Brokers compare wordings at renewal in two passes. First, a mechanical comparison (Word's legal-blackline Compare, or Compare Files in Acrobat Pro) so no change is missed. Then an AI explains each difference: which clause, what it changes for this client, how material it is. Every flagged change is checked against the wording before it reaches the renewal report.
The reason for two passes: asked simply to "spot the differences" between two 60-page wordings, an AI summarises confidently and can miss a narrowed definition, a condition moved into another section, or an endorsement sitting in the schedule rather than the wording. A text-compare tool understands nothing about insurance, but it doesn't skip pages. The diff guarantees coverage, the AI explains, and the broker decides.
Gather the full set, not just the wording
Changes hide in the documents nobody thinks to compare. For each policy, collect both years of:
- Schedule: limits, sub-limits, excesses or deductibles, retroactive date, territorial and jurisdiction limits, named insureds.
- Policy wording, with its version reference. Check whether the insurer has moved to a new version or just re-issued the old one.
- Endorsements on both the expiring and renewal terms, including any the insurer is removing.
- Subjectivities and conditions attached to the renewal quote, such as surveys, security requirements or risk-improvement deadlines.
- Mid-term changes to the expiring policy, so you compare against what was actually in force, not what was issued a year ago.
If the wording version reference hasn't changed, the comparison is mainly schedule and endorsements. If it has, compare everything.
Schedules are short, so it is tempting to ask the AI for a quick summary of them instead of a diff. That is where one illustrative review of a small joinery workshop's combined policy went wrong. The AI's summary said "excesses unchanged at $250". Laid out line by line from both schedules, the picture was different:
| Schedule item | Expiring | Renewal |
|---|---|---|
| Standard excess | $250 each and every claim | $250 each and every claim |
| Escape of water excess | Not stated (standard applies) | $1,000 each and every claim |
| Theft excess | $250 | $250, plus a new condition requiring an alarm to be set |
| Business interruption indemnity period | 24 months | 12 months |
The summary was true of the one line it read and wrong about the policy. Three of the four rows changed, and the indemnity period halving is the kind of change a joinery firm with long lead times on specialist machinery needs to hear about. Ask the AI to extract both schedules into a table with one row per item and a "not stated" value where a line is missing, then compare the rows yourself. The extraction is the part it's good at.
Pass one: a mechanical diff that can't skip a page
For Word documents, use the Review tab, then Compare, then "Compare two versions of a document (legal blackline)". Word creates a third document showing every insertion and deletion, leaving the originals untouched. For PDFs, Acrobat Pro, Premium or Studio has a Compare files tool that shows the old and new files side by side with changes highlighted; tick "Compare text only" to ignore layout noise.
Two practical points. Scanned PDFs need text recognition first, or the comparison finds nothing. And if the insurer has redesigned the document, the diff will be noisy, flagging whole pages as changed because the numbering moved. In that case, compare section by section: definitions against definitions, exclusions against exclusions. It's slower but still mechanical, which is the point.
AI can help sort a noisy diff, as long as you check the sorting. Say a redesigned 64-page wording produces 312 flagged changes. Paste the change list and ask the AI to put each item into one of two groups: "numbering, formatting or layout only" and "wording changed". It might return 271 and 41. Before trusting the split, pick ten items at random from the 271 and read them against the documents. If all ten really are formatting, work through the 41. If even one is a genuine change, such as "his or her" becoming "any person" inside an exclusion, the sorting can't be trusted for this document and you go back to section-by-section comparison.
Pass two: the explanation prompt
Give the AI both versions of one section at a time, plus the list of changes from the diff, and ask for a structured explanation. Use a business plan (ChatGPT Business, Claude Team or equivalent) that doesn't train on your content by default.
You are assisting an insurance broker reviewing renewal changes for a
[class of business] policy. Below are the expiring and renewal versions
of the [section name] section, and the list of changes from a text
comparison. For EACH change in the list, return a table row:
- Clause reference (both versions)
- Expiring wording (quote exactly)
- Renewal wording (quote exactly)
- Plain-English effect on the policyholder
- Materiality: High / Medium / Low, using the rubric below
- Question to raise with the insurer, if any
Rules: quote wording verbatim. Do not say whether any specific claim
would or would not be covered. If a change depends on a definition or
clause elsewhere, name it and write "check cross-reference". If you are
unsure of the effect, say so.
Rubric: [paste]
Expiring: [paste] Renewal: [paste] Diff list: [paste]An illustrative row from a cyber policy's conditions section, and what the broker fixes:
Clause: 6.2 (expiring) / 6.2 (renewal)
Expiring: "...notify us as soon as reasonably practicable..."
Renewal: "...notify us within 72 hours of discovery..."
Effect: The policyholder must now report incidents within a fixed
72-hour window rather than a flexible timeframe.
Materiality: Medium
Question: NoneThe explanation is accurate but the grade is wrong. If clause 6.2 is a condition precedent to liability (the general conditions section says so, and the AI wasn't shown it), missing the 72-hour window could defeat a claim entirely. That makes it High, with a question for the insurer about how "discovery" is defined. The fix for next time: always include the general conditions and definitions sections alongside whichever section you're explaining.
A materiality rubric the whole team applies the same way
Without a shared rubric, one broker's "Medium" is another's "Low", and the AI grades inconsistently too. A filled-in version to adapt:
| Grade | Typical changes | Action |
|---|---|---|
| High | New or widened exclusion; limit or sub-limit reduced; key definition narrowed; new condition precedent or warranty; retroactive date changed on claims-made cover; excess or deductible increased significantly; cover removed | Explain in the renewal report, discuss with the client, consider alternatives |
| Medium | Notification or claims-handling procedure changed; cover extension added with its own conditions; wording clarified in a way that may narrow interpretation; new subjectivity with a deadline | Mention in the renewal report; raise with insurer if unclear |
| Low | Renumbering, formatting, insurer contact details, regulatory boilerplate, wording clarified with no change in meaning | Record in the file only |
Paste the rubric into the prompt every time, and when the team disagrees about a grade, update the rubric rather than overriding case by case.
An IT support firm's cyber and PI renewal, change by change
Consider an illustrative broker handling the renewal of a combined cyber and professional indemnity policy for a 30-person IT support firm. The insurer has issued a new wording version. The mechanical diff flags 41 changes across the definitions, exclusions, conditions and schedule. The broker runs the explanation prompt section by section, about 20 minutes of work, then spends 25 minutes checking the rows graded High and Medium against the wording itself.
Four changes matter:
- Ransomware sub-limit halved in the schedule, a High change that the AI graded correctly.
- A new exclusion for losses arising from vulnerabilities where a vendor patch had been available for more than a stated number of days, High, graded correctly, and a real issue for a firm that manages clients' patching.
- The definition of "computer system" narrowed to systems the insured owns or leases, which may exclude client systems the firm manages. The AI graded it Medium and flagged "check cross-reference". The broker upgrades it to High: for this client, most of the exposure sits on client systems.
- Notification within 72 hours, as in the example above, upgraded to High.
The remaining 37 changes are renumbering, clarified boilerplate and one new extension for social engineering fraud, with its own sub-limit, graded Medium. In the renewal report, the client gets one page: the four material changes in plain English, what each means for how they work, and two options the broker is seeking from other insurers. The working table, diff and sign-off go on file.
What would have happened with AI alone? In a test on the full wordings pasted at once, the same model listed the sub-limit and the patching exclusion, mentioned "some changes to definitions", and missed the notification change entirely because it sat in the general conditions, 40 pages from the claims section. The mechanical diff is what caught it.
Turning the table into words a client will read
The explanation table is written for the broker. The renewal report is written for a business owner who has ten minutes and no interest in clause numbers. AI helps with this translation too, provided you give it the corrected rows, not the raw ones. A before and after for the patching exclusion above. The table's "effect" column read:
New exclusion 4.14 removes cover for loss arising from exploitation of a vulnerability for which a patch was available for more than the stated period prior to the loss.
The renewal report version, drafted by AI from the corrected row and edited by the broker:
This year's policy won't pay for an attack that exploits a security weakness if a fix had been available from the software maker for longer than the period stated in the policy and hadn't been applied. Because you manage patching for your own clients, this matters in two ways: for your own systems, and for any claim a client makes against you after an attack on theirs. We'd like to talk through your patching process before renewal, and we're asking two other insurers whether they apply the same exclusion.
The second version names the consequence, connects it to how this client actually works and says what happens next. The prompt that produces it is short: "Rewrite these corrected rows for the client's renewal report. Plain English, no clause numbers, explain why each change matters for a business that [what they do], and end each with what we are doing about it." What the broker adds is the sentence about their own clients' systems, because the AI can't know that the firm's biggest exposure sits on other people's networks.
Where AI comparison goes wrong
- Cross-references. A clause that looks unchanged can change meaning because a definition it relies on changed. Always include the definitions section.
- Endorsements that override the wording. An endorsement can reinstate or remove cover the wording describes. Compare endorsements as their own section, and ask the AI which wording clauses each endorsement affects. Watch for the ones that quietly disappear: for an illustrative groundworks contractor, the renewal dropped an endorsement waiving the insurer's subrogation rights against the principal contractors it works for. Nothing in the wording changed, so the diff was clean. But the firm's subcontract terms require that waiver, and without it the firm would have been in breach of contracts it had already signed.
- Document hierarchy. Where the schedule and wording conflict, the policy usually says which prevails. Make sure the AI knows.
- Confident generalisation. Phrases like "broadly similar cover" in an AI summary are a warning sign; ask for the specific differences or treat the section as unreviewed.
- Coverage opinions. The prompt forbids them for a reason. Whether a claim would be covered depends on facts, law and the insurer's position; that is the broker's professional judgement, sometimes with a claims specialist or lawyer.
The questions to the insurer need the same care. AI will draft them quickly from the High rows, but its first instinct is to ask insurers for coverage opinions they won't give in writing. An illustrative before and after for the "computer system" definition:
- AI draft: "Please confirm that an attack on a client's network managed by our client would be covered under the renewal wording."
- Edited: "The renewal narrows 'computer system' to systems the insured owns or leases (definition 2.7). Our client manages networks it neither owns nor leases for around 40 clients. Is the intention to exclude incidents originating on those networks? If so, is an extension available, and on what terms?"
The edited version quotes the clause, gives the underwriter the fact that makes it matter and asks about intent and options. Those are questions an underwriter can answer, and the answers belong on file.
Many of these apply equally to contracts outside insurance; what AI contract review catches and misses covers the general pattern.
Comparing a different insurer's wording, not a new version
When you are moving a client to another insurer, a diff is useless: the two wordings share almost no text. Switch to a clause map. List the cover topics that matter for the class, then ask the AI to find and quote the relevant clause in each wording, side by side:
For each topic below, quote the relevant clause from Wording A and
Wording B with clause numbers, then note any difference in scope.
If a wording has no clause on a topic, write "NOT FOUND". Topics:
[insuring clause for equipment; definition of equipment; hired-in
equipment; equipment in vehicles overnight; theft conditions; excess;
worldwide or territorial limits; claims notification]For an illustrative video production company's equipment policy, the topic list is where the broker's knowledge lives: overnight vehicle theft conditions and hired-in kit are where such claims commonly go wrong, so they head the list. Treat every "NOT FOUND" as a question for you to answer by reading, not proof that the cover is missing; the clause may sit under a heading the AI didn't expect.
Records, confidentiality and your own professional indemnity
Keep the evidence of each review: the diff output, the AI explanation table as corrected, and a note of who checked and signed off, on what date. If a client later disputes whether a change was explained, that file is your answer. Subjectivities need a diary entry as well as a line in the report. If a renewal is subject to, say, an electrical inspection certificate within 60 days of inception, ask the AI to list every subjectivity with its deadline and the evidence the insurer wants, then diary each one in your broking system with a chase date two weeks earlier. A subjectivity missed on day 61 can affect cover, depending on its wording, and no comparison table will remind you of it in month two. Keep client names out of the AI tool where you can, use business plans, and read your insurer agreements for any restrictions on sharing documents with third-party tools; whether chat-with-PDF tools are safe for client files covers the questions to ask.
The broker remains responsible for the advice, whether or not AI helped prepare it, so the check at source is not optional. Once the comparison is in place, it feeds directly into an automated renewal timeline, and it is one of the larger time savings mapped in where AI saves time in a small brokerage.
Broker questions about AI wording comparisons
Can I rely on the insurer's own summary of changes?
Use it as a starting point, not a substitute. Insurer change summaries tend to cover the changes the insurer considers significant, and they may not mention reworded definitions, moved clauses or endorsements. The mechanical comparison shows every textual change; the insurer's summary tells you which ones they chose to highlight. Where the two differ, that difference is worth a question.
Is it safe to upload policy wordings to ChatGPT or Claude?
Standard policy wordings are usually published documents, so the main sensitivity is the schedule and anything client-specific. Use a business plan such as ChatGPT Business or Claude Team, which doesn't train on your content by default, and check your agreements with insurers and clients for any restrictions. Remove client names from schedules if your policy requires it.
Should the comparison table go to the client?
Not in raw form. The working table belongs in your file as evidence of the review. The client should get a short, plain-English section in the renewal report covering the material changes, what they mean for this business, and what you recommend, with the less material changes listed briefly or available on request.
What if the wording is too long for the AI to read in one go?
Split it. Compare and explain section by section, such as definitions, insuring clauses, exclusions, conditions and endorsements, rather than asking for the whole wording at once. You get more reliable explanations, and it becomes obvious if a section was skipped. Long documents are where models are most likely to summarise loosely.
Further reads
- What Does AI Cost a Small Insurance Brokerage in Year One? — What tools like these cost a small brokerage in the first year.
- AI Submission Intake: Stop Re-Keying Data Into Insurer Portals — The data-entry side of the same renewal and new-business work.
- Does Your Business Insurance Cover AI Mistakes? — Whether your own cover responds if an AI-assisted review misses something.
- Questions to Ask Before Buying AI That Touches Client Data — What to ask before a wording-comparison vendor sees client documents.
- How Insurance Brokers Use AI to Handle Claims Enquiries — Wording changes show up again when claims are made.
- How to Check AI Is Doing Good Work, Not Just Fast Work — Track whether AI comparisons stay accurate over time.
- Do I Need AI in My Business? A Decision Guide for Owners — A decision guide for owners: the three situations where AI is worth it, when waiting is sensible, a ten-question score and a 30-day test.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft Support, compare document differences using the legal blackline option; Adobe Acrobat help, compare two versions of a PDF (Acrobat Pro); ChatGPT Business and Claude Team plan pages on data use.