Blog Industries

AI for insurance brokers: use cases and rules (2026).

How brokers and MGAs use AI on submissions, quotes, renewals and client questions, where claims and underwriting fit in, and the rules that apply to AI in insurance.

Fig. 0A submission read, checked and handed to a broker

For insurance brokers and MGAs, the first AI system worth building is usually submission intake: reading the emails, ACORD forms, schedules and loss runs that arrive in every format, checking them for gaps, and loading clean data into your system for a person to review. Renewals and client coverage questions come next. Claims and underwriting benefit too, but the quickest wins sit in the broker’s inbox.

This guide covers where AI in insurance pays off for distribution, where it’s the wrong tool, the rules that apply as of October 2026, and how to start.

Where does AI help brokers and MGAs most?

Submissions and clearance

A commercial submission is rarely one document. It’s an email with an ACORD application, a statement of values (SOV), several years of loss runs, a fleet or driver list and a broker’s covering note. An AI system reads all of it, extracts the risk details, and checks them against each other: a location on the SOV that isn’t on the application, a loss run that stops a year early, a revenue figure that differs between documents. It then drafts the questions back to the client or retail broker. For an MGA, it can also check the submission against the binder’s appetite and referral rules before an underwriter opens it.

Quotes and comparison

Once the risk data is clean, an agent can prepare the quote requests, enter them into carrier portals or rating tools where that’s possible, and read the quotes that come back. It then lays the terms side by side against what the client asked for: limits, deductibles, exclusions and endorsements. The broker still advises. The system’s job is to stop anyone retyping data into five portals and to stop a missing exclusion going unnoticed in a long quote PDF. Selling and following up with prospects is covered in AI for sales teams.

Renewals

Renewal work is a calendar of small, repeated tasks: pull the expiring policy, ask the client what has changed, chase the answers, remarket if needed, and send the renewal terms with the changes explained. An agent can run the chasing, compare expiring and renewal terms clause by clause, and draft the covering email. A large book of small accounts is a good fit, because the volume is high and the steps are the same each year.

Client and carrier questions

Clients ask “am I covered for…?”, “can you send my certificate?” and “what changed at renewal?” An assistant on email, web chat or WhatsApp can answer from the actual policy wording and endorsements, quote the clause it relied on, and issue routine documents such as certificates of insurance. It hands over to a person when the answer turns on judgement, when the client is upset, or when the question is really a claim. The design it needs is the one in our guide to retrieval that admits when it doesn’t know. For the full support picture, see AI customer service.

Mid-term changes and bordereaux

Mid-term adjustments, endorsement requests and change-of-address emails are structured work in free-text form. For MGAs, the same applies to bordereaux: reading what comes in from coverholders and sub-agents, validating it against the binder’s terms, and flagging the lines that break them before the report goes to the capacity provider.

Where do claims and underwriting fit?

They’re supporting sections for a broker or MGA, but they matter because the same document reading carries over.

  • Claims intake (first notice of loss). An agent takes the first report by web, app, WhatsApp or email at any hour. It asks the questions for that claim type, collects photos and documents, checks the policy is in force, and opens a complete file. A broker handling claims for clients, or an MGA with delegated claims authority, starts with everything needed instead of chasing it.
  • Claims documents. Repair estimates, invoices, medical reports and police reports are read and cross-checked against the policy and each other. Inconsistencies, such as a repair date before the incident or an invoice that doesn’t match the estimate, go to a person.
  • Underwriting triage. For MGAs with delegated authority, the system can sort submissions by appetite, summarise the risk and draft referral notes. The underwriter prices and decides.
  • Fraud signals. Machine-learning models trained on claims history flag suspicious patterns. Language models help by spotting inconsistencies in documents. Both should send cases to an investigator, never decline them.

Which kind of AI fits which job?

Not every insurance problem needs a language model. This table maps the main broker and MGA jobs to the right tool.

Use caseRight kind of AIData it needsMain risk
Submission intake and gap checkingDocument AI with an LLM for the covering email and questionsPast submissions, ACORD forms, SOVs, loss runs, your appetite rulesA wrong value loaded into the system and priced on
Quote comparisonDocument AI plus rules that check terms against the client’s requestQuote PDFs, the client’s requirements, carrier termsA missed exclusion or sub-limit
Renewal chasing and renewal packsLLM agent with limits on what it can sendExpiring policy, client contact history, agency management system dataWrong or stale client details sent out
“Am I covered?” questionsRetrieval over the wordings and endorsements, with clause referencesFull policy documents per client, current endorsementsA confident answer the wording doesn’t support
Rating and pricing individualsActuarial and machine-learning models, with an LLM only to explainHistoric claims and exposure dataUnfair outcomes and regulatory scrutiny
Fraud screeningMachine-learning anomaly models, with documents read by an LLMClaims history, document metadataFalse positives treated as decisions
Bordereaux validationDocument AI plus deterministic validation rulesBinder terms, coverholder reportsBreaches reported late or not at all

Which systems does it need to connect to?

The AI is rarely the hard part. The work is connecting it to where the data lives.

  • Agency or broker management system. Platforms such as Applied Epic, Vertafore and EZLynx hold clients, policies, renewals and activity. Reading from and writing to them, through their APIs or integrations, is what turns an extraction into saved time.
  • Policy admin and rating. For MGAs and insurers, systems such as Guidewire and Duck Creek hold the policy and claim record. Integration with these is typically most of the project.
  • Carrier portals. Many have no API for quoting, so some steps stay with a person or run through tightly controlled automation.
  • Email and document stores. Most submissions start in a shared inbox, so mailbox access and clean filing of attachments matter as much as the model.

An AI system sits alongside these. It doesn’t replace your agency management system.

Where should a person stay in charge?

Some steps carry an advice, coverage or fairness duty, and those should keep a named person behind them. The simplest way to design this is by autonomy level.

LevelWhat the system doesExamples
Acts aloneCompletes the task within limits set in codeFiling attachments, sending certificates from a template, answering status questions
Drafts, a person approvesPrepares the output and a person sends or confirms itRenewal emails, quote comparisons, questions back to a client
Prepares, a person decidesGathers the facts and explains the optionsCoverage questions that depend on interpretation, underwriting referrals, claims decisions
Stays with a personNo AI decisionAdvice on what cover to buy, declining a risk, binding, pricing an individual

Three more habits matter in practice. The limits belong in code, not in the prompt: the system should not be able to send binding language or confirm cover it can’t find in the wording. Inbound emails and attachments are untrusted input, so a hostile PDF can’t steer the agent (see prompt injection is an input-validation problem). And every output that touched a decision needs a record of the inputs, the system’s output and the person who decided.

Which rules apply to AI in insurance?

These rules change, so the dates below are as of October 2026, and a qualified professional should decide how they apply to your firm.

  • United States. The NAIC’s model bulletin on the use of AI systems by insurers, adopted by the NAIC in December 2023 for state insurance departments to issue, sets out what regulators expect of insurers: a written AI program, governance, controls that scale with the potential harm to consumers, and oversight of AI and data supplied by third parties. Regulators may ask for documentation during an investigation or examination. It is addressed to licensed insurers, so a broker is most likely to meet it through carrier and MGA partners and their vendor due diligence. Colorado’s SB26-189, which replaces its earlier AI Act, covers automated decision-making technology used in consequential decisions, insurance included. Its obligations phase in, with developer duties and the attorney general’s disclosure rules due by 1 January 2027. Deployers must give notice at the point of interaction and a plain-language description of the system’s role within 30 days of an adverse decision, and consumers can correct data and ask for human review. The California Privacy Protection Agency says businesses using automated decision-making technology for significant decisions must comply with its ADMT requirements from 1 January 2027.
  • United Kingdom. The FCA says it does not plan extra regulation for AI and relies on existing frameworks, including the Consumer Duty and the Senior Managers and Certification Regime, as it does for banks and lenders. For an FCA-regulated broker, the practical point is that AI-assisted processes are judged against the same customer-outcome and accountability expectations as manual ones. Under the UK GDPR as amended by the Data (Use and Access) Act, the ICO’s guidance on automated decision-making (still in draft, updated 31 March 2026) covers decisions based solely on automated processing with a significant effect, including enabling people to make representations, human intervention and the ability to contest. Our guide to UK AI rules goes further.
  • European Union. AI used for risk assessment and pricing of natural persons in life and health insurance is high-risk under Annex III of the AI Act. The Commission’s implementation timeline shows the Annex III rules applying from 2 December 2027, after the Digital Omnibus changes, and transparency rules under Article 50 from 2 August 2026. Article 50 requires systems that interact directly with people to be designed so those people are told they’re dealing with AI, unless that is obvious from the context, which is why we build client-facing assistants to say so. EIOPA’s opinion on AI governance and risk management (6 August 2025) is addressed to national supervisors and takes a risk-based, proportionate approach to applying existing insurance-sector rules to AI. See our EU AI Act guide.

A submission or renewal assistant that doesn’t score or price individuals isn’t what that Annex III entry describes. A tool that does is a different matter, and worth a legal opinion before you build it.

What makes insurance AI trustworthy?

RequirementHow it’s built
Answers grounded in the policyRetrieval over the actual wordings and endorsements, with clause references, and “I can’t confirm that” when the wording doesn’t say
Accurate extractionConfidence checks on every field, cross-checks across the application, SOV and loss runs, and review for anything uncertain
Fair outcomesOutcome monitoring by customer group before launch and in production
ExplainabilityA record of the inputs, the system’s output and the person who decided, for every decision it touched
Fraud resistanceUploaded documents and messages treated as untrusted input

How do you start, and what do you measure?

Pick the flow with high volume, a clear number and a person reviewing the output. For most brokers that’s submission intake or renewals. For MGAs it’s submission triage.

  1. Choose one line of business and one flow. For example, commercial property submissions or small-account renewals.
  2. Collect a test set. A few hundred past submissions or renewals, including the messy ones, with the correct answers filled in by your team.
  3. Build the reading and checking first. Extraction, cross-checks and gap questions, with every uncertain field going to a person.
  4. Run it beside your team. Compare the system’s output with what staff would have done, and fix the failures before it handles live work.
  5. Add the next step only when the numbers hold. Quote entry, renewal chasing or the client assistant, in that order.

Track these:

  • Turnaround: time from submission received to a complete, clean file.
  • First-pass completeness: the share of submissions that need no follow-up questions.
  • Extraction accuracy: fields corrected by a person, per submission.
  • Renewal response time: days from the first renewal notice to terms sent.
  • Answers without a person: the share of client questions resolved correctly with no handover, checked on a sample.

The broker’s best day is one where every submission arrives complete. That’s a realistic first goal for AI.

Questions brokers and MGAs ask

Should we build or buy?

Buy for common, standardised tasks. Build when the system has to work with your own wordings, systems, appetite rules and workflow, or when speed on submissions is part of how you compete. Our services page covers what we build.

Can an AI assistant tell a client they’re covered?

It can quote the relevant wording and the clause it came from, and say when the wording is silent. A statement that a particular loss is covered is a coverage position, which should come from a person. Design the assistant so it can’t make that statement on its own.

Can AI detect fraud?

Machine-learning models trained on your claims history flag unusual patterns, and language models spot inconsistencies in documents. Both should send cases to an investigator, not decline them.

Does it work with our agency management system?

Usually, through its APIs, integrations, document stores and email. Integration with the management or policy system is typically most of the work.

Working in insurance?

Modulus Labs AI builds submission and renewal agents, document AI and policy assistants for brokers, MGAs and insurers, with decision records and human review built in. Tell us where the paperwork piles up, and we’ll tell you what we’d build first.