---
title: How do you govern an AI agent that extracts data from documents for underwriting?
url: https://opencontroller.lyzr.com/blog/govern-document-extraction-agents-in-underwriting
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How do you govern an AI agent that extracts data from documents for underwriting?

**Classify each document before you extract anything, set a confidence threshold per field rather than per document, check extracted values against master data, and route anything below threshold or in conflict to a human reviewer whose decision is recorded with the agent version and evidence. The agent proposes values; a rule or a person decides whether they enter the underwriting file.**

Document extraction is one of the most common first agents in insurance and lending. It reads applications, medical reports, financial statements or broker submissions, pulls out fields and fills a system of record. It looks low risk because a person still makes the underwriting decision. The risk sits in the data that person sees. This guide is part of our [rogue agent and runtime control guide](/blog/how-to-stop-a-rogue-ai-agent).

## Why classify documents before extraction?

An extraction prompt tuned for a bank statement will still return something when given a payslip. It will just be wrong in confident ways. A classification step that labels the document type first, and rejects types the agent was never tested on, removes a whole class of silent errors.

Make classification its own step with its own threshold. If the classifier is not sure what a document is, stop and send it to a person. Do not let the extractor guess.

## What confidence thresholds should you use?

Set thresholds per field, based on the cost of an error in that field. A wrong policyholder middle name is a nuisance. A wrong sum insured, date of birth or declared medical condition can change the price or the decision.

| Field type | Example | Suggested handling |
|---|---|---|
| Identity fields | Name, date of birth, ID number | High threshold; always check against master data |
| Financial values | Income, sum insured, revenue | High threshold; cross-check totals and units |
| Risk declarations | Medical history, prior claims | Human review for every value, regardless of score |
| Descriptive fields | Occupation, address | Moderate threshold; spot-check samples |
| Free text summaries | Underwriter notes | Labelled as AI-generated; never auto-accepted |

Calibrate the thresholds on a labelled test set before go-live and record the error rate at each threshold. A score is only useful if you know what it predicts. Keep an explicit "undecided" outcome for values that fall between accept and reject, so they cannot slip through.

## How should extracted values be checked?

Use deterministic checks wherever a source of truth exists:

1. **Compare against master data.** Customer records, policy administration data or broker files. A mismatch is a review trigger, not an automatic overwrite.
2. **Check internal consistency.** Totals should add up. Dates should be in order. Currencies and units should match the document.
3. **Check ranges.** Values outside plausible ranges for the product are flagged.
4. **Check provenance.** Every extracted value should point to the page and region it came from, so the reviewer can see the source.
5. **Check for missing fields.** A blank is better than an invented value. Test that the agent returns blanks when a field is absent.

These checks do not need a model. They are cheap, explainable and easy to evidence.

## What should a reviewer approval record contain?

When a person accepts, corrects or rejects a value, record:

- document ID and document type
- field name, extracted value and confidence score
- the master-data value it was compared with, if any
- the reviewer's decision and corrected value
- reviewer identity and timestamp
- the agent version, model version and prompt version that produced the value
- the reason code for any correction

This record is your evidence that a human was meaningfully in the loop. It also becomes training and evaluation data. A rising correction rate on one field is an early warning that something changed, often a new document template from a broker or a model update. See our guide to [evidence for human review of AI agent outputs](/blog/evidence-for-human-review-of-ai-agent-outputs).

## Which rules apply to underwriting agents?

In the US, the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted by the NAIC in December 2023 and since adopted by many states, expects insurers to maintain a written AI programme covering governance, risk management and third-party oversight. In the EU, the AI Act lists AI systems used for risk assessment and pricing in life and health insurance as high-risk in Annex III. Treat these as reasons to keep the records above, and check with counsel how they apply to your products. Our post on [AI governance evidence for insurers](/blog/ai-governance-evidence-for-insurers) goes further.

## How do you keep the agent safe in production?

Run evaluations continuously, not only before launch. Score a sample of production extractions against reviewer corrections every week. Watch three numbers: straight-through rate, correction rate per field, and the share of documents classified as unknown. Pin the model version and re-run the test set before any change. If the correction rate on a critical field crosses an agreed level, stop straight-through processing for that field until the cause is found.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents and versions, runs multi-turn and red-team evals, and uses LLM judges to score production traffic. Live groundedness and task-adherence guardrails flag outputs not supported by the source, and behavioural-deviation monitoring can alert on, quarantine or stop an agent across connected platforms. Policies are written as code. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can an extraction agent write directly into the policy system?**
Only for fields with calibrated thresholds and deterministic checks, and only when both pass. Risk declarations should always go through a person.

**What is a good straight-through processing rate?**
It depends on document quality and field risk. Set it from your calibration data, not a target, and report it alongside the correction rate.

**Does a human reviewer remove the need for AI governance?**
No. The reviewer only sees what the agent shows them. Classification, thresholds, provenance and version records are what make the review meaningful.

## Key facts

- Classify the document type before extraction and reject untested types.
- Set confidence thresholds per field by error cost; route risk declarations to a human every time.
- Check extracted values against master data, internal consistency, ranges and provenance.
- A reviewer approval record should hold the field, value, score, comparison, decision, reviewer, timestamp and agent, model and prompt versions.
- The NAIC Model Bulletin on insurers' use of AI was adopted by the NAIC in December 2023; the EU AI Act Annex III lists life and health insurance risk assessment and pricing as high-risk.
- Opencontroller runs evals, LLM judges on production traffic and live groundedness and task-adherence guardrails.
- Opencontroller starts from $250K per deployment per year.

## Related

- [How do you validate an AI agent's output before it acts?](https://opencontroller.lyzr.com/blog/validate-ai-agent-output-before-it-acts.md)
- [What evidence should you keep when a human approves an AI agent's output?](https://opencontroller.lyzr.com/blog/evidence-for-human-review-of-ai-agent-outputs.md)
- [What AI governance evidence do insurers need for regulators?](https://opencontroller.lyzr.com/blog/ai-governance-evidence-for-insurers.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
