The slow part of open enrollment is not the conversation with the client. It is building the side-by-side comparison: pulling the premium, deductible, out-of-pocket max, copays, coinsurance, and network type out of each Summary of Benefits and Coverage and lining them up so the differences are obvious. Reading those SBCs and tabulating the numbers is slow, repetitive work, and it is exactly what AI is good at. The trap is that the same tool that builds you a clean grid in a minute will also misread a deductible, fill a blank with a "typical" figure it has seen before, or tell the client which plan to pick, and that recommendation is licensed work that only the advisor can make. This is a plain look at where AI genuinely helps with SBC comparison, where it does not, and a set of prompts you can paste in and test on your own real SBCs.
The honest picture
AI is genuinely useful for the reading and tabulating, the part that is slow and mechanical:
- What AI does well today: read an SBC, pull the premium, deductible, out-of-pocket max, copays, coinsurance, and network type into a structured row, line several plans up side by side, and draft a neutral, plain-language summary of how they differ. It compresses the document work, not the judgment.
- What AI does not do: recommend. It cannot tell a client which plan to choose, whether a deductible is right for their situation, or what coverage fits their family, and it cannot advise on suitability. That is licensed work, and the responsibility stays with the advisor. AI extracts and tabulates what is on the SBC documents you give it; the licensed advisor advises and recommends.
The right way to think about it: AI is a fast reader that organizes what the SBCs say, not an advisor. The grid is a draft. The recommendation and the suitability call are yours.
The line: it will recommend a plan or misread a number if you let it
The useful job and the job that gets you in trouble sit right next to each other, so the line is worth stating plainly. Watch for these failure modes:
- Extracting the wrong number. It pulls the wrong deductible, out-of-pocket max, or copay off a dense SBC, or copies a figure into the wrong plan's column. A clean-looking grid with one wrong number is worse than no grid.
- Inventing a value. Asked for a field the SBC does not state, it fills the blank with a "typical" or "standard" figure it has seen elsewhere instead of leaving it empty.
- Recommending which plan to choose. Asked to compare, it slides into "Plan B is the best value" or "most people in your situation pick the HMO." That is a recommendation, and it is the line.
- Interpreting coverage as advice. It reads a coverage detail and turns it into guidance about what the client should do, which is suitability advice.
The fix is one discipline: AI extracts and tabulates ONLY what is on the SBC documents provided, and the licensed advisor advises and recommends. Keep AI strictly on the extracting side: it lays the facts out in a grid and cites where each one came from. You make the recommendation.
The setup that keeps the recommendation with the advisor
Two habits make AI much safer here, and the prompts below build them in:
- Feed it the actual SBCs and nothing else. Give it the documents and tell it to read only what is on them. A model told "use only these SBCs, copy each figure exactly, leave blank anything not on the document" invents far less than one asked to describe a plan from memory, where it will confidently report numbers no SBC contains.
- Tell it not to recommend. Be explicit that it extracts and tabulates but never advises on suitability or tells the client which plan to pick. You want a comparison grid and the differences laid out, not a verdict.
How to test it on your own work
Do not trust a polished demo, including this one. Pull two or three of your own real SBCs from a recent enrollment, a timer, and the prompts below. Rate each output 1 to 5 on usefulness and accuracy, and check every extracted figure against the source SBC. Compare the time against how you build the grid today. Keep what wins. Use your brokerage's enterprise or no-train settings before you paste any client document, and confirm you are comfortable with that. Treat this as a workflow guide, not licensing or compliance advice: confirm anything about your obligations with your brokerage's compliance contact or counsel.
Paste-ready prompts
Copy these as written. Bracketed text is what you swap per file.
Test 1: Extract one SBC into a clean row (text model)
I am attaching one Summary of Benefits and Coverage. Extract the key fields into
a single clean row, copying each figure exactly as written on the SBC. Rules:
- Include: monthly premium (if shown), deductible, out-of-pocket maximum, primary
and specialist copays, coinsurance, and network type (HMO, PPO, etc.).
- Extract only what is on the SBC I paste. Copy each figure exactly. Recommend
nothing. Leave blank anything not on the document, do not fill it with a
typical or standard value.
- For each field, note the page or section of the SBC you took it from.
SBC: [attach or paste it]
Watch for: did it copy every number correctly, and did it leave blanks empty instead of inventing a "typical" figure?
Test 2: Build the side-by-side comparison grid (text model)
Here are several Summaries of Benefits and Coverage for plans I am comparing:
[attach or paste each one, labeled]. Build one side-by-side comparison grid with
one column per plan and one row per field: premium, deductible, out-of-pocket
maximum, primary copay, specialist copay, coinsurance, network type. Rules:
- Extract only what is on the SBCs I paste. Copy each figure exactly, with the
plan it came from. Leave a cell blank where that SBC is silent on the field.
- Recommend nothing. Do not say which plan is better, cheaper, or a better fit.
- Flag any mismatch for me, for example a field one SBC reports differently than
the others, so I can check the source.
Watch for: did every figure land in the right plan's column, did it leave silent fields blank, and did it stay on tabulating without sliding into "Plan B is the best value"? If it recommended anything, that is the line.
Test 3: Neutral plain-language summary for an employee (text model)
Using only the comparison grid above, write a plain-language summary, about 150
words, of how these plans differ, for an employee deciding at open enrollment.
Rules:
- Describe the differences in deductible, out-of-pocket max, copays, and network
only, using the exact figures from the grid.
- Recommend nothing. Do not call any plan the best, cheapest, or right choice,
and do not suggest what most people pick.
- Add no fact that is not in the grid. Leave out anything the grid does not cover.
Watch for: does the summary stay neutral and accurate, and did it avoid any "best plan" language, so the recommendation stays with you?
Test 4: Accuracy audit (text model)
Compare this comparison grid against the source SBCs I provided. For each
problem, quote the exact cell or line.
1. Any value in the grid that is not on the source SBC it is attributed to.
2. Any cell filled with a figure the SBC does not state.
3. Any recommendation or "best plan" language anywhere in the output.
Do not fix anything. Only flag. Treat any invented value or any recommendation
as a defect.
Source SBCs: [attach or paste]
Comparison grid: [paste]
Watch for: does it catch the number that does not trace back to any SBC, and does it flag any sentence that crept toward recommending a plan? Run it on a grid you already built by hand.
What success looks like
If your own testing shows real time savings, the next step is a small pilot: run a week of enrollments through the prompts and measure the hours. If that holds up, the natural next step is a simple agent, running on your brokerage's own cloud, that you use in plain language. The most useful version ingests each SBC, extracts the premium, deductible, out-of-pocket max, copays, coinsurance, and network into a comparison grid, drafts the neutral summary, and cites the page or line of the SBC behind every value, so you verify each figure and make the actual recommendation before anything goes to the client. It extracts and tabulates and cites; you advise and recommend.
The principle holds the whole way through: AI gives you a faster read and a cleaner grid. It does not carry your license, and it does not make the recommendation. Keep that line clear and the rest is upside.
This is general information about workflow tools, not insurance or benefits advice, and not a recommendation about any plan.
Want a straight answer for your benefits practice?
I build practical AI and custom software for businesses, on Google Cloud. If you want a second set of eyes on how AI could fit your plan-comparison workflow, or on a tool you are considering buying, tell me what you are working with. No pitch, just a straight answer.