Renewal season is a comparison problem before it is an advice problem. Only one of those two can be automated.
Key takeaways
Brokerages use AI on renewals to do the reading, not the advising. The tool ingests the expiring policy and the renewal terms, extracts the fields that define cover — limits, deductibles, sublimits, endorsements, exclusions, named locations, premium — and produces a difference list. A licensed broker then reads the difference list, decides which changes matter to that client, and makes the recommendation.
That split is not a compliance nicety. It is the only version that survives contact with a book of two hundred renewals, because the bottleneck in renewal season is not advice capacity. It is the hours spent discovering which accounts need advice at all.
Extraction. Pulling structured fields out of declarations pages, schedules and endorsement wordings that arrive as PDFs of wildly varying quality. This is mechanical, high-volume and well suited to a model with a defined schema and a confidence score.
Comparison. Diffing this year against last, including the changes insurers do not headline — a sublimit that moved, a new exclusion inserted into a form, a deductible that became percentage-based. A machine is better at this than a tired person at 7pm in November, because it does not skim.
Advice. Deciding whether a change is material to this client’s operations, whether to remarket, and what to recommend. This is licensed work and it stays with a person. A model has no view on whether the insured has just opened a second warehouse.
A commercial brokerage runs roughly two hundred renewals a quarter across four producers. Historically each file got the same twenty minutes, which meant the complicated ones got twenty minutes too.
The rebuild starts with a fixed field schema — about thirty fields per policy, including one that people forget: the source page number for every extracted value, so a broker can jump straight to the wording rather than hunting for it. Extraction runs per policy, not per batch, so every file carries its own confidence score.
The comparison step then classifies each difference into three buckets: no client impact (premium moved within tolerance, insurer rebranded a form), notify (deductible changed, minor sublimit movement) and discuss (exclusion added, limit reduced, coverage basis changed). The output is a queue, not a report.
The third bucket is the small one, and that is the whole point: the producers spend November on the accounts where a conversation changes the outcome, and the other files get an accurate, generated summary that a broker checks and sends. Nobody saved time by advising less. They saved it by finding the files faster.
A broker who fails to obtain or explain appropriate cover can be liable to the client, and the tool that produced the comparison is not a defence — our sister firm sets out the shape of that claim in can I sue my insurance broker for failing to get me the right coverage. Two controls follow from that. First, the exception queue is worked and cleared, with the reviewer recorded; an unreviewed queue is worse than no queue, because it looks like diligence. Second, the client-facing renewal summary is signed off by the licensed individual whose advice it represents.
Conduct standards for intermediaries are set provincially. CISRO describes itself as a forum of Canadian regulatory authorities dedicated to consistent qualifications and conduct-of-business standards for insurance intermediaries, and publishes a list of its members — that list is the fastest way to reach the body that actually regulates you. In Ontario, FSRA is the regulator for the life and health insurance sector including accident and sickness, and in British Columbia the Insurance Council of BC publishes licensee requirements including continuing education. Write your own regulator’s name into your procedure rather than the phrase “the rules”.
Commercial renewal files carry personal information about directors, drivers and employees. Under clause 4.1.3 of Schedule 1 to PIPEDA the brokerage remains responsible for that information when it is transferred to a third party for processing, and must use contractual or other means to provide comparable protection. The Office of the Privacy Commissioner’s generative AI principles apply to your deployment, not merely to the vendor’s development. If your brokerage has never appointed someone to own this, the privacy-officer question is the place to start.
Choose one line of business with consistent documentation — commercial property or fleet auto — and one quarter. Run the extraction and comparison in parallel with the current process. Measure two things: how many files cleared with no broker touch, and how many material differences the comparison caught that the manual review missed. The second number is the one that justifies the project, and it is usually the surprise.
Onboarding a new commercial client is a different job with different controls; that is covered separately in AI for insurance broker client onboarding.
Can the tool recommend coverage?
No. It can rank files by how much the cover moved, and it can draft the summary a broker edits. The recommendation is licensed advice and belongs to a licensed person.
What about wordings the model has never seen?
Unfamiliar forms are where extraction confidence drops, which is the system working. Route low-confidence files to a person with the source page attached; do not let a low-confidence value post silently.
Do we need to tell clients?
Check your provincial regulator’s conduct expectations. Whatever they say, the summary that goes to the client should be one a broker is prepared to defend as their own work, because that is what it is.
A 30-minute call is enough to tell you whether AI pays for itself here.
