systems and the quality of your documents. We can only say that once we have seen them.
How does it change the role of the credit manager?
The purpose is not to replace the credit manager, but to move their time. The traditional day is about data collection, document checks, manual look-ups and writing the same kind of assessment again and again. A credit manager with a good preparation system spends more time on cases that require judgement, on the customer relationship, on the portfolio and on complex negotiations. That is the work that requires a human, and it is what creates value and keeps customers.
How do you calculate the gain?
Use a simple model with your own figures:
- Number of cases per month × time saved per case × hourly rate = direct saving.
- Number of cases that previously dropped out because of waiting time × margin per case = revenue that can be saved.
- Fewer errors and reopened cases = time and risk avoided.
Set the cost of building and running the solution against that. The price depends on how many systems need to be connected, how uniform the documents are, and how much the solution is to do on its own. Agent360 prices according to the task.
How Agent360 can help
Agent360 does not supply a finished credit assessment system, and we do not make credit decisions. We build AI solutions in your own systems and call it Your AI Department:
- We think it through with you. We map where the time goes in your case handling, and where a solution gives the most with the least risk.
- We build it with you. On top of the systems you have, e.g. case system, CRM, mail archive and document archive. You decide what the system may do on its own. Everything else waits for approval.
- We keep track of it. Tested, reviewed and in line with your rules for data and AI.
If you are a director or manager in a financial company and want to understand for yourself what AI can and cannot do, we have a personal AI mentoring programme, where you learn on your own tasks, with a mentor at your side. You do not need to be able to code.
See more examples on the use-cases page, e.g. data transfer between systems.
Frequently asked questions about AI in credit assessment
Q: What is AI-driven credit assessment? It can mean two things: AI that reads and summarises documents and prepares a case, and machine-learning models that calculate a risk score. The first is the most obvious place to start. The second requires model validation, explainability and, for natural persons, compliance with the rules for high-risk AI.
Q: Is it lawful to use AI for credit assessment in Denmark? Yes, but with requirements. GDPR Article 22 restricts fully automated decisions with significant effect on a person, and the EU AI Act classifies credit assessment of natural persons as high-risk. The AML Act sets requirements for customer due diligence. Bring your compliance department and a lawyer in from the start.
Q: What does it cost to implement? It depends on your systems, your documents and how much the solution is to do on its own. Agent360 prices according to the task. In the 20-minute free consultation we can say roughly what it requires, and whether it is worth building.
Q: Does AI replace credit managers? No. AI can prepare the cases, but responsibility for the decision and the assessment of complex cases lies with the employees. The gain is that they spend more time on the cases that need them.
Q: What happens when the AI gets it wrong? The system must be built for that. That means sources and justifications are visible, a human approves, and errors can be traced and corrected. During the pilot phase you should compare the system's suggestions with the employees' assessments and study the differences.
Q: What is Open Banking, and should we use it? Open Banking is the framework that, through the Payment Services Directive (PSD2), allows customers to give authorised third parties access to their account data. It can provide fresher data than old annual accounts, but it requires the customer's consent, an authorised provider and a clear framework for what the data is used for. It is a choice you make once you have mapped the need.
Q: Can you build on top of our case system? We build on the systems you have. Which connections are possible depends on the system and the provider. We contact your providers and help with access, and we will tell you if it is not the right route.
Q: How do we handle deletion and retention? GDPR gives a right to deletion, but the AML Act and the accounting rules may require that certain information is retained. Agree retention and deletion for each type of information with your compliance department before the system is built.
Conclusion: infrastructure, not a quick fix
Speed and security are not an either-or in credit assessment, but you do not reach both by letting a model decide alone. It happens by letting AI take the preparatory work, keeping the decision with people, and building the solution so that it can be explained. The companies that succeed start with one type of case, measure before and after, and bring compliance in from day one.
Want to hear what we could build for your credit department? Book 20 minutes of free consultation with Agent360. We listen to where the time goes, and tell you honestly what is worth building.
Sources
- Finanstilsynet: Brug af kunstig intelligens i den finansielle sektor (2024)
- Finanstilsynet: God praksis ved brug af AI i den finansielle sektor (maj 2024)
- EU's regulation on artificial intelligence (the AI Act), annex III on high-risk systems
- GDPR (Regulation 2016/679), Article 22 on automated decisions
Contents
Which of your manual steps could disappear?
We map where data is moved by hand between your systems, and what can realistically be automated.
Book a mapping sessionRelated articles

AI in the Call Centre: How to Handle Peak Loads Without Hiring More People
AI in the call centre: handle peak loads, follow-up work and knowledge search without losing quality. Frameworks, rules and measures before you build.

AI for Car Sales & Leasing: Filling the Test-Drive Calendar
AI for car dealers and leasing companies: faster replies, follow-up on expiring leases and old leads, test-drive booking and service. Done with control.

AI for Events and Courses: Registration, Reminders and Delivery
How AI can support administration around events and courses: registrations, waiting lists, reminders, the instructor pack and the rules for invitations.