No. A mortgage approval commits a large sum over decades, secured on someone's home, and the decision belongs to an underwriter. An AI agent can gather documents, verify income, check the case against lending policy and prepare a clear summary. The approval, policy exceptions, the offer and the release of funds stay with named people.
What makes a mortgage different from other lending?
Three things set it apart. The amount is large relative to the borrower's income. The term runs for decades, so a weak decision has a long tail. And the loan is secured on a home, so the worst outcome for the borrower is losing it.
The process also has more moving parts than a personal loan: the property valuation, the borrower's deposit and its source, the solicitor or conveyancer, and the date of completion when the money moves. Each part has its own risks. A mistake in any of them can mean lending against a property worth less than assumed, to a borrower who cannot sustain the payments.
So the conditions under which an agent could approve a mortgage do not really arise. What it can do is make the underwriter's work faster and more complete.
What are the conditions for the agent's role?
The agent's work is safe when:
- it works from documents the applicant supplied and data from trusted sources, and records which is which;
- every check it runs follows the written lending policy;
- it marks anything it cannot resolve rather than choosing an interpretation;
- the underwriter sees the evidence behind each statement, not just the summary.
What may the agent do, and what stays with a person?
| Stage | The agent | Stays with a person |
|---|---|---|
| Document collection | Requests and chases documents, checks they are complete and legible | Nothing routine |
| Income verification | Matches payslips, tax documents and bank credits | Judging variable or self-employed income |
| Deposit and source of funds | Traces the deposit through statements and flags gaps | Deciding whether the source is acceptable |
| Affordability | Runs the affordability model under policy | Overrides and stress-test exceptions |
| Valuation | Reads the valuation report and flags conditions | Accepting the valuation |
| Policy check | Lists every policy rule and whether the case meets it | Approving exceptions |
| Decision and offer | Drafts the case summary and the offer | The underwriter approves; the offer is issued by a person |
| Completion | Prepares the funds release and checks conditions | Releasing funds |
Document gathering is where agents save most time; the process is described in how to gather documents for mortgage applications with an AI agent.
A worked example
Take a mortgage lender assessing an application for 280,000 on a house valued at 350,000. The applicant is employed with a regular salary and an annual bonus.
The agent checks that payslips, the employer's reference and bank credits agree. The base salary matches. The bonus appears in one year's records but not the other. The lender's policy counts only a proportion of bonus income, and only where it is sustained. The agent calculates affordability with and without the bonus and shows both.
It also traces the deposit. Most came from savings over three years, but 20,000 arrived two weeks ago from an account in another name. The agent flags it as needing a gifted deposit letter and an explanation. The underwriter reviews both points, asks for the letter, and approves on the basis of salary alone.
Why is completion no place for autonomy?
Completion is when the lender sends the money to the conveyancer, and it is the most exposed point in the process. Fraudsters target it with spoofed payment details, and a payment sent to the wrong account is gone. The argument is made in mortgage completion is no place for autonomy. The agent can check that every condition of the offer is met and prepare the release. A person releases it, after verifying the payment details through a known channel.
What controls make it safe?
- No decision capability. The agent's role scope ends at the case summary.
- Evidence-linked statements. Every figure in the summary links to the document it came from.
- Policy as a checklist. Each rule is listed with pass, fail or unclear, so exceptions are visible.
- Separate approval for the agent itself. Before the agent handles live applications, it should go through your production approval, as set out in how to approve a new AI agent for production.
For the general case, see whether an AI agent can approve a loan, and for the wider limits, AI agents in banking: what they may and may not do.
Questions people ask
Can the agent issue a decision in principle?
A decision in principle is a lighter, indicative check. If the lender's approved model produces it under policy, the agent can deliver the result and explain that it is not an offer. It should not produce one by its own reasoning.
Can the agent talk to the applicant's conveyancer?
It can send and receive routine updates and document requests. Changes to payment details or completion arrangements go to a person and are verified independently.
Can the agent handle product transfers for existing borrowers?
Switching an existing borrower to a new rate with no extra borrowing is lower risk, and many lenders run it with less review. Whether the agent can complete it depends on the lender's policy and the evidence for that capability.