Insurance and lending
AI underwriting: how it is used, where it goes wrong and what to control
A plain guide for underwriting leaders, risk owners and compliance teams at insurers and lenders. It explains what AI underwriting covers, where it helps, where it fails, and the controls to agree before a model or agent touches a decision.
Artificial intelligence underwriting means using machine learning models and AI agents to support the decision to accept, price or decline a risk, whether that is an insurance policy or a loan. In practice the AI more often prepares the decision than makes it: reading applications and documents, pulling data together, scoring or ranking cases, and drafting notes for an underwriter. AI in insurance underwriting and in mortgage lending faces the same core questions. Who owns the system, can you explain a decision to an applicant, and have you tested for unfair outcomes? Start with preparation tasks where an underwriter still decides, and widen the AI's role only after the checks work.
plain answer
What is AI underwriting?
Underwriting is the work of deciding whether to take on a risk and on what terms. For an insurer that means whether to write a policy and at what price. For a lender, including a mortgage lender, it means whether to approve a loan and on what terms. AI underwriting uses models and AI agents inside that work.
The AI might be a scoring model trained on past decisions, a language model that reads and summarizes a file, or an agent that gathers data from several systems and prepares a case. Often a person still makes the final call. How much of the decision the system is allowed to make is the key design choice.
use cases
Where is AI used in insurance and mortgage underwriting?
These are common uses, with the part a person should still own.
| Use | What AI does | What a person should still own |
|---|---|---|
| Document intake | Reads applications, statements and supporting documents and pulls out the fields. | Checking unclear or missing items, and sampling the extraction for errors. |
| Data gathering | Collects data from internal systems and approved outside sources into one case file. | Deciding which sources are allowed and acting on conflicts between them. |
| Triage and ranking | Sorts cases by complexity or likely fit so simple ones move faster. | The rules for what counts as simple, and review of cases the system routes away. |
| Risk scoring | Produces a score or recommendation from the case data. | The decision, the reasons given to the applicant, and fairness testing. |
| Case notes and summaries | Drafts a summary of the file and the reasoning for the underwriter. | Reading the source material for anything the summary leaves out. |
risks
What goes wrong with AI in underwriting?
Four problems come up again and again. First, unfair outcomes. A model trained on past decisions can carry past patterns forward, and a data field that looks neutral can stand in for a protected characteristic. Second, decisions nobody can explain. If an applicant or a regulator asks why a case was declined, the team needs more than a score. Third, no clear owner, so results drift and nobody notices. Fourth, outside models and data vendors adopted without review, so the insurer or lender is relying on a system it does not understand.
Regulators are paying attention to this. The National Association of Insurance Commissioners adopted a model bulletin on insurers' use of AI systems in December 2023. It says decisions or actions that affect consumers and are made or supported by AI systems must comply with applicable insurance laws, including those on unfair trade practices and unfair discrimination. It names underwriting and pricing among the stages of the insurance life cycle where AI is used, and it lists inaccuracy, unfair discrimination, data vulnerability and lack of transparency and explainability as risks to consumers. It is a model for states to adapt, so check what your state requires. Lenders face their own rules on credit decisions and should confirm them with counsel.
controls
What controls should an insurer or lender have before using AI in underwriting?
A small set, agreed in writing, covers most of the risk.
1. An inventory
A list of every model, agent and AI tool used in underwriting, with an owner for each and the data it uses.
2. Defined limits
For each use, what the system may do alone, what needs an underwriter's approval, and what it may never do.
3. Explainable reasons
A way to give the reasons for a decision in plain terms to an applicant, a regulator or an auditor.
4. Fairness testing
Tests for uneven outcomes across groups, run before release and again on a schedule, with a named person who acts on the results.
5. A record of actions
Logs of the inputs, the output and who approved, kept long enough to answer a later question.
6. Vendor review
A check of any outside model or data source, including what it was trained on and what your contract says about your data.
7. A way to switch off
A tested step that stops the system and moves the work to a person-run process.
kriv fit
Where does Kriv AI fit?
Kriv AI helps regulated teams put AI to work with controls designed in from the start. For underwriting, that usually means picking one use, mapping what the system may and may not do, setting up the record and review steps, and training the team that will run it. Our pages on responsible AI in insurance underwriting and on bias in AI underwriting go deeper on those two topics. We scope work after a short discovery call.
Straight answers
Frequently asked questions about AI underwriting: how it is used, where it goes wrong and what to control
What is AI underwriting?
It is the use of machine learning models and AI agents to support the decision to accept, price or decline an insurance risk or a loan. The AI often prepares the case for an underwriter instead of deciding it.
How is AI used in insurance underwriting?
Common uses are reading applications and documents, gathering data into one file, sorting cases by complexity, scoring risk and drafting case notes.
Can AI make the final underwriting decision?
It can be set up that way, but the decision then carries the most risk. Many teams keep an underwriter as the decision maker until the checks have a track record, and they confirm what their regulator expects.
What is the main risk of AI in mortgage lending?
Unfair or unexplained decisions. A model can carry past patterns forward, and a lender must be able to give the reasons for a decision. Confirm the credit rules that apply to you with counsel.
Where should an insurer or lender start?
With a preparation task such as document intake or case summaries, with a named owner and a sample review by an underwriter. Widen scope only after the checks are working.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call