Credit assessment and agentic AI under the EU AI Act

When creditworthiness meets high-risk AI

AI systems for evaluating natural persons' creditworthiness or establishing their credit score are explicitly listed in Annex III of the EU AI Act as high-risk ^[1]. That classification brings a long list of obligations: risk management, data governance, technical documentation, automatic logging, human oversight, accuracy requirements and a conformity assessment before the system can be placed on the market. For systems under Annex III, those requirements apply from 2 December 2027 ^[1].

Think of it like a driving licence for a lorry. You can drive a car on an ordinary licence, but once the vehicle is heavy enough the rules change. Credit assessment of people crosses that line. The classification isn't about whether the AI is good at its job; it's about the potential consequences when it gets things wrong.

The academic article from the University of Helsinki examines how these rules apply when AI agents, not just traditional models, start making credit decisions ^[6]. The short version: if an agent can assess creditworthiness, it's a high-risk AI system, and all the obligations in the Act apply to what the agent does.

What the Act asks for

Article 10 says training, validation and test data for high-risk AI must be relevant, representative and subject to documented data governance, including the data's origin ^[7]. Article 12 requires the system to technically allow for automatic recording of events over its lifetime, for traceability ^[2]. Providers must keep the logs they control for at least six months under Article 19, and deployers must do the same under Article 26(6), unless another law requires longer ^[3].

Article 27 adds another layer for certain deployers. Public authorities, private operators delivering public services, and anyone using high-risk AI for credit assessment of natural persons or for pricing and risk assessment in life and health insurance must carry out a fundamental rights impact assessment before putting the system into use ^[4]. That assessment looks at the processes the deployer has in place for human oversight, and the measures to inform and support people who will be affected by the system.

Finally, Article 86 gives individuals a right to a clear and meaningful explanation when a decision based on high-risk AI under Annex III (except critical infrastructure) produces legal or similarly significant effects for them ^[5]. The explanation must cover the AI system's role in the decision-making procedure and the main elements of the decision.

Those are minimum obligations. Under the GDPR, Article 22 already restricts decisions based solely on automated processing, including profiling, when they produce legal effects or similarly significantly affect someone ^[8]. Credit decisions often fall into that category, and when they do, the data controller must at least give the person the right to obtain human intervention, to express their point of view and to contest the decision.

When the agent does the assessment

The Helsinki paper looks at what happens when an agentic AI system, one that can plan and act, is given tools to assess creditworthiness ^[6]. Say a lender builds an agent that can pull a customer's transaction history, check their income against thresholds and decide whether to approve a loan. The agent isn't just assisting; it's making the assessment.

Under the AI Act, that agent is a high-risk AI system. The logging duty in Article 12 applies to what the agent does, which means each assessment needs to be recorded ^[2]. The explainability duty in Article 86 applies, so the customer has a right to know what the agent considered and why it reached its conclusion ^[5]. The fundamental rights impact assessment under Article 27 applies before the agent is put to work ^[4].

The tricky part is that agents act in ways that are harder to predict than a traditional scoring model. A model runs the same calculation every time. An agent might decide to fetch additional information, or skip a step, depending on what it finds. That makes logging more important, not less, because the log is the only way to know what the agent actually did in a particular case.

The evidence question for credit logs

Here's where it gets practical. The law requires the logs. It says they must be kept for at least six months. It doesn't say how you prove the log you show an auditor or a supervisor six months later is the one the system wrote on the day.

Say a customer disputes a credit decision and asks for an explanation under Article 86. You pull the log and it shows the agent ran a check, found an income figure and applied a threshold. The customer says the income figure is wrong and suggests the log has been edited. You know it hasn't, but how do you show that?

Access controls and monitoring are the foundation, and they matter. But they don't answer the question directly, because someone with the right access could still change a log. The missing piece is a way to show that what you're presenting is exactly what was recorded at the time, without anyone having to take your word for it.

How a digital witness would work in this case

The thing to seal is each assessment record as the agent writes it to the log: the customer reference, which tools the agent called, the data it retrieved, the decision it reached and the identifier of any human reviewer. Sealing takes a fingerprint of the record, not the record itself, so the assessment data never leaves your systems.

The fingerprint is sealed on WitniumChain at the moment the record is written. Anyone you later share the record with, whether it's the customer, an auditor or a court, can calculate the fingerprint again and check it against the sealed one. If they match, the record hasn't changed since it was sealed. Change a single character and it won't match.

That doesn't show the decision was right, fair or in line with the GDPR. It doesn't replace the conformity assessment, the fundamental rights impact assessment or the human oversight the law requires. What it does is let you prove that the log you're presenting is the one the system wrote, without anyone having to trust that your organisation hasn't quietly edited it in the meantime.

You can seal through the REST API, the SDK or the MCP server. The content stays with you. The first 2,000 witnesses a month are free. Read the docs or get access now.

How it would work

1. The agent writes an assessment record. Your system sends a fingerprint to WitniumChain and gets a witness ID back.
2. When the record is questioned, you share it with the witness ID.
3. The recipient calculates the fingerprint and checks it. If it matches, nothing has changed.

What this means for you

If you're deploying AI for credit assessment of natural persons, here are the questions to ask now:

• Have you identified which of your systems fall under the high-risk definition in Annex III, and do you know when the obligations apply to you?
• Can your systems log what they do in enough detail to meet Article 12, including which data an agent retrieved and which tools it called?
• Do you have a process for keeping those logs for at least six months, and can you show an auditor that a log from March hasn't been edited since March?
• If you're required to carry out a fundamental rights impact assessment under Article 27, have you started mapping the human oversight processes and the information you'll give to affected individuals?
• Can you give a customer a clear explanation under Article 86 of what the AI did and why it reached that conclusion, in a way that doesn't require them to trust that your records are complete?

The high-risk requirements are detailed, and the penalties for non-compliance go up to 15 million euros or 3 per cent of global annual turnover ^[1]. The safe approach is to assume you'll need to prove what your AI did, and to put the logging and the protection in place before 2 December 2027.

What to watch

2 December 2027 is when the high-risk requirements for Annex III systems, including credit assessment, begin to apply ^[1]. The AI Office has published guidance on the high-risk classification and is expected to issue further materials on conformity assessment and fundamental rights impact assessments. If you're building or deploying agentic AI for credit decisions, the safest assumption is that the agent is a high-risk system and that everything it does needs to be logged and explainable.