You wouldn't accept spreadsheet output without checking formulas.
Don't accept GenAI output without verification either.
Ask, Verify, Refine
This is the only workflow that matters.
Ask: Give GenAI a clear task
"Draft an email explaining why we're increasing rates on commercial auto by 8%."
Verify: Check every factual claim
- Is the rate increase actually 8%?
- Did you cite the right loss ratio?
- Are the effective dates correct?
- Does the tone match your company's communication style?
Refine: Edit the output until it's yours
Remove generic language.
Add specific context.
Adjust technical depth for your audience.
Sign your name to it only when you would have written it yourself.
Never Accept Raw Output
Raw GenAI output always has problems.
Sometimes small.
Sometimes catastrophic.
Small problems:
- Generic phrasing that doesn't match your voice
- Technically correct but contextually wrong
- Missing caveats specific to your situation
Catastrophic problems:
- Fabricated citations
- Incorrect calculations presented as fact
- Logical gaps covered by confident language
You won't catch catastrophic problems by skimming.
You catch them by verifying every claim that matters.
Treat It Like a Junior Analyst
Imagine you hired someone smart but inexperienced.
They've read all the textbooks.
They haven't worked on your company's data.
You'd assign them to:
- Draft standard documents
- Summarize routine analyses
- Format data for presentations
You wouldn't ask them to:
- Make final pricing decisions
- Choose reserve methodologies
- Explain unique aspects of your portfolio
That's exactly how to use GenAI.
What you'd tell a junior analyst:
"Take these loss triangles and draft the reserve summary section. I'll review the development factors you chose and make sure the ultimates look reasonable."
What you'd tell GenAI:
"Here are the reserve estimates by line of business. Draft three paragraphs explaining the year-over-year changes. I'll verify the numbers and adjust the explanations."
Same supervision model.
Same verification requirements.
The Verification Checklist
Before using any GenAI output:
For numbers:
- [ ] Did I independently verify every calculation?
- [ ] Are the figures sourced from my own data?
- [ ] Would I stake my credentialing on these numbers?
For assumptions:
- [ ] Is this assumption appropriate for my specific context?
- [ ] Have I documented why this assumption makes sense?
- [ ] Can I defend this assumption to a regulator or peer reviewer?
For explanations:
- [ ] Is this explanation accurate for my company's situation?
- [ ] Have I removed generic language that doesn't apply?
- [ ] Would a colleague recognize this as my work?
If you can't check all the boxes, keep refining.
Concrete Example: Reserve Memo
You need to draft a reserve adequacy memo.
Ask:
"Draft an executive summary for a reserve analysis showing $2.3M favorable development in workers compensation and $1.1M adverse development in general liability."
Verify:
Check that:
- Development amounts match your actual results
- Any mentioned drivers (claim counts, severity) are accurate
- No fabricated statistics appear
- Reserve adequacy conclusion aligns with your analysis
Refine:
- Replace "industry trends suggest" with your actual experience
- Add specific context about known large claims
- Adjust tone to match previous memos
- Include required regulatory language
Before/After:
GenAI Draft: "Our reserve analysis indicates overall adequacy with mixed development patterns across lines of business. Workers compensation showed favorable emergence driven by lower than expected severity, while general liability experienced adverse development primarily due to higher claim frequencies in recent accident years."
Your Final Version: "Reserves remain adequate at $23.4M, representing 105% of actuarial best estimate. Workers compensation released $2.3M due to faster claim closures in the 2023 accident year. General liability added $1.1M, driven by three severity claims in products liability, two of which are still in litigation."
The structure came from GenAI.
The content came from you.
When to Not Use GenAI
Some tasks don't benefit from GenAI at all:
- Building experience triangles
- Selecting development factors
- Applying credibility weighting
- Documenting unusual case reserves
- Explaining one-off exceptions
These require your judgment from the start.
GenAI can't draft something when there's no pattern to follow.
The Iron Rule
If you can't verify it, don't use it.
If you wouldn't show it to a peer without caveats, revise it.
If you couldn't defend it under questioning, rewrite it.
GenAI is a starting point.
Never an endpoint.
Think of GenAI as autocomplete with excellent vocabulary.
It will suggest the next sentence.
You decide if that sentence is true.