What problem does it solve?
A mortgage file passes through several people before a lender can say yes. A processor collects documents, an underwriter checks income, assets, credit and the property against the guidelines of the investor who will buy the loan, and conditions go back to the borrower until the file is clean. Each investor has its own overlays, so the same borrower can be eligible for one buyer and not for another. The process is slow and sequential, and the underwriter is the scarce person in it.
Better, a US digital lender, puts the average for the traditional mortgage industry at 21 days. The document and income steps have specialist tools (see the related page on income and document verification). The step that remains is the decision: reading the whole file against the guidelines, listing what is missing and saying whether the loan can be approved. That is the part a decision engine takes on.
- Better says the average for the traditional mortgage industry is 21 days, against a median of 2 minutes and 24 seconds for its own Tinman engine.Better Announces First Conversational Credit Decision Engine in ChatGPT with OpenAI (2026)
How does it work?
- Keep one live record of the loan. The system holds a snapshot of the file, with its facts, documents, actions and outstanding conditions, and updates it as documents arrive.
- Read the documents. Specialized agents read the credit, income, collateral and asset documents in the file and write the facts they find into the record.
- Apply the guidelines. An orchestrator routes the file to agents that check it against the guidelines of the investors and programs the lender sells to, including the lender's own overlays. Better says its engine covers the guidelines of more than 45 institutional investors.
- Return a decision ready status. The result is an approval, a referral or a list of open conditions, supported by a decision tree.
- Refer what needs judgment. Files with exceptions, conflicting documents or unusual income go to an underwriter with the evidence already assembled.
- Feed back. Underwriter overrides and investor repurchase or defect findings go back to improve the rules and the readers.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Emerging
- Channels
- API and system to system, Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cycle time | Not pooled | 2.4 minutes | 1 | 1 organization |
| Productivity gain | Too few to pool | about 50% | 1 | 1 vendor |
Value drivers: Speed and cycle time, Lower cost to serve, Employee productivity, Customer experience.
Indicative value
A mortgage lender funding 12,000 loans a year
USD 216,000 to USD 1.5 million
Underwriting capacity released per year
How this is calculated
Formula: loans * automatedShare * hoursPerLoan * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Funded loans per year loans, loans per year | 12,000 | 12,000 | The reference lender. |
| Share of loan files whose underwriting is automated automatedShare, fraction of loan files | 0.2 | 0.4 | The high value follows the line on NEO Home Loans' site that Better's Tinman automates underwriting for nearly 40% of loan files, which is a product claim and not a measured NEO result. The low value is an editorial assumption for a first year. Source |
| Underwriter hours per file today hoursPerLoan, hours per loan | 2 | 4 | Editorial assumption. Replace with a time study of your own underwriting desk. |
| Fully loaded cost of an underwriting hour costPerHour, USD per hour | 45 | 80 | Editorial assumption. Replace with your own fully loaded cost. |
What it leaves out: Capacity released, not cash saved, unless headcount changes. It leaves out the effects of faster decisions on pull through and on borrower experience, the cost of the platform, integration and model governance, and repurchase risk if automated decisions are wrong.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Better Home & Finance Holding Company
United States · Banking · 2026
Better built Tinman, an AI platform trained on its own mortgage data that keeps a live snapshot of each loan file and uses an underwriting orchestrator to route work to specialized agents that read documents, apply investor guidelines and return a decision ready status with a decision tree. In March 2026 it released the engine as an app inside ChatGPT Enterprise for loan officers at other lenders, who connect their own guidelines, pricing and CRM data. Better reports that loans from Tinman partners made up $646 million, 44% of its funded loan volume, in the fourth quarter of 2025.
- Cycle time: 2.4 minutes, median per loan, reported at the March 2026 launch
"in as little as 47 seconds and a median of 2 minutes and 24 seconds compared to the average of 21 days for the traditional mortgage industry."
Claimed by: organization
NEO Home Loans
United States · Banking · 2025
NEO Home Loans, a US mortgage lender with a network of loan officers and branches, runs its origination on Better's Tinman platform. NEO's website describes Tinman as automating underwriting for nearly 40% of loan files and cutting fulfillment costs by up to $1,400 per funded loan, a product description from Better that NEO repeats and not a result measured at NEO. Better reports that within six months of fully rolling out, NEO increased loans funded per officer by 91%, per processor by 17% and per underwriter by nearly 50%. The partnership began in January 2025 and had grown from a $1.50 billion to a $2.97 billion run rate by March 2026, according to Better.
- Productivity gain: about 50%, loans funded per underwriter, within six months of full rollout
"Within six months of fully rolling out, NEO increased loans funded per officer by 91%, per processor by 17%, and per underwriter by nearly 50%, powered by Tinman AI"
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
- Current guidelines and overlays for each investor and product, in a form that can be versioned
- A set of past loan files with the underwriter's final decision and conditions, to test the engine against
- Document images and extracted data for credit, income, assets and property
- Records of repurchase demands, defects and quality control findings on past loans
Systems to integrate
- Loan origination system and borrower portal
- Credit bureau, income and asset verification and appraisal providers
- Pricing engine and investor eligibility data
- Document storage and condition tracking
- Quality control and secondary marketing systems
Complexity: High
The engine has to read every document type in a mortgage file and encode the guidelines of each investor and program the lender sells to, and those guidelines change often. It needs a loan origination system that exposes the file and its conditions by API, a rules and overlay store the credit team owns, pricing and credit data feeds, and a model risk process. Better says Tinman is trained on more than a decade of its own mortgage data.
- 1
Run the engine in shadow beside underwriters
For a few hundred files, let the engine decide without acting on it, and compare its decision, conditions and cited rules with the underwriter's. Sort the disagreements by cause.
- 2
Automate the clean files first
Start with salaried borrowers, standard products and complete documents. Refer anything that is self employed, has unusual assets or has a document conflict.
- 3
Make every decision traceable to a rule
Store the facts used, the guideline and version applied, and the document page each fact came from, so that a quality control reviewer or an investor can reconstruct the decision.
- 4
Own the guideline layer
Put one named person or team in charge of investor guidelines and overlays, with a change log and a test set that runs before every update.
- 5
Add adverse action reasons
For a decline or counteroffer, produce the specific principal reasons from the rules that failed, in language a borrower can read, and test them with compliance.
- 6
Widen by program and investor
Add products and investors one at a time, with a defect rate target and a release gate for each, rather than turning on every program together.
Guardrails
- No automatic decline; declines and counteroffers are confirmed by an underwriter
- Every decision stores the guideline version, the facts used and their document source
- Fair lending testing on decisions, referrals and conditions across protected groups before release and on a schedule
- Borrower data masked before it reaches a model that does not need it
- Investor and agency requirements for AI governance checked before use on loans sold to them
KPIs to instrument
- Share of files decided without an underwriter touch, and share referred
- Time from complete file to decision, and conditions per file
- Defect and repurchase rate of automated versus manual loans
- Agreement between the engine and a second underwriter on a held out sample
- Loans funded per underwriter, tracked with quality
Human in the loop
Licensed underwriters decide referrals, exceptions and declines, and review a weekly sample of automated approvals. The credit team approves guideline changes, and quality control compares automated and manual loans for defects and repurchase demands.
Common failure modes
- Guideline drift
- Investors can change overlays often. An engine running on last quarter's rules approves loans that no buyer will take. Tie each guideline to an owner and a review date.
- Speed reported without quality
- Productivity per underwriter rises when more files are automated, even if more of them are defective. Report defects and repurchases alongside throughput.
- Unexplainable decline
- A decision that cannot be turned into specific reasons for the borrower fails adverse action and fair lending review. Keep the rule trail for every decision.
- Chat front end hides the control point
- When loan officers query the engine from a general purpose assistant, access, data retention and logging move outside the loan origination system. Keep decisions in the system of record.
What are the risks and rules?
EU AI Act
High risk
Annex III point 5(b) lists AI systems used to evaluate the creditworthiness of natural persons or establish their credit score. A system that approves, refers or conditions a mortgage for a consumer falls under it, with risk management, data governance, logging, human oversight and conformity assessment obligations.
Rules that apply
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 5(b) names creditworthiness evaluation of natural persons as high risk.
- Guidelines on loan origination and monitoring (EBA/GL/2020/06) (European Banking Authority, Europe). Sets expectations for credit decisions, automated models and creditworthiness assessment in EU lending.
- Consumer Financial Protection Circular 2022-03: adverse action notification requirements for credit decisions based on complex algorithms (issued May 2022, withdrawn May 2025) (Consumer Financial Protection Bureau, North America). States that a creditor using a complex algorithm must still give specific and accurate reasons for an adverse action. The circular was withdrawn in 2025, and the underlying Equal Credit Opportunity Act and Regulation B requirement still applies.
Controls to put in place
- AI inventory entry with an accountable credit and compliance owner
- Model risk validation of the decision rules and readers, with periodic revalidation
- Versioned guideline and overlay library with change log and test set
- Decision record per loan: facts, document sources, guideline version and reasons
- Fair lending monitoring of approvals, referrals, conditions and declines
- Vendor and model provider due diligence, including data retention terms
Frequently asked questions
- How fast can AI underwrite a mortgage?
Better says its Tinman engine can complete the process in as little as 47 seconds, with a median of 2 minutes and 24 seconds, against an average of 21 days that it gives for the traditional mortgage industry. These are the company's figures from its March 2026 launch and are not an independent measure. The 21 days is an industry wide figure from the company, not a like for like comparison with the engine's timing.
- Does it replace underwriters?
Not in the figures reported here. Better reports that NEO Home Loans increased loans funded per underwriter by nearly 50% within six months of rolling out Tinman, which means the underwriters are still there and handle more loans. NEO's website also repeats Better's description of Tinman as automating underwriting for nearly 40% of loan files, a product claim and not a figure measured at NEO.
- Is AI mortgage underwriting high risk under the EU AI Act?
Yes for consumer lending. Annex III point 5(b) lists AI used to evaluate the creditworthiness of natural persons. The lender needs risk management, data governance, logging, human oversight and a conformity assessment, and in the EU the EBA loan origination guidelines also apply.
- What happens when the engine declines a borrower?
The lender still owes the borrower specific reasons. In the United States the Equal Credit Opportunity Act and Regulation B require them even when a complex model made the decision, so the engine has to keep the rule trail that produces the reasons.
How to cite this page
Blits.ai AI Use Case Library, "AI mortgage underwriting decision engine", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/mortgage-underwriting-decision-engine. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published