No. The agent prepares the credit case, and an approved credit model or a credit officer decides. An AI agent can gather evidence, run the checks policy requires, draft the credit summary and prepare documents. Approving outside automated policy, setting terms and releasing money stay with people, because a disbursed loan cannot be called back on a hunch.
Why does a loan approval need more than an agent's judgement?
A loan approval commits the lender's money and the borrower's future income. Once the funds are drawn, the lender has a contract it cannot unwind because it changed its mind. If the borrower could not afford the loan, the harm arrives months later as arrears, collections and a damaged credit record.
Lenders already automate many small, standard approvals. They do it with credit models that have been built, validated and approved under model governance, inside a written credit policy. That is different from an agent reasoning its way to a decision. The EU AI Act lists creditworthiness assessment and credit scoring of natural persons as high-risk, and GDPR Article 22 and its UK equivalent restrict decisions based solely on automated processing with legal or similarly significant effects. Those safeguards attach to the decision process, which needs to be governed as such.
What may the agent do with the applicant's data and money?
| Step | The agent | Where it stops |
|---|---|---|
| Collecting the application | Gathers data and documents, notes gaps | Nothing routine |
| Verifying evidence | Checks identity, income and bank statements against each other | Deciding a discrepancy is acceptable |
| Running checks | Runs credit search, affordability and fraud checks under policy | Interpreting conflicting results |
| Scoring | Passes the case to the approved model and records the result | Overriding the model |
| Drafting the credit summary | Sets out the case, risks and policy exceptions | The credit decision |
| Pricing | Applies the approved pricing grid | Discretionary pricing; see whether an AI agent can price a loan |
| Documents | Prepares the agreement and conditions | Agreeing non-standard terms |
| Disbursement | Prepares the payment | Releasing funds |
The applicant's data is used for the assessment the applicant applied for, and nothing else. The agent never moves money. Disbursement is a separate capability held for a person.
A worked example
Take a business lender. A bakery with two shops applies for 50,000 to buy a new oven and fit out a third site.
The agent collects two years of accounts, six months of bank statements and the management accounts. It reconciles turnover in the accounts with the bank credits and finds them consistent. It runs the credit search on the company and the directors, and calculates debt service cover using the lender's method.
The draft credit summary notes two points the officer needs to weigh. A single wholesale customer accounts for a large share of income, and the third site's lease has not yet been signed. The agent does not decide whether those risks are acceptable. It holds the case for the named credit officer, who approves on condition that the lease is signed before drawdown. The agent prepares the facility letter with that condition. When the signed lease arrives, it checks the document and prepares the disbursement, which a person releases.
What controls make it safe?
- The decision belongs to a governed process. The agent runs the approved model and prepares cases for officers. It has no capability to approve on its own reading.
- Conditions tracked to drawdown. Conditions attached to an approval are checked before any money moves, and the agent holds disbursement until they are met.
- Separation of duties. The agent that prepares the case does not release the funds.
- Integration through defined actions. The agent reaches the origination system through specific, permissioned actions rather than broad access; see connecting an AI agent to your loan origination system.
- Reasons recorded at the time. The factors behind each decision are filed when it is made, so any later explanation is accurate.
For the stages of the process, see what loan origination is. Mortgages raise further issues, set out in whether an AI agent can approve a mortgage. The wider limits are in AI agents in banking: what they may and may not do.
Where Pryme fits
In Pryme, gathering evidence, drafting the credit summary and preparing disbursement are separate capabilities with separate depths. Approving a loan on the agent's own reading is outside its role scope, and disbursement is held for a named person. Each verdict is filed, so a credit committee or examiner can see exactly what the agent prepared and who decided.
Questions people ask
Can the agent approve small loans the model has already passed?
If the lender's approved process lets the model decide small, standard applications, the agent can carry that decision through and record it. The decision is the model's under policy, not the agent's, and the safeguards the law requires still apply.
Can the agent tell an applicant they are likely to be approved?
No. An indication is heard as a promise. The agent can explain what happens next and how long it usually takes.
What happens after the loan is paid out?
Ongoing administration, such as statements, payment changes and arrears contact, is loan servicing, where agents often do more of the routine work. See what loan servicing is.