Yes, with the customer's knowledge and for a defined purpose. An AI agent can run a soft search to show likely eligibility, and a full search once the customer submits an application, at depth 3. It must not run a full search before the customer applies, search people who have not agreed, or turn the report into a lending decision.
Why does the type of search matter so much?
Credit reference agencies usually offer two kinds of search. A soft search is visible only to the customer and does not affect how other lenders see them. A full, or hard, search is recorded on the customer's file and is visible to other lenders. Several full searches in a short time can make a customer look as if they are applying everywhere, and can lower their chances of being accepted.
So running the wrong kind of search is not a harmless mistake. An agent that runs a full search because a customer asked "would I qualify?" has left a mark on their file that they did not choose. The agent's first job is to know which question the customer is asking: "what are my chances?" or "I want to apply."
What conditions decide it?
The agent may run a search when:
- the customer has been told, in the bank's approved wording, what kind of search will be run and what it means for their file;
- a soft search is used for eligibility checks and quotes;
- a full search is run only on an application the customer has submitted;
- the person searched is the applicant, or a joint applicant or guarantor who has agreed separately;
- the purpose is recorded with the search.
Searches on a partner, a relative or a company director who has not agreed are not the agent's to run.
What may the agent do, and where must it stop?
| Task | The agent | Where it stops |
|---|---|---|
| Eligibility check | Runs a soft search with the customer's agreement, depth 3 | A full search |
| Application check | Runs the full search once the application is submitted, depth 3 | Before submission |
| Joint applicant or guarantor | Runs a search after their own agreement | Searching on another person's word |
| Reading the report | Extracts facts for the credit file, depth 1 | Deciding what they mean for the loan |
| Discrepancies, such as an unknown account on the file | Records and asks the customer | Fraud or identity concerns |
| Explaining the report to the customer | Shows what the bank used | Interpreting their overall credit standing |
| Repeat searches | Reuses a recent search within its validity | Running another without a reason |
Reading the report is not scoring it. The step from facts to a decision is covered in what credit scoring is, and in the EU AI Act, credit scoring of individuals is listed as high-risk under Annex III.
An example: "would I get a loan?"
Take a customer who messages: "Would I get a loan of 5,000 if I applied?" The agent explains that it can run a soft search that will not affect their file, and asks if they agree. They do. The soft search and the bank's eligibility rules show a likely acceptance at a stated rate range.
The customer then applies. Only now does the agent explain that a full search will be recorded, and runs it once the customer confirms. The report shows a credit card the customer does not recognise, opened three months earlier. The agent does not decide whether this is fraud. It asks the customer about it, records the answer ("I've never had that card"), and holds the application for a person on the credit team with an identity concern flagged. It also tells the customer how to contact the credit reference agency about the entry.
What controls make it safe?
- Search type tied to the stage. The agent's tool runs a soft search for eligibility and a full search only when an application record exists.
- Consent recorded per person. Each person searched has their own recorded agreement.
- Report facts sealed. The report arrives from the agency stamped with its source and date, and the agent cannot edit it.
- No decision capability. The agent passes facts to the credit decision; it does not approve or decline.
- A referral note for anything odd. Discrepancies go to a person using the bank's credit referral note template.
The credit check is one input; the other is whether the customer can afford the loan, covered in whether an AI agent can check affordability for a loan and checking income and affordability evidence with an AI agent. The general limits are in AI agents in banking: what they may and may not do.
Where Pryme fits
In Pryme, running a soft search and running a full search are separate capabilities, each with its own certificate. Attest replays the agent on the lender's past enquiries and the scope gate fails any run where it runs a full search on an eligibility question. The report arrives with signed provenance from the agency, and the agent's summary links back to it.
Questions people ask
Can the agent run a credit check on a business?
Yes, using business credit data under the same principles. Searches on the directors as individuals need their own agreement.
Can the customer see the report the bank used?
The customer can usually get their own report from the agency. The agent can explain which facts the bank relied on.
How long can a credit check be reused?
That depends on the lender's policy. The agent reuses a search within its validity period and records that it did.