What problem does it solve?
Loan application forms can lose customers. They ask for terms the applicant does not understand, in a language that may not be their first, and they rarely explain which product fits. Some applicants give up, and many completed applications arrive with missing documents or wrong figures. According to Google Cloud, Oper Credits reports that most loan applications in Belgium are returned for missing or incorrect information, and each return is another round between the customer and the credit team. Where branches are few, a phone or messaging channel can be the most practical way to apply.
The bank also has to be careful. Intake collects personal and financial data, touches product suitability and responsible lending duties, and sits right next to the credit decision. An assistant that drifts from collecting information into telling people they will or will not be approved creates regulatory and fairness risk.
- Oper Credits, a mortgage digitisation company serving about 20 banks, says only 30 to 40% of loan applications in Belgium are complete and compliant on first submission, as reported by Google Cloud.1,302 real-world gen AI use cases from the world's leading organizations (2025)
How does it work?
- Understand the need. The assistant asks what the customer wants to finance and explains the relevant products from approved, current product content, in the customer's language.
- Check the basics first. Published eligibility rules (age, residency, minimum income, product limits) are checked by a rules service before the customer invests time, and the assistant explains any rule that is not met.
- Capture the application in conversation. It asks one question at a time, fills the application fields, and reads uploaded payslips, statements and identity documents with document AI, asking again when something is unclear.
- Validate completeness. Before submission it checks that every required field and document is present and consistent, so the application arrives complete.
- Hand over a structured file. The complete application goes into the loan origination system with a summary; the credit decision is made there, by the bank's governed credit process and people.
- Keep the customer informed. The assistant tells the customer what happens next and when, and hands over to a lending specialist on request or when the case is complex.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Web chat, Mobile app, WhatsApp, Phone and voice
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Revenue growth, Customer experience, Speed and cycle time, Inclusion and access, Lower cost to serve.
Indicative value
A lender receiving 50,000 personal loan applications a year
USD 54,000 to USD 600,000
Application follow up cost avoided per year
How this is calculated
Formula: applications * incompleteShare * followUpAvoided * minutesPerFollowUp * costPerMinute. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Applications received per year applications, applications per year | 50,000 | 50,000 | The reference lender. |
| Share of applications that need follow up for missing or wrong information incompleteShare, fraction of applications | 0.3 | 0.5 | Editorial assumption. According to Google Cloud, Oper Credits, a mortgage digitisation company, says only 30 to 40% of loan applications in Belgium are complete on first submission, which implies 60 to 70% need follow up; this range stays below that to be conservative for other markets and products. |
| Share of that follow up the assistant prevents followUpAvoided, fraction of follow up | 0.3 | 0.6 | Editorial assumption, replace with your own pilot result. |
| Staff minutes per follow up minutesPerFollowUp, minutes per application | 20 | 40 | Editorial assumption covering calls, emails and re checking documents. |
| Fully loaded cost of a lending staff minute costPerMinute, USD per minute | 0.6 | 1 | Editorial assumption, replace with your own fully loaded cost. |
What it leaves out: Counts avoided follow up work only. It leaves out additional completed applications (the larger prize, but hard to predict), faster time to decision, the cost of the AI and the integration with origination.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Absa Bank
South Africa · Banking · 2025
Absa runs Abby, a customer facing assistant built on Salesforce Agentforce, in production. At a Salesforce event in June 2025 Absa's Relationship Banking technology chief described it asking business customers about their needs, such as working capital, and recommending matching products. The report says that, according to Absa's website, the agent can help customers apply for loans, open investment accounts and make international payments. Absa staff set guidelines on what the agent may not do, and Salesforce says it urges its partners to route topics such as pricing to employees. Salesforce could not share the impact on Absa customers, and no outcome figures were disclosed.
No outcome disclosed.
Figure
United States · Banking · 2025
Figure, a US fintech that offers home equity lines of credit, uses Gemini models to run chatbots that simplify and speed up the lending experience for consumers and for its own staff. No outcome figures were published.
No outcome disclosed.
Lloyds Banking Group
United Kingdom · Banking · 2025
Lloyds Banking Group uses Vertex AI to scale machine learning work across more than 300 data scientists and AI developers. In the same entry Google Cloud reports that the bank cut income verification in mortgage applications from days to seconds and has put 18 generative AI systems into production. This is back office work at the application stage, not a customer facing assistant.
No outcome disclosed.
Oper Credits
Belgium · Technology and software · 2025
Oper Credits, a Belgian mortgage digitisation company that serves about 20 banks in six countries, uses Vertex AI to automate document verification that used to take several hours of manual work. According to Google Cloud, only 30 to 40% of loan applications in Belgium are complete and compliant on first submission, and most are returned for missing or incorrect information. The company aims to raise that to 90%; the aim is a target, not a reported result.
No outcome disclosed.
Rocket Mortgage
United States · Banking · 2025
Rocket Mortgage runs an AI Digital Assistant across chat and voice that takes prospective borrowers from first questions to preapproval: it answers questions, collects information, pulls credit, presents personalised rates and loan options and hands the client to a human banker. According to the vendor, the programme started as a proof of concept and has grown to more than 400,000 successful chat conversations and over one million outbound dials a month. Clients who start with the assistant close at three times the rate of those who do not. Sierra also reports that clients who use both the AI chat and a banker convert four times better, without stating the comparison group.
- Interactions handled: at least 400,000, successful chat conversations per month
"What started as a proof of concept in May has grown to more than 400,000 successful chat conversations and over one million outbound dials each month, and both are rising fast."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Approved, versioned product content and eligibility rules
- The application data model of the origination system
- Sample documents per type for testing extraction
- Past applications with the reasons they were returned or declined
Systems to integrate
- Loan origination system
- Document storage and document AI
- Identity verification and authentication
- CRM for leads and follow up
- Contact centre or lending specialists for handover
Complexity: Medium
Capturing an application in conversation is well understood. The work is in the eligibility rules service, reliable document reading, writing clean records into loan origination and keeping the assistant on the right side of the line between information and a credit decision.
- 1
Pick one product and one channel
Start with a simple product (a personal loan or a card) on the channel your applicants use most, often the app or WhatsApp, and measure completion and quality against the web form.
- 2
Separate rules from conversation
Put eligibility checks and required fields in a rules service the assistant calls. The model explains; it does not decide whether someone qualifies.
- 3
Map every field to a question and a check
For each application field write the plain language question, the validation and what the assistant says when the answer does not fit.
- 4
Make documents conversational
Let the customer upload a photo, read it, show what was extracted and ask them to confirm, instead of retyping.
- 5
Define the handover and the wording
Agree the phrases the assistant may use about outcomes ("your application is complete and will be reviewed") and ban the ones it may not ("you are approved").
- 6
Monitor fairness from day one
Track drop off and completion by language, channel and customer segment, so the assistant does not become a filter that disadvantages some groups.
Guardrails
- The assistant never states or implies a credit decision
- Eligibility checks come only from the versioned rules service
- Product explanations come only from approved, current content
- Personal and financial data masked in prompts and logs, and kept in region
- Handover to a specialist on request, for complex cases and on vulnerability signals
KPIs to instrument
- Application completion rate versus the web form, by segment and language
- Share of applications complete at first submission
- Time from first message to complete application
- Handover rate and reasons
- Complaints mentioning the assistant
Human in the loop
Credit officers or the bank's governed credit models decide every application. Lending specialists take over complex or vulnerable cases, and a sample of conversations is reviewed each week for accuracy, suitability and fair treatment.
Common failure modes
- Implied approvals
- The assistant says something that sounds like a decision. Restrict outcome wording and test for it.
- Stale rules or rates
- The assistant quotes last month's criteria. Version the rules and content with owners and review dates.
- Silent filtering
- Customers are discouraged before they apply, unevenly across groups; in the US, Regulation B bars discouraging applicants on a prohibited basis. Measure drop off by segment and review the eligibility wording.
- Bad data, faster
- Extraction errors enter origination unnoticed. Show extracted values to the customer for confirmation and sample them.
What are the risks and rules?
EU AI Act
Depends on design
Explaining products and capturing an application is limited risk with an Article 50 disclosure. If the assistant evaluates creditworthiness or filters applicants on its own assessment, it falls under Annex III point 5(b) and becomes high risk, so keep the decision in the governed credit process.
Rules that apply
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 5(b) lists AI systems used to evaluate the creditworthiness of natural persons or establish their credit score as high risk, except systems used to detect financial fraud; intake must stay outside that scope or meet the high risk duties.
- Regulation B, § 1002.4 General rules (Consumer Financial Protection Bureau, North America). Section 1002.4(b) bars statements to applicants or prospective applicants that would discourage them on a prohibited basis, and the official commentary names interview scripts that do so; an intake assistant's eligibility wording falls under the same rule.
- Responsible lending (Australian Securities and Investments Commission, Asia Pacific). Example of national responsible lending obligations that an intake assistant's questions and wording must support.
Controls to put in place
- AI disclosure at the start and a clear route to a person
- Versioned eligibility rules and product content with change control
- Log of every question, answer and document captured
- Fairness monitoring of drop off and completion by segment
- Data protection impact assessment for the personal and financial data collected
Frequently asked questions
- Can a chatbot approve a loan?
- It should not. The assistant collects and checks information; the credit decision stays with the bank's governed credit process. An assistant that assessed creditworthiness itself would be a high risk system under the EU AI Act.
- Are banks using AI assistants for loan applications today?
- Yes, though outcome data is scarce. Absa's Agentforce based assistant Abby is reported to help customers apply for loans, and Figure uses Gemini powered chatbots in its home equity lending. Rocket Mortgage's assistant, built with Sierra, goes further than intake: it also pulls credit and takes clients to preapproval, and the vendor says it handles more than 400,000 chat conversations a month. Oper Credits, which serves about 20 banks, uses Vertex AI to automate document checks on mortgage applications.
- What should we measure first?
- Completion rate against your web form and the share of applications that arrive complete, both split by language and customer segment. They show value and fairness at the same time.
How to cite this page
Blits.ai AI Use Case Library, "Conversational AI for loan application intake", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/conversational-loan-application-intake. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published