On this page (12 sections)
You can build a real-time insurance underwriting pipeline that returns calibrated risk probabilities in roughly 120 milliseconds from Europe, without generating text or risking hallucinated labels. Laya Studio is a hosted decision API for the open-source Laya System 1 model. You send structured applicant data and typed questions, and it returns a calibrated probability for each question in one forward pass.
Key takeaways
- Calibrated probabilities mean a 0.70 risk score reflects observed outcomes, not the model's confidence.
- The API runs in Switzerland with zero data retention, which supports GDPR and insurance data residency.
- The first 100,000 questions per month are free; after that, it costs $5 per million questions.
- It speaks the TypeSafe Jev
/v1/systemoneprotocol, so you can migrate without rewriting your client.
Why does real-time underwriting need a different stack?
Traditional underwriting cannot keep pace with instant quote expectations. It depends on human reviewers, static rule sets, and batch processes that take hours or days. A real-time pipeline must score risk the moment an application arrives, return a calibrated probability, and explain why—before the customer abandons the flow.
Sixfold describes its underwriting AI as covering commercial P&C and life and health, which shows the industry is already moving toward AI-assisted risk assessment at scale. Luca's 2026 guide on AI underwriting frames the shift as a transformation in how insurers and lenders assess risk and price policies, not a distant pilot.
The catch is speed. If your model takes seconds to score, you have lost the customer. If it returns a free-text label, you cannot audit it. You need a decision API that returns numbers, not narratives.
What is System 1 AI, and where does it fit in financial services?

System 1 AI means fast, single-pass decision-making. It does not deliberate, chain through steps, or generate text. It takes an input and returns a scored answer. In financial services, that pattern fits underwriting, fraud detection, and credit scoring because those decisions need to be fast, repeatable, and auditable.
We wrote about applying System 1 AI to real-time fraud detection in fintech, and the same logic applies here: the model answers a specific question about risk, not a general conversation. It is not a replacement for an underwriter's judgment on complex cases, but it handles the high-volume, routine decisions that clog a desk.
Step 1: Define the underwriting questions your API will answer

Before you call an API, write down the exact underwriting decisions as typed questions. Do not ask the model to "write an underwriting summary." Ask it to score risk, approve or decline, or pick a tier.
Use this checklist:
- Identify the decision points: auto-approve, risk tier, or referral to a human underwriter.
- Phrase each decision as a typed question:
yes_no,score, orchoice. - Define the allowed outputs for each question (for example,
standard,preferred,declined). - Set the probability threshold that triggers escalation to a human.
If you skip this step, you will end up parsing free text and arguing with the model about what it meant.
Step 2: Send the application payload to the decision API
Send a JSON payload with the applicant's structured data and your questions. The API returns calibrated probabilities in one forward pass, typically around 120 ms from Europe.
Example request:
Example response:
The API does not generate text, so it cannot invent a policy rationale or hallucinate a coverage decision. It returns probabilities only.
Step 3: Interpret calibrated probabilities and route the decision
A calibrated probability is a number you can act on. If the model says there is a 0.68 probability of high risk, that means roughly 68 out of 100 similar cases turned out high-risk in the data it was calibrated on. It is not a vague "confidence" score.
Use thresholds you control to turn probabilities into actions:
| Probability band | Action | Human review? |
|---|---|---|
| 0.00 – 0.30 | Auto-approve standard tier | No |
| 0.31 – 0.65 | Refer to underwriter | Yes |
| 0.66 – 1.00 | Decline or require manual review | Yes |
These thresholds are business rules, not model outputs. You can tighten or loosen them as your loss ratios change.
How do explainable AI and calibrated probabilities protect the underwriter?
Explainability means showing which inputs moved the probability, so an underwriter can audit and defend the decision. Calibration means the probability matches observed outcomes, which regulators and auditors can check. In a regulated industry, an unexplained decline is a compliance risk.
We have written about implementing explainable AI with calibrated probabilities in regulated environments. The core idea is that the model should make the underwriter's job easier to defend, not harder.
Where should underwriting data live to stay compliant?
Insurance applications contain personal and health data. Where that data is processed matters as much as the model's accuracy. Laya Studio runs in Switzerland with zero data retention, meaning applicant text is not stored after inference. That simplifies GDPR and data residency requirements for EU and UK insurers.
We covered the Swiss data advantage in the context of AI regulation beyond the EU AI Act. If your compliance team requires data to stay within specific borders, verify the hosting location before you integrate any classification API.
What metrics prove the underwriting pipeline is working?
Accuracy is not enough. Track latency, calibration, and operational impact.
| Metric | Target | Why it matters |
|---|---|---|
| p95 latency | Under 200 ms | Keeps the quote flow interactive |
| Calibration error | Under 0.05 | Predicted probabilities match real outcomes |
| Escalation rate | Under 15% | Most routine decisions are automated |
| Agreement with human underwriters | Over 85% | The model reflects your underwriting philosophy |
If your p95 latency creeps above your quote SLA, the pipeline is too slow regardless of how accurate it is. If calibration error is high, your thresholds are pricing risk incorrectly.
Can the same pipeline handle claims processing?
Yes. The same decision API can score claim severity, fraud likelihood, or routing priority at first notice of loss. It is not an underwriting-only tool; it is a general decision API for structured risk questions. AI in claims processing follows the same pattern: structured input, typed questions, calibrated output.
How do I scale without breaking latency or budget?
Scale by sending parallel requests and caching static applicant attributes that do not change between quotes. Pricing is per question: 100,000 free per month, then $5 per million. An insurer writing five million quotes a year would spend about $25,000 on classification after the free tier, before engineering overhead.
For a full cost comparison across Laya Studio, Jev, OpenAI, Google, and AWS, see our 2026 pricing comparison. If you are already on TypeSafe Jev, the migration path is straightforward because of protocol compatibility.
If you are ready to replace slow, opaque underwriting rules with calibrated, explainable decisions, explore the Laya Studio decision API.
FAQ
What is real-time AI insurance underwriting?
It is the use of a fast decision model to score insurance risk the moment an application arrives, returning calibrated probabilities that drive auto-approval, referral, or decline decisions without making the customer wait.
How do calibrated probabilities work in insurance?
A calibrated probability means the number reflects observed frequency. If 100 applicants receive a 0.70 risk score, about 70 of them should turn out to be high-risk. This lets underwriters set thresholds based on actual loss expectations.
Is this API suitable for claims processing?
Yes. The same structured decision API can score claim severity, fraud likelihood, or routing priority. It is not limited to underwriting; any structured risk question works.
Where is underwriting data processed?
Laya Studio runs in Switzerland with zero data retention. Applicant data is not stored after inference, which helps with GDPR and residency requirements.
How does this compare to using an LLM for underwriting?
An LLM generates text, which can hallucinate labels and is hard to calibrate. This API returns probabilities only, does not generate text, and is designed for structured underwriting decisions. For cost and latency, see our comparison of specialized versus general models.
Sources
Topics
- real-time AI insurance underwriting
- AI for insurance risk assessment
- automated underwriting with AI
- calibrated probabilities in insurance
- low-latency decision API insurance
- System 1 AI financial services
- explainable AI for insurance decisions
- real-time policy pricing AI
