Not with its own judgement. An AI agent should never produce a creditworthiness score from its own reading of a customer's data. It can gather the inputs, call the bank's validated scoring model, report the result with its reason codes, and explain it. The model belongs to model risk governance, and the decision to lend to the credit process.
Why can't the agent just work out a score?
Because it would be an unvalidated credit model. Banks score credit with models that are built on historical outcomes, tested for accuracy and stability, checked for unfair effects on groups of customers, documented, independently validated and monitored. Supervisors expect this. Model risk guidance such as SR 11-7 in the United States, PRA SS1/23 in the UK and OSFI Guideline E-23 in Canada all set out what governing a model involves.
A language model asked "how creditworthy is this person?" will give an answer. It will sound reasoned. It will not be validated, its errors will not be measured, and it may vary between runs on the same facts. Using it would put an ungoverned model at the centre of a decision that affects people's access to credit. Under the EU AI Act, creditworthiness assessment and credit scoring of natural persons are listed as high-risk in Annex III, with obligations on both the provider and the deployer.
What can the agent do around the score?
A lot, as long as it stays on the right side of the model. The agent can:
- collect and check the application data the model needs, and flag missing or inconsistent inputs;
- call the scoring service with those inputs and receive the score and reason codes, at depth 1;
- record exactly which model version scored the application;
- explain the factors to the customer, using the bank's approved descriptions of each reason code;
- prepare the file for a credit officer where the score sits in a referral band.
Our explainer on what a creditworthiness assessment is sets out how the score fits into the wider assessment, alongside affordability and policy rules.
What stays with the bank's model and people?
| Step | Who | Why |
|---|---|---|
| Calculating the score | The validated model | Governed, tested and monitored |
| Changing the model or its cut-offs | Credit risk, with validation | Material model change |
| Overriding a score | A credit officer, with recorded reasons | Overrides are monitored for fairness |
| Explaining a decline | The agent, from approved reason codes | In some markets specific reasons are required |
| Deciding to lend | The bank's credit decision process | The decision, not just the score |
In the United States, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used. The agent can deliver those reasons accurately because they come from the model, not from the agent's guess about what mattered.
An example: a referral-band applicant
Take an applicant for a 6,000 loan whose score falls in the bank's referral band. The agent has gathered the inputs and called the model; the reason codes are a short credit history and high use of existing credit limits.
The agent prepares the referral. It summarises the inputs, the score, the model version and the reason codes. It also notes something outside the model: the applicant's bank statements, which they uploaded, show a regular salary that started four months ago, after a period of study. The agent does not adjust the score to reflect this. It records the observation for the credit officer, who can weigh it and, if policy allows, override with a recorded reason.
The record shows the model's output untouched and the officer's decision beside it.
How is scoring a customer different from scoring the agent?
The word "score" causes confusion. Certification scores an agent's answers against expected outcomes; see how you score an AI agent's answers. That score tells you how well the agent did on past cases. It is not a licence to act on the next one, a point made in an eval score is not a bound. Neither kind of score should be produced by the agent about itself or about the customer.
What controls make it safe?
- No scoring capability. The agent can call the model; it has no tool that writes a score.
- Model version recorded. Every application records which validated model scored it.
- Reason codes from the model. Explanations to customers use the model's codes and approved wording only.
- Observations kept separate. Anything the agent notices outside the model goes to the officer as a note, not into the score.
The general limits are in AI agents in banking: what they may and may not do.
Questions people ask
Can the agent tell the customer their score?
It can tell them the factors behind the decision, using approved wording. Whether the numerical score is shared depends on the bank's policy.
Could an agent ever become part of the scoring model?
Only by going through the same model development and validation as any other credit model. At that point it is a model governed as such, not an agent improvising.
Can the agent suggest how a customer could improve their chances?
It can share the bank's general guidance about the factors lenders look at. It should not promise that any action will change the outcome.