The line sits between recommending and deciding. An agent may gather the evidence, run the checks it is permitted to run, and draft a recommendation with its reasons. The credit decision, and the reasons given to the applicant, belong to a person who has read the evidence and can disagree with the agent. Everything on the agent's side of the line is preparation.
What counts as the decision?
It is easy to say "the person decides" and harder to pin down what deciding means. A credit decision is not only the moment someone clicks approve or decline. It includes every step that settles the outcome before that click.
An agent that filters applications before a person sees them is deciding for the applications it filters out. An agent whose recommendation is accepted without review in nearly every case is deciding in practice, whatever the process map says. An agent that sets the terms, such as the limit or the rate, within which the person then approves, has decided the terms.
So the useful test is not who presses the button. It is whether the outcome could be different because a person considered the case. If not, the agent decided.
Why does the decision stay with a person?
Three reasons, which reinforce each other.
Law and supervision. Credit decisions about individuals attract specific rules in many markets. Under GDPR and UK GDPR, Article 22 restricts decisions based solely on automated processing with legal or similarly significant effects; GDPR Article 22 and AI agents in credit decisions covers how that applies. The EU AI Act lists creditworthiness assessment and credit scoring of natural persons as high-risk in Annex III. In the United States, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used.
Reasons. An applicant who is declined is often entitled to know why, and the reasons must be true and specific. A person who has weighed the evidence can give reasons they stand behind. An agent's explanation is its account of its own output, which can be plausible and still not be the real reason.
Accountability. Credit decisions shape people's lives and the lender's balance sheet. Someone in the lender needs to own each one.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What sits on each side of the line?
| The agent may | Stays with a person |
|---|---|
| Collect application data and documents | Deciding to approve or decline |
| Verify income and identity evidence against strong sources | Setting or overriding the limit, rate and terms |
| Run checks it is permitted to run, with consent recorded | Deciding to run a hard search where policy requires a person |
| Calculate affordability and policy measures | Weighing factors the policy does not capture |
| Flag anomalies, inconsistencies and missing evidence | Deciding whether an anomaly matters |
| Draft a recommendation with its basis and reasons | Writing or confirming the reasons given to the applicant |
| Draft the approval or decline letter | Approving the letter before it is sent |
The left column is substantial. On many applications, gathering and checking is most of the time spent. An agent that prepares credit applications in an origination system can hand an underwriter a complete, verified file. Whether the agent should run a credit check depends on the type of search and the consent behind it.
How do you keep the person's role real?
The danger is a process that looks human-led and is not. Four design choices keep the person's role meaningful.
- Show the evidence, not just the recommendation. The underwriter sees the facts, their sources and what was verified, so they can form their own view.
- Make disagreement easy. Overriding the recommendation should take no more effort than accepting it, and should prompt for a reason that is filed.
- Watch override rates. A team that never overrides the agent may be rubber-stamping. A team that overrides constantly may have an agent that is not fit for the work. Both are signals.
- Let reasons come from the person. The agent can propose reasons, but the person confirms that they are the actual basis of the decision.
The detail of where automated support tips into automated decision is in automated decisions under GDPR: where agents cross the line.
What about small or pre-approved credit?
Some lenders pre-approve limits for existing customers, or offer small credit lines through simple rules. These are still credit decisions. The rule that sets the pre-approval was a decision, made by someone, and it should be owned, documented and reviewed like any credit policy. An agent applying that rule to an individual is executing a decision a person already made, which is a different thing from the agent deciding.
That distinction is worth keeping clear. Applying an approved rule can, with evidence, run at depth 3. Forming a judgement about an individual's creditworthiness stays with a person.
Where Pryme fits
In Pryme, drafting a credit recommendation and recording a credit decision are separate capabilities. The agent can earn depth 2 or 3 for gathering and drafting; recording an approval or decline is held for a named underwriter. The facts relied on, their provenance and the underwriter's reasons are filed together. See AI agents in banking: what they may and may not do.
Questions people ask
Can an agent decline obviously ineligible applications?
Checking hard eligibility rules, such as minimum age, can be automated as a published rule. Declines on judgement, or on rules involving assessment, should go to a person.
Should applicants be told an agent was involved?
Where people interact with an AI system, the EU AI Act requires that they be told in many cases, and other markets set their own expectations. Being clear about the agent's role also helps applicants understand the process.
What if the agent's recommendation and the underwriter disagree?
The underwriter decides, and the reason for the disagreement is filed. Those disagreements are valuable cases for the agent's next certification run.