How Should an Insurance Broker Use AI While Staying Compliant in 2026?

What the Compliance Question Actually Asks

An insurance broker using artificial intelligence is not automating only quoting, renewal searches, or document processing. The system may collect personal information, recommend coverage, draft customer communications, compare carriers, submit applications, service claims, and influence the allocation of professional responsibility. Each function creates a different combination of privacy, insurance, advertising, cybersecurity, consumer-protection, and licensing duties. The direct answer is that a broker may use AI, but should do so under a documented governance process tied to the broker’s licensed activities and the specific jurisdictions where customers are located. Human review remains the most defensible control for decisions that require licensed judgment, an authorized signature, or a material explanation to a customer.

Also worth reading: What is AI legal broker liability insurance and how does it protect technology intermediaries? · What is the definitive AI cyber insurance broker checklist for 2026? · How Should an AI Agent Coverage Review Evaluate Insurance and Legal Risk in 2026?

There is no single US rule called “AI insurance broker compliance,” and no federal AI-insurance-broker switch that becomes active on a particular date in 2026. Compliance instead comes from existing laws applied to new technology: state insurance codes, the NAIC’s model bulletin on AI systems by insurance companies and producers, privacy and biometric laws, the Gramm-Leach-Bliley Act when applicable, state data-breach duties, unfair-deceptive-practices rules, and the broker’s contractual obligations. Because the research date is September 28, 2026, counsel should verify the enacted text and regulator guidance current on that date rather than rely on a vendor’s summary of the law.

The main risk is often described as “the algorithm made the recommendation.” That explanation is not a legal defense. A broker remains responsible for a recommendation communicated under the broker’s name, just as a company generally cannot avoid responsibility for an employee acting within assigned functions. A defensible system therefore records the information used, the recommendation produced, the review performed, the approval authority, and the reason a recommendation was changed. Automation can improve speed and consistency, but it does not transfer professional accountability to the model developer.

The Rules That Govern AI-Assisted Brokerage

State law is the starting point. Insurers and producers are ordinarily licensed and regulated at state level, although surplus lines, captive agents, reciprocity, and other arrangements can change the applicable regime. The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers and Producers in December 2023. It is a model framework, not automatically binding law in every state; its value is that it identifies governance expectations involving nondiscrimination, customer notice, data governance, testing, and documentation. A broker should map each system use to the licensing and recordkeeping rules in every state in which it operates.

Privacy law applies according to the information processed. The California Consumer Privacy Act, as amended by the CPRA, generally applies to covered businesses that meet its thresholds and handles categories such as identifiers, contact data, professional information, inferences, and sensitive personal information. California defines “sell,” “share,” and certain targeted-advertising practices in ways that may matter when broker data is supplied to insurers, lead networks, or analytics providers. The FTC Act can prohibit unfair or deceptive practices even when a statement is not technically false. Federal statutes such as the Fair Credit Reporting Act can apply if the broker makes consumer reports, obtains reports, or uses eligibility decisions based on such information.

Anti-discrimination testing may also be required when AI influences access to coverage, terms, pricing, or claims handling. General references to fairness testing are not enough; the broker should determine whether the use is subject to disparate-impact or discriminatory-pricing provisions in the relevant state or sector. A commercial general liability quote and an auto quote can invoke different rules. This is why a national model cannot safely replace a state-by-state legal inventory.

No general federal or state threshold makes AI use lawful below a certain number of customers. A one-person brokerage may still be covered if it handles regulated or sensitive information. Conversely, large use does not automatically trigger every AI statute. Compliance turns on conduct, data, scale, jurisdiction, and affected persons, not simply the label attached to a tool.

Where Human Review Matters Most

Human review should be strongest where a customer could suffer a material loss of money, coverage, privacy, or opportunity. Examples include declining or limiting coverage based on incomplete answers, selecting exclusions, quoting terms that differ from the actual policy, interpreting policy language, negotiating a claim position, handling cancellation or nonrenewal notices, and making eligibility decisions using protected or proxy characteristics. A licensed person should understand the recommendation and be able to explain its basis in policy language. Merely clicking “approve” after displaying dozens of generated suggestions is unlikely to provide meaningful review if the broker cannot validate the output.

Review intensity should reflect the consequence and reversibility of the action. A low-risk internal tag that organizes “for review” emails may need only sampling and an audit trail. Binding a customer to a policy, communicating a claim denial, or transferring sensitive data to a new vendor merits a defined approval process. The broker can set review levels, such as automatic use for low-risk clerical work, sampled review for routine recommendations, and transaction-by-transaction review for material decisions. These are internal governance choices rather than statutory safe harbors.

The reviewer must have enough time and information to exercise independent judgment. A process that pressures staff to approve thousands of AI outputs in a few minutes can become a rubber stamp. The interface should expose source policies, carrier terms, missing information, and the reasons for a recommendation. When the model is uncertain or documents conflict, the system should route the matter to a person instead of inventing a plausible answer.

Human review also does not solve every problem. It cannot make a biased dataset fair, cure unlawful data collection, or establish that an automated notice was received. The organization should test both the model and the workflow around it. A high approval rate may indicate productive automation, but it may also reveal that reviewers are not meaningfully challenging the system. Metrics should include override rates, error severity, customer corrections, complaints, disparate outcomes, and incidents rather than measuring only time saved.

Data Collection, Security, and Vendor Contracts

A broker should begin by classifying the information the AI tool will receive. Ordinary business data may include names, email addresses, home addresses, policy numbers, vehicle details, health-related information, financial records, and geolocation. Some of this information is sensitive even when it is not publicly available. The broker should minimize what is sent to a model, remove fields that do not improve the task, restrict raw records where a retrieval-based answer is sufficient, and avoid placing unnecessary special-category or government-identifier data into a general-purpose prompt.

Security controls should be proportionate to the sensitivity and volume of the information. At minimum, the broker should require encryption in transit and at rest, role-based access, multifactor authentication for privileged access, logging, secure deletion, vulnerability management, tested backups, and documented incident-response procedures. The exact controls should be supported by a risk analysis rather than a generic promise that the vendor uses “bank-grade security.” FTC guidance emphasizes reasonable security and the need to match safeguards to the sensitivity and volume of consumer information.

Vendor contracts should allocate responsibility for training data, model changes, subprocessors, retention, deletion, data location, government demands, intellectual property, output accuracy, security incidents, cooperation with regulators, and the broker’s right to retrieve records. A contract should state whether customer data is used to train shared or provider models and whether that use is opt-in, opt-out, or technically prohibited. Deletion promises should distinguish removal from active systems, backups, derived embeddings, and logs.

The NAIC model framework treats third-party service providers as important extensions of an insurer’s or producer’s governance. Brokers should maintain a vendor inventory, assess material providers before deployment, and review changes after upgrades. Subprocessor and model-provider changes can alter the system’s data practices without changing the broker’s customer-facing workflow. A broker should not assume that a contract with the immediate AI vendor automatically binds every upstream provider or cloud host.

A Practical Compliance Workflow

The first practical step is to inventory every AI use case, including tools embedded in email, customer relationship management, transcription, underwriting intake, social media, and claims software. Record the business purpose, customer population, data inputs, output, decision impact, hosting model, vendor, and states touched by the workflow. A spreadsheet may be adequate for a small team, while a regulated enterprise may need a formal system register. An unregistered spreadsheet tool is still a system creating legal and operational risk.

The next step is to perform a jurisdiction and legal assessment. The team should ask whether the tool is merely clerical or performs an activity reserved to a licensed producer, whether it communicates to the public on the broker’s behalf, whether it makes eligibility decisions, and whether customer information is sold, shared, disclosed, or profiled. Counsel or a qualified compliance professional should translate those findings into specific controls, owner assignments, and review periods rather than issuing one blanket approval.

Before launch, test the system on representative and adverse scenarios. Testing should include missing data, contradictory documents, unusual addresses, multilingual inputs, prompt injection inside uploaded files, stale policy versions, fabricated citations, and requests to circumvent underwriting rules. The organization should define acceptable error rates by task, with stricter thresholds for coverage-impacting decisions. A possible starting control is 100% review during the first 60 to 90 days, followed by risk-based sampling only after performance is demonstrated; that is an internal example, not a legal requirement.

Each material transaction should preserve an audit record containing the model and prompt version, relevant source documents, generated result, human reviewer, corrections, final communication, and timestamp. Complaints, errors, and incidents should be logged and fed back into testing. A policy requiring annual review may be insufficient for a vendor that changes its model or data practices every month. Higher-risk tools should be reassessed after a material update, new data source, workflow change, regulatory change, or security incident.

Comparison of Governance Approaches

A broker can choose among several AI governance models, but the labels are organizational descriptions rather than recognized legal exemptions. The right choice depends on customer impact, licensing obligations, available expertise, and how the AI is deployed. In many brokerages, a controlled copilot is easier to govern than an autonomous agent because the copilot drafts or recommends while an authorized person approves the customer-facing action.

An autonomous agent is not unlawful merely because it is autonomous, but it increases the number of systems authorized to affect a customer. Permission design becomes critical: the agent should not be able to bind coverage, issue final coverage interpretations, change beneficiary information, or export records merely because it can access the same platform as a licensed employee. “Human in the loop” should mean a person can understand, reject, and reverse the action before it becomes difficult or impossible to undo.

A traditional outsourced or employee-based service remains a valid alternative where a broker lacks the resources to govern AI. It may be slower and more expensive per transaction, but established workflows can make responsibility, training, and review clearer. Buying a specialized insurance AI platform may also reduce legal-model development work, although the broker still needs to assess the platform and supervise its outputs. No vendor’s SOC 2 report, insurance policy, or contractual warranty proves that a deployment complies with every state insurance rule.

Common Compliance Mistakes and Their Corrections

A frequent mistake is treating model use as an invisible productivity feature. Employees often add unapproved AI tools to a customer-management or document system, creating an unregistered processor of personal and policy data. The correction is to establish an approved-tool process, restrict access to reviewed services, and require business owners to obtain privacy, security, licensing, and vendor assessment before customer data is entered. Discipline should be consistent, but a control that merely threatens punishment will not uncover shadow use. A simple reporting channel and safe migration path can encourage disclosure.

Another mistake is assuming that citations in a generated answer are authoritative. AI systems may cite nonexistent cases, obsolete regulations, incorrect policy sections, or real sources in the wrong context. Every external legal or policy statement used in customer work should be checked against the controlling source. This rule is especially important when comparing exclusions, deductibles, endorsement language, or renewal requirements. A fluent answer has no evidentiary advantage over a verified document.

Brokerages also confuse a disclaimer with informed consent. A footer saying “decisions are made by AI” may not explain what data is used, who reviews the result, or how a person can obtain assistance. Notices should be clear, timely, accessible, and tailored to the actual practice. Customers should not be required to waive statutory rights as a condition of using an ordinary service.

The most serious strategic mistake is allowing the AI to operate across jurisdictions through one global prompt while local rules remain unknown. Auto insurance, health coverage, workers’ compensation, commercial property, life insurance, and surplus lines may implicate different regulators and restrictions. A system approved for one workflow should not silently expand into regulated advice or another state without a documented reassessment.

When to Act and What Implementation May Cost

A broker should act before uploading real customer data or allowing AI to contact a customer. A short pre-launch gate of 2 to 4 weeks may be realistic for a small brokerage evaluating a low-risk internal copilot, but that is an implementation estimate, not a regulatory deadline. A multi-state deployment using sensitive health, financial, or identity data can require several months of legal analysis, contracting, security review, integration, employee training, and validation. The delay is justified when the alternative is exposing regulated data to an unapproved system.

The broker should pause deployment after a material incident, such as a fabricated policy explanation, unauthorized disclosure, access by a former employee, discriminatory recommendation, incorrect coverage advice, or customer complaint showing that review failed. The response should contain the problem, preserve logs, correct customer harm, notify affected persons or regulators where required, and determine whether the system caused the event or merely surfaced an existing data error. Voluntary regulator contact should be based on the facts and applicable reporting rules, not an assumption that every AI error must be disclosed publicly.

Pricing varies mainly with integration, data volume, model usage, security requirements, and vendor type. Public freemium copilots may cost an individual user $0 to roughly $20 to $30 per month, while enterprise seats, API usage, and storage can run from tens of thousands to several million dollars annually. A specialist insurance workflow platform may add implementation and data-migration fees to subscription or per-policy pricing. External legal reviews commonly depend on scope and jurisdictions, so fixed figures would be misleading.

The most reliable budget line is not the software license but the full governance cost: staff training, audits, evaluation datasets, model monitoring, incident response, privacy notices, contract review, and licensed-person review time. A cheap tool that requires expensive manual correction is not economical, and an expensive platform is not compliant by default. Before purchase, ask for measurable service levels, data-use restrictions, audit rights, deletion terms, security evidence, model-change notice, and examples of insurance-specific evaluation.

The Defensible Standard for AI-Assisted Brokerage

The best compliance posture is controlled, explainable, and proportional automation. A broker should know exactly where AI appears in the workflow, what information it receives, what recommendation it produces, and which person remains accountable. Customers should receive accurate information in language they can understand, while sensitive records remain protected through access, security, retention, and vendor controls. The organization should be able to reconstruct a material decision months later and explain both the model’s contribution and the human approval.

No percentage of AI use can be quoted as universally safe. A 10% automated workflow may still be unlawful if it screens protected classes without a valid basis, while a much larger automation program may be defensible if every action is low risk, independently verified, and tightly bounded. Claims such as “AI is 100% compliant” or “human oversight guarantees compliance” should therefore be treated as marketing statements requiring evidence.

For most brokerages, the prudent sequence is to begin with low-risk internal assistance, establish a register and approval process, test against real failure modes, and expand only after documented performance. If a tool creates or changes insurance coverage, communicates an authoritative interpretation, handles regulated data, or acts without readily reversible human control, the broker should obtain jurisdiction-specific legal advice and maintain a licensed reviewer. That approach does not eliminate AI risk; it makes the remaining risk measurable and manageable.