Systems and records
Insurance renewal change detection data: what changed since last term
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
An insurance renewal comparison dataset should preserve both policy terms in full, including declarations, forms and endorsements with edition dates, plus a field-level record of every difference. Each difference needs evidence spans from both terms, a materiality label tied to a written guideline or notice rule, the notice actually sent to the insured, and a dated outcome record showing how the change played out.
License terms
Insurance renewal change
Key takeaways
- Renewal comparison differs from policy checking: the two terms are supposed to differ, and the question is which differences the insured needed to know about.
- Align forms by number and edition and schedules by stable keys before comparing text.
- Label materiality with a small fixed set, citing both versions and the rule or guideline used.
- Include counterexamples: premium changes explained by exposure, and unchanged limits narrowed by new endorsements.
- Verify labels against dated outcomes such as affected claims or remarketing, not a renewed status field.
What should an insurance renewal comparison dataset preserve?
It should preserve both terms of the policy side by side: the expiring and renewal declarations, the full form and endorsement schedule for each term with edition dates, a field-level list of every difference, a materiality label for each difference with the evidence and rule behind it, the notice that went to the insured, and what happened after renewal. Without the outcome, the dataset teaches a model to spot differences, not to judge which ones mattered.
Insurance renewal change detection data is the record of that comparison. An agent, account manager or underwriting assistant answering what changed since last renewal has to do more than diff two PDFs. Forms are renumbered, schedules are reordered, a limit can stay the same while a new endorsement adds a sublimit beneath it, and a premium can move for reasons the insured caused. The dataset has to capture enough context to tell those cases apart.
The rest of this guide gives the packet a team would assemble for each renewal, the change types worth labeling, an illustrative worked case and the counterexamples that keep the labels honest.
Why is renewal comparison a different task from policy checking?
Policy checking compares documents within one term: the quote, binder and issued policy should agree, and a mismatch is an error to fix. Renewal comparison compares two terms that are supposed to differ, and the question is which differences the insured needed to know about and whether they were told.
Several jurisdictions write part of that judgment into law, which gives the dataset an external anchor for materiality. New York's Insurance Law Section 3426 requires conditional renewal notices for covered commercial policies to be mailed or delivered at least sixty, but not more than one hundred twenty, days before expiration, and to state the specific reasons and the amount of any premium increase. California's Insurance Code Section 678.1 addresses an insurer that will not renew or will condition renewal on changed terms, such as a reduction in limits, elimination of coverages or an increase in deductibles.
Outside the United States the same idea appears in conduct rules. The UK Financial Conduct Authority's ICOBS 6.5 requires firms renewing covered general insurance contracts to give the renewal premium alongside a comparable figure, and to include a statement that the consumer should check that the cover still meets their needs.
These rules differ in scope and detail, and a dataset should record which ones applied to each policy rather than assume one standard. They are useful because they turn some materiality calls from opinion into a checkable fact: the notice was required, and it was or was not sent on time with the required content.
Which renewal changes count as material?
A change is material when it alters what is covered, how much is paid, what the insured must do, or what the insured pays, in a way a reasonable insured would want to decide on. Cosmetic differences, such as a reprinted form with no wording change, are not material even though a naive diff flags them.
| Change type | Illustrative example | Usually material when | Evidence to cite |
|---|---|---|---|
| Limit | Per-occurrence limit unchanged, new sublimit for a named peril | The sublimit applies to a loss type the insured has had or is exposed to | Declarations of both terms and the new endorsement |
| Deductible or retention | Wind deductible changes from a flat amount to a share of insured value | Always, for the affected peril | Deductible schedule of both terms |
| Exclusion added or removed | New exclusion endorsement for a category of loss | Always; note whether the prior term covered it | Endorsement schedule and endorsement wording |
| Form edition | Coverage form moves to a newer edition | The new edition changes wording that affects this insured | Form numbers, edition dates and a clause-level comparison |
| Schedule of locations, vehicles or persons | A location drops off the schedule | The item is still owned or used by the insured | Schedules of both terms and the insured's exposure update |
| Conditions and warranties | New protective device or inspection condition | Breach could reduce or void cover | Conditions section of both terms |
| Premium or rating basis | Premium rises with no exposure change | The increase is not explained by insured-reported changes | Premium summary, rating worksheet if held, exposure updates |
| Insurer or program | Renewal written by a different carrier in the same group | Forms, claims handling or financial terms change with it | Declarations and policy jacket of both terms |
What the renewal comparison and material change labeling packet contains
Assemble one packet per renewal, with one change record per difference. The header carries the two terms and the jurisdictional context; the change records carry the comparison and the label. Fields copied from source documents stay separate from fields a reviewer assigns.
At a glance
What the renewal comparison and material change labeling packet contains
| Field | What it holds | Format or allowed values | Why it matters |
|---|---|---|---|
| renewal_ref, line, jurisdiction | Pseudonymous policy key, line of business, state or country of the risk | Stable token; line code; jurisdiction code | Decides which notice rules applied |
| expiring_term and renewal_term | Effective and expiration dates, carrier, form and endorsement schedule with edition dates | ISO 8601 dates; form ID plus edition for every item | The two versions being compared |
| change_id and change_type | One difference between the terms | Controlled list matching the change types above | Lets evaluators slice by change type |
| before_span and after_span | The exact text or value in each term | Document ID with page and line or character offsets; null when added or removed | Forces citation to both versions |
| materiality_label | Reviewer's judgment | material_notice_required, material, immaterial, cosmetic, cannot_determine | The training target |
| rule_ref | Notice rule or internal guideline that drove the label | Statute or guideline ID with version date, or none | Separates legal triggers from judgment |
| notice_record | What the insured was told and when | Document ID, send date, delivery method, or none found | Tests whether disclosure matched materiality |
| outcome_record | What happened after renewal | accepted, remarketed, mid-term change, claim affected by the change, non-renewed, unknown | Verifies the label against events |
| reviewer and guideline version | Who labeled it and under which rules | Role code; guideline version and date | Audit and relabeling |
How do you align two policy versions before comparing them?
Align by structure first, then compare text. Match forms by form number across terms, then by edition date, and only then compare wording; match schedules by a stable key such as an address or vehicle identifier rather than by row order. A diff run on raw document text will report reordering as change and miss a substituted endorsement that happens to sit in the same position.
Clause-level research datasets show the shape that works. ContractNLI asks whether each hypothesis is entailed by, contradicted by or not mentioned in a contract and requires the evidence spans for that decision. CUAD, an expert-annotated contract review dataset, labels clauses across 41 categories so a model learns to find the provision before it judges it. A renewal packet borrows both ideas: find the corresponding provision in each term, cite it, and only then label the change.
Keep edition dates as first-class data. Two forms with the same number but different editions are different documents, and the dataset should never let a model assume they match. Where the agency or insurer holds a form library, record the library version used for the comparison so the alignment can be rerun.
Expect some provisions to have no counterpart. An endorsement added at renewal has no before_span, and one dropped at renewal has no after_span. Those one-sided records are often the most important changes in the packet, so the schema should allow them explicitly rather than forcing a false match.
How should the materiality guideline for labelers be written?
Write it per line of business, as a short set of tests a reviewer can apply in order, and version it. A single guideline for every line produces labels that are right on average and wrong for the account in front of you: a sublimit on water damage matters to a restaurant in a way it may not to a parking operator.
Start each test from the insured's position. Does the change reduce what would be paid for a loss this insured could plausibly have? Does it add a duty the insured must perform, such as a protective device condition or a reporting deadline? Does it change who or what is insured? Does it change the price for a reason the insured did not cause? A yes to any test makes the change material; a statutory notice trigger upgrades it to material_notice_required.
Calibrate before labeling at scale. Have two experienced reviewers label the same sample of renewals independently, compare their labels change by change, and rewrite the tests wherever they disagree. Repeat until disagreements are rare and explainable, then record the guideline version on every label so later changes to the guideline can be traced.
Keep the account context the reviewer used. If a reviewer marked a dropped location as material because the insured still owned it, the packet should cite the exposure record that shows ownership. A label that depends on facts outside the two policy terms is only reproducible when those facts travel with it.
Worked case: a commercial property renewal (illustrative)
This case is invented for illustration and does not describe any real policy or insured. A small manufacturer's commercial property policy renews with the same carrier. The account manager is asked what changed since last renewal, and the packet records five differences. The jurisdiction in the example requires a conditional renewal notice for reduced coverage.
Starting state: two terms with identical building and contents limits on the declarations. Linked evidence: the declarations, endorsement schedules and the carrier's renewal letter for both terms. Decision: three changes are labeled material, one of them notice-required. Verified outcome: the insured remarketed the account after receiving the notice, which the agency's activity log records.
| Change | Before and after | Label | Evidence and outcome check |
|---|---|---|---|
| New water damage sublimit | No sublimit; sublimit endorsement added under the unchanged building limit | material_notice_required | Endorsement schedule both terms; notice sent within the required window; insured remarketed |
| Coverage form edition | Same form number, newer edition date | immaterial | Clause comparison shows renumbering and a definitions reorder only |
| Location schedule | Warehouse dropped from schedule | material | Insured's exposure update confirms the warehouse was sold; agent note records confirmation |
| Equipment breakdown deductible | Waiting period lengthened | material | Deductible schedule both terms; mentioned in renewal summary, no statutory notice rule applied |
| Policy jacket | Reprinted with updated carrier address | cosmetic | Jacket text otherwise identical |
Counterexamples that keep materiality labels honest
Include cases that look material and are not, and cases that look unchanged and are, in roughly equal measure. Without them, a model learns that any flagged diff is important and that an unchanged declarations page means nothing changed.
One false positive to include is a premium increase. If the insured reported higher sales or added a location, the change follows the exposure and the label is immaterial for disclosure purposes, though the dataset should still cite the exposure update that explains it. One false negative to include is the unchanged limit with a new sublimit or exclusion endorsement beneath it, as in the worked case: the declarations page looks the same and the cover is narrower.
Record cannot_determine honestly. When an endorsement is listed on the schedule but its wording is not in the file, the reviewer cannot judge materiality, and the dataset should say so rather than guess from the endorsement title.
A third counterexample worth including is the change that was material and disclosed but drew no reaction. It still belongs in the material class. Materiality describes what the insured needed to know, not whether they acted on it, and a model trained only on changes that caused remarketing will under-flag the quiet ones.
What dated record shows how a renewal change played out?
A status field that says renewed proves almost nothing about whether the change was handled well. The outcome has to come from a dated record that shows what the insured knew and did, or how the change played out.
The strongest outcome evidence is a later event that the change governed: a claim in the renewal term that was limited or excluded by the new endorsement, or a mid-term change request that restored the prior cover. Next comes an agency activity record showing the insured was told and responded, such as a remarketing request or a signed acknowledgment. Weakest is the absence of complaint, which should be labeled unknown rather than treated as acceptance.
Outcome windows matter. A renewal extracted a month after inception has had little time to produce claims or mid-term changes, so its outcome is mostly unknown. Record the extraction date and the length of the observed window on every packet, and let evaluators filter to renewals with a full term of history when they need verified outcomes.
- Link the notice document and its send date to the change it disclosed, not just to the policy.
- Link any claim in the renewal term to the change records whose wording it touched.
- Record mid-term endorsements in the renewal term, since they often reverse a change the insured noticed late.
- Keep the extraction date, so open outcomes can be refreshed rather than mislabeled.
Rights, limits and preparation before licensing renewal records
Renewal packets combine documents with different owners. The policy forms and endorsements are usually the insurer's or a rating organization's, standard industry forms carry their own license terms, and the insured's schedules and exposure data concern a client. ACORD, for example, states that its forms may only be used with a valid license and that users of vendor software must obtain one directly from ACORD. An agency's own comparison notes, labels and activity records are the part it is most likely to control, and counsel should confirm the rest before any licensing.
Treat these records as client information. Remove or tokenize insured names, addresses, vehicle identifiers and individual contact details, and keep the token key inside the company. De-identification lowers privacy risk; it does not give the right to license a document someone else owns.
Bring in specialist review before labeling at scale. An insurance operations expert should set the materiality guideline for each line of business, and a compliance reviewer should confirm which notice rules apply in each jurisdiction. The NAIC's model bulletin on insurers' use of AI systems, adopted by NAIC members in 2023, sets expectations for a written AI program and the documentation regulators may request, which a packet with cited evidence and versioned guidelines supports.
A brokerage or agency deciding whether its renewal history could become a licensed dataset can have SourceX take selected records through rights review, contracting, preparation and delivery. The opening step is a short company assessment, nothing is uploaded for it, and no record is delivered before the company signs an agreement on scope and terms it has approved.
Frequently asked questions
What is insurance renewal change detection data?
It is a set of renewal comparisons where each difference between the expiring and renewal terms is recorded with evidence from both versions, a materiality label, the notice sent to the insured and the outcome. It trains and tests tools that answer what changed since last renewal.
Is comparing declarations pages enough?
No. Many material changes sit in endorsements, conditions or form editions while the declarations look identical. A useful comparison aligns the full form and endorsement schedule for both terms, with edition dates, before judging any change.
Who should label materiality?
Experienced account managers or underwriters should apply a written guideline per line of business, with a compliance reviewer confirming which notice rules apply. Annotators without insurance experience can align documents and mark spans but should not decide materiality.
Do statutory renewal notice rules define materiality?
Only partly. Rules such as New York Insurance Law Section 3426 or California Insurance Code Section 678.1 make certain changes notice-triggering for covered policies, which gives an objective label for those cases. Many material changes fall outside any statute and still need a guideline and a reviewer.
What outcome should a renewal change label be checked against?
Use dated records: a claim in the renewal term affected by the change, a mid-term change restoring cover, or an activity record showing the insured was told and responded. A plain renewed status should be treated as weak evidence.
Sources
- New York Insurance Law Section 3426 requires conditional renewal notices for covered commercial policies to be mailed or delivered at least sixty, but not more than one hundred twenty, days before expiration and to state the specific reasons and the amount of any premium increase. Source
- California Insurance Code Section 678.1, in Chapter 11, addresses insurers that will not renew or will condition renewal on changed terms such as a reduction in limits, elimination of coverages or an increase in deductibles. Source
- FCA ICOBS 6.5 requires firms renewing covered non-group general insurance contracts to give the renewal premium with a comparable figure and a statement that the consumer should check the cover still meets their needs. Source
- ContractNLI asks whether each hypothesis is entailed by, contradicted by or not mentioned in a contract and requires evidence spans for the decision. Source
- CUAD is an expert-annotated contract review dataset with labels spanning 41 categories. Source
- ACORD states its forms may only be used with a valid license and that users of vendor software must obtain a license directly from ACORD. Source
- NAIC members adopted a model bulletin on insurers' use of AI systems in 2023 that sets expectations for a written AI program and documentation regulators may request. Source
See if your company qualifies
A short company assessment. No data uploads are needed.