AI automation works in insurance broking where the work is high-volume and the decision is not a regulated one: policy data intake, renewal follow-up, claims document assembly, and quote aggregation across carrier portals. It works badly the moment it touches advice. The constraint is rarely the model — it is that your data sits behind portals with no usable API, and that anything touching a quoting or renewal decision must leave an audit trail a regulator can read.
We build and operate this software for a living, so read the rest with that bias in mind. It includes the parts we tell brokers not to automate.
The adoption gap in broking is an integration gap, not a budget gap
The 2025 Best Practices Study, published by the Independent Insurance Agents & Brokers of America (Big "I") and Reagan Consulting, found that 84.2% of brokerages above $100 million in revenue had invested in generative AI, against 60% of firms in the $25–100 million band. That is reported in Risk & Insurance's coverage of the study (22 August 2025). The same study year recorded a record 10.7% organic growth and 26.1% EBITDA margins across Best Practices agencies, so this is not a story about who could afford it.
It is a story about who had someone to do the integration. The report cites a 2025 Gartner survey finding that 91% of high-maturity firms have appointed dedicated AI leaders. Large brokerages have an operations function that can own a system after launch. Smaller firms buy a tool, and it sits at the edge of the workflow because nobody owns the plumbing between it and the agency management system.
When one Best Practices agency did close that gap — deploying a platform across loss run analysis, coverage analysis, Statement of Value comparison, and an "Ask Anything" agent — the study reports 7.5 hours saved per staff member per week and a 244% return on investment. Reagan Consulting's research on Broker Tech Ventures partner firms separately found producers under 35 with tech-enabled tools carried books averaging $168,000 larger than peers without them. Real numbers from a named study — and not a forecast for your agency, because they came from a firm that solved the integration problem first.
Sort your workflows by evidence burden, not by hours saved
Every tools listicle in this category ranks workflows by time savings. In a regulated business that is the wrong first sort, because the workflows with the biggest time savings are frequently the ones carrying the heaviest documentation obligation, and the documentation obligation is what kills the project in month four.
Sort on two axes instead: how often the work happens, and whether the output touches a decision a regulator would care about — quoting, renewal, underwriting, or anything a client could reasonably read as advice.
The top-left quadrant is where you start. The top-right is where the value is — reachable, but only with logging built in from the first line of code rather than retrofitted after a market conduct question. The bottom-right is where careers go wrong.
The three workflows to automate end to end
Policy data intake. The same details get typed into the management system, a carrier portal, and a spreadsheet. This is pure data movement — extract, validate against the schema you already enforce, write to the system of record, escalate what fails. There is no judgment in it, which is exactly why it is safe to automate fully.
Renewal reminder and follow-up sequencing. Renewal outreach tracked in inboxes and spreadsheets is where books quietly leak. Sequencing reminders, document requests, and follow-ups off policy status is scheduling logic, not advice, provided the message templates are pre-approved and the automation never invents coverage language. Keep the "send" on a human where the client relationship warrants it; automate the tracking and the drafting underneath.
Claims document collection and bundling. Assembling the receipts, forms, and approvals a carrier requires for a submission is checklist work with a known target format. Automate the collection, the completeness check, and the packaging; leave the submission decision with the adjuster-facing broker.
For a comparison outside insurance: on a logistics shipment-intake workflow we built, reading inbound emails, validating details, and creating labels and bookings, the result was 95% less manual work with zero monthly entry errors. That is freight, not broking — same problem shape, different regulatory surface. We would rather name the domain than let the number imply a broking benchmark it is not.
The two workflows to automate with a decision log
Quote aggregation across carrier portals. Pulling rate slips from multiple insurer portals, normalising them into a comparable structure, and surfacing the shortlist is the highest-value automation available to most brokers. It also lands squarely on a regulated decision, because the ordering of that shortlist shapes what the client is offered. Build it so that every run records which portals were queried, what was returned, which normalisation rules were applied, and why an option ranked where it did. If you cannot reconstruct a comparison from six months ago, you have built a liability.
Compliance reporting. Aggregating data from the management system, formatting to the regulator's template, and routing through a QA step before filing is well-suited to automation precisely because the output format is fixed. Version the templates, keep the transformation logic readable, and keep a human sign-off on the filing itself.
Both of these are covered in more depth on our insurance broker automation briefing, which maps them alongside the KPIs worth tracking — renewal response rate inside 48 hours, quotes delivered per broker per week, and manual compliance hours avoided per reporting cycle.
Where automation should assist and never decide
Coverage recommendations, appetite judgements, and declinations stay with a licensed human. An assistant that retrieves the relevant policy language, pulls the prior-year comparison, and drafts the rationale is genuinely useful. A system that selects the recommendation and sends it is not a productivity tool, it is an unlicensed adviser with your firm's name on it. The test we use: if a client could act on the output without a broker reading it first, a broker reads it first. That rule costs less than one complaint.
The portal problem, and what to do about it
Most broking automation projects fail on access, not intelligence. Carrier portals were not built to be read by software, and agency management systems vary enormously in what they expose — some have workable APIs, some a subset of fields, some nothing useful at all.
Three honest options, in order of preference:
- Use the API where one exists and covers the fields. Confirm the field coverage before you design the workflow, not after. Document authentication, refresh frequency, and error handling per integration.
- Use browser automation where it does not. This is unglamorous and it is what actually ships. It needs credential handling, change detection when a portal's layout moves, and a defined failure path when a session breaks. Treat it as a maintained integration, not a script.
- Change the input. Sometimes the cheapest fix is asking a carrier for a scheduled export instead of scraping a screen. Ask before you build.
Working with the systems you already run — Applied Epic, EZLynx, whatever is in place — using browser automation when APIs do not exist is the core of our insurance automation work. No rip-and-replace, because a management system migration and an automation programme should never be the same project.
What the NAIC bulletin actually asks of you
The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers was adopted in December 2023. As of March 2025, 24 states had adopted it with little or no material change, according to a Quarles analysis published 2 April 2025, and adoption has continued since. It is principle-based rather than prescriptive: it does not mandate a testing methodology or an annual filing.
What it does ask for, in substance:
- A written programme governing responsible use of AI systems that make or support decisions in regulated insurance practices, with named accountability across business, compliance, legal, and data functions.
- Validation, testing, and retesting of outputs before and after deployment.
- Vendor diligence on third-party data and third-party AI systems, including contract terms that preserve audit rights and cooperation with regulatory inquiries.
- Consumer notice and documented procedures for responding when a regulator asks how a decision was reached.
Read the vendor-diligence point twice if you are buying a tool rather than building one: the obligation does not transfer to the vendor. If a vendor cannot tell you what model made a determination and on what inputs, you have inherited a problem you cannot answer for.
What it costs to run, not just to build
The build fee is the number everyone quotes and the smaller half of the total. Running cost in these systems has three layers: model usage, runtime infrastructure and orchestration, and the operations layer of human review and exception handling. Our breakdown of AI agent cost covers how those move and why cost per workflow is the only unit worth tracking.
In broking the operations layer is bigger than in most sectors, because the exception rate on portal-sourced data is high and the review requirement is regulatory rather than optional. Budget for it. A workflow that quietly needs 40 minutes of daily human review is not the workflow you were sold. To sanity-check the arithmetic first, our ROI calculator gives you a defensible starting frame.
We do not publish a rate card, because a quote-aggregation build across four carrier portals and one across fourteen are different projects. We would rather say that than imply a number we cannot stand behind for a workflow we have not seen.
A 90-day sequence that survives contact with production
- Days 1–15: pick one workflow from the top-left quadrant. Measure the current version honestly — volume per week, minutes per instance, current error rate. Without a baseline you cannot prove anything later, and proving it is how you get funded for the second workflow.
- Days 16–45: build with the human in the loop. The automation proposes, a person approves, and every approval and override is logged. You are collecting your exception taxonomy and your training evidence at the same time.
- Days 46–75: narrow the human gate. Where override rates are near zero on a defined class of input, let those run through. Where they are not, you have found the class of work that stays manual, which is a result, not a failure.
- Days 76–90: instrument and hand over. Ownership, monitoring, an alert when volume or exception rate moves, and a scheduled review. Then start the next workflow.
This mirrors the three-phase delivery we run everywhere: strategy and readiness, applied build with a human-in-the-loop QA gate, then production delivery with observability and exception handling. The same sequence appears in our accounting firm automation guide, for the same reason — regulated document workflows fail in the exception path, not the happy path.
How to tell a real implementation from a demo
Ask a prospective partner five questions. What happens when a carrier portal changes its layout overnight? Where is the decision log written and who can read it? What is the expected exception rate in month one, and who works the queue? What does this cost to run per month at our volume, not to build? What does the handover look like if we part ways?
A demo answers none of those. An implementation team answers all five without hedging. If you want to test us on them, book a strategy call — bring one slow workflow and we will map it, give you a straight feasibility read, and tell you if the honest answer is that it should stay manual.
Frequently asked questions
What can AI actually automate in an insurance brokerage today?
Reliably: policy data intake into the management system, renewal reminder and follow-up sequencing, claims document collection and bundling, quote aggregation across carrier portals, and compliance report assembly. Less reliably, and only with a human gate: coverage comparison summaries and client-facing drafting. Not at all: coverage recommendations, appetite decisions, and declinations.
Will automation work with our legacy agency management system?
Usually yes, but not always through an API. Where a management system exposes workable endpoints covering the fields you need, use them. Where it does not, secure browser automation against the interface your staff already use is the normal route, and it needs to be maintained as a real integration rather than a one-off script.
Does the NAIC AI model bulletin apply to brokers or only to insurers?
The bulletin is directed at insurers, but its practical reach extends through the distribution chain: insurers carry vendor-diligence obligations for third-party systems, and brokers running automation that touches quoting or renewal decisions get pulled into those documentation requirements. Check the position in each state you are licensed in — 24 states had adopted the bulletin with little or no material change as of March 2025, and more have since.
How long before an automated broking workflow pays for itself?
It depends on volume and on how much human review the workflow still needs after go-live. The one published figure worth anchoring on is the 244% ROI reported for a Best Practices agency in the 2025 study, alongside 7.5 hours saved per staff member per week — from a firm that deployed across four defined functions rather than buying a general-purpose tool.
Should we buy an insurance-specific AI tool or build a custom workflow?
Buy where the workflow is standard across firms and the tool integrates with your management system without custom glue. Build where the value is in your specific carrier mix, your appetite rules, or your service model — which is usually quote aggregation, because no two brokers carry the same panel.
What is the single most common reason these projects stall?
Nobody owns the system after launch. The pilot works, the person who championed it moves on, a portal changes, exceptions pile up unworked, and within two quarters the team is back to manual with a licence still being paid for. Assign ownership before you assign budget.
Figures cited here carry the publication dates of their sources: the 2025 Best Practices Study via Risk & Insurance (22 August 2025) and Big "I" (12 August 2025), and NAIC model bulletin adoption via Quarles (2 April 2025). Regulatory adoption in particular is still moving — verify the current position for your jurisdictions before relying on it.