AI governance belongs inside the underwriting workflow—not beside it.

AI in insurance has moved beyond experimentation. The NAIC says insurers are using AI in areas including renewal evaluation and inspection, while its Big Data and Artificial Intelligence Working Group has been developing an evaluation tool for regulatory examinations.

For MGAs, the implication is practical. A model can influence a delegated underwriting decision even when it does not make the final decision. If its output changes triage, referral, pricing inputs or the information shown to an underwriter, it belongs inside the control environment.

Begin with the decision, not the tool

Governance inventories often start with vendors and models. An underwriting-led approach starts with decisions: what decision is being supported, whose authority applies, which evidence is retained, and what happens when the system is wrong or unavailable.

This framing keeps governance proportional. A summarization aid and an automated eligibility determination should not receive identical treatment.

A minimum viable control set

Leadership teams do not need to wait for a perfect enterprise framework. They can establish a practical baseline now.

• Name an accountable business owner for each use case.
• Record data sources and material transformations.
• Define when human review is required.
• Retain enough evidence to reconstruct consequential decisions.
• Monitor drift, overrides and unexpected outcomes.
• Include vendors and models in change-management processes.

The hard boundary is not “human in the loop”

A nominal human approval does not make an AI-assisted process well governed. If the interface anchors judgment, suppresses contradictory evidence or makes an override burdensome, the system may exert more practical authority than the person signing off.

Governance should therefore test the decision environment, not merely the organization chart. Leaders need to know what information the underwriter saw, what the model omitted, how often people disagreed and whether disagreement changed the outcome.

The counterpoint: controls can overwhelm low-risk uses

Not every summarization, extraction or workflow aid deserves the same control burden as pricing, eligibility or claims decisions. Treating all AI use as equally consequential encourages shadow adoption and directs assurance resources away from the systems that matter most.

MGA Index expects the durable approach to be decision-based tiering: controls rise with the system’s influence on customer outcomes, delegated authority and financial exposure.

• Classify the decision before classifying the technology.
• Measure automation bias as well as model accuracy.
• Give underwriters a visible and reviewable path to disagree.

Regulatory attention is moving toward examination evidence

The NAIC says its AI Systems Evaluation Tool is intended to help regulators gather information about insurer use, governance, risk mitigation, potentially higher-risk models and data inputs. As of March 2026, twelve states were piloting the tool. A separate working group is developing a framework for third-party data and models used in property and casualty pricing and underwriting.

The practical implication is that a policy document will not be enough. An insurer may need to show which systems influence regulated decisions, how risks are classified, what validation was performed and how the organization monitors outcomes. An MGA-operated workflow can become part of that record when it supports delegated decisions.

Vendor oversight cannot be outsourced to a contract

Contractual representations matter, but they do not reveal how a tool performs on the MGA’s documents, classes and exception patterns. Validation should compare automated and manual performance on the same samples and weight errors by underwriting consequence. A wrong mailing address and a wrong limit do not belong in one undifferentiated accuracy rate.

The operating record should also capture material human repairs. When underwriters repeatedly correct output without recording the intervention, the organization loses both its earliest warning of model weakness and its best source of improvement data.

Proportionality is part of good governance

The countercase is bureaucracy. Treating summarization, extraction, eligibility and pricing systems as equally consequential can slow benign use and push adoption into unapproved channels. Controls should follow the decision affected, the scale of use, the difficulty of reversal and the potential customer or financial consequence.

This approach makes “human in the loop” testable. The question becomes whether the person had relevant information, time, authority and a usable path to disagree—not whether a nominal approval box existed.