For most finance teams in the insurance sector, the quarterly reserve analysis is a ritual of controlled chaos. You are pulling loss run data from the claims system, triangulating paid and incurred losses across accident years, and then reconciling actuarial estimates with finance’s own judgment. The friction is relentless: spreadsheets with broken links, version control issues when the actuarial team sends “final_final_v3.xlsx,” and the sheer manual effort of drafting the narrative that explains why reserves moved by 4% this quarter. The most painful part is the disconnect between the quantitative output and the qualitative explanation—your CFO wants the story behind the numbers, and you are spending hours writing boilerplate paragraphs that all sound the same.
ChatGPT compresses this workflow from days to hours. It does not replace the actuary or the financial analyst; it removes the mechanical overhead that makes the job tedious. The model can ingest your historical reserve triangles, your prior commentary, and your current period data, then draft a coherent, structured analysis that mirrors your firm’s tone. It handles the first-pass drafting of the narrative, flags inconsistencies between paid and incurred trends, and even generates the required regulatory language for your actuarial opinion summary. You remain the reviewer and the decision-maker, but you stop being the typist.
The key is learning how to brief the model correctly. The prompts below are structured to force ChatGPT to act like a senior financial analyst who understands reserve volatility, loss development patterns, and the difference between IBNR and case reserves. They are not simple “write this for me” commands; they are detailed briefs that give the model context, constraints, and a clear definition of what success looks like. Use them as templates, adapt them with your actual data, and you will get output that is ready for senior review—not a generic essay that needs a full rewrite.
Why Standard Prompts Fail in Reserve Analysis
Most people ask ChatGPT to “analyze my reserves” and get a generic response about the importance of accuracy and prudence. That is useless. The problem is that reserve analysis is domain-specific: the model needs to understand development triangles, loss trends, and the regulatory environment. A vague prompt produces vague output. The following pre-box formats solve this by giving the model a structured brief, including the specific files it should reference and the measurable outcome you expect. This turns the model from a text generator into a junior analyst that drafts your memo correctly the first time.
First, read these files completely before responding:
[reserve_triangle_Q2_2026.xlsx] — the paid and incurred loss development triangle by accident year and line of business
[prior_quarter_narrative_Q1_2026.docx] — the memo from last quarter that was approved by the CFO, showing tone and structure
[actuarial_report_Q2_2026.pdf] — the external actuary’s opinion, including the range of reasonable estimates and point estimates
Here is a reference for what I want to achieve:
[Upload a de-identified version of your best prior reserve memo as markdown, or describe it: a 2-3 page document with sections for Overview, Key Drivers, Changes by Line of Business, and Outlook]
Here’s what makes this reference work:
The tone is confident but not speculative; it quantifies every claim (e.g., “Reserves increased by $2.1M due to adverse development in commercial auto”); it separates paid and incurred trends; it explicitly reconciles the actuarial point estimate with the booked reserve; it flags uncertainty without alarmist language.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: A 1,500-word memo with four sections (Overview, Key Drivers, Line of Business Detail, Outlook)
Recipient’s reaction: They should be able to skim the first 500 words and understand the reserve movement without asking clarifying questions
Does NOT sound like: A generic AI essay, a textbook definition of reserves, or a sales pitch; no phrases like “It is important to note” or “In today’s dynamic environment”
Success means: The CFO approves the memo without requesting structural changes, only minor numeric edits
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
The prompt above is designed for the quarterly narrative. It forces the model to ask clarifying questions before producing anything, which is critical because reserve analysis is not a one-shot generation task. The model needs to know whether you are booking to the actuarial point estimate or to a conservative position within the range. It needs to understand which lines of business are driving the change. The execution plan step also ensures that ChatGPT outlines its approach before writing, giving you a chance to correct course early rather than receiving a full draft that misses the mark.
Automating the Loss Development Factor Calculation
The second major pain point is the mechanical calculation of loss development factors (LDFs) and the selection of the tail factor. This is tedious, error-prone, and often requires cross-checking against prior selections. While ChatGPT is not a spreadsheet, it can process the raw data if you provide it in a structured text format (CSV pasted into the prompt) and ask it to perform the chain-ladder calculations. It can also explain the rationale for selecting a particular tail factor based on industry benchmarks, which is helpful when you need to justify your selection to an auditor. The prompt below handles this scenario.
First, read these files completely before responding:
[claims_data_GL_2026.csv] — cumulative paid losses by accident year and development period, formatted as a triangle
[prior_ldf_selections_2025.xlsx] — last year’s selected LDFs and the rationale notes for each factor
[industry_benchmarks.md] — published LDF benchmarks for general liability from a reputable source (e.g., CAS or reinsurer data)
Here is a reference for what I want to achieve:
[Upload a markdown file showing a completed LDF analysis with a table of age-to-age factors, a volume-weighted average, a selected factor, and a written rationale for the tail factor]
Here’s what makes this reference work:
It shows the calculation steps (not just the final numbers); it compares volume-weighted and simple averages; it explicitly states why the selected factor differs from the average (e.g., “Selected 1.35 over 1.28 due to recent large loss activity”); it includes a tail factor rationale that cites industry data.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: A 1,000-word analysis with a table of calculated factors and a written rationale for each selection
Recipient’s reaction: The auditor should be able to follow the logic and accept the selections without requesting additional documentation
Does NOT sound like: A math textbook or a hand-wavy “this feels right” justification
Success means: The auditor signs off on the LDF selections without follow-up questions, and the analysis is reproducible from the raw data
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
When you use this prompt, be prepared to paste your triangle data as a comma-separated list within the conversation, not as a file attachment, since the model reads text directly. The model will perform the arithmetic, but you should verify the calculations in your own spreadsheet afterward—the value here is the documentation and the rationale, not the raw math. The prompt also forces the model to compare your selections to industry benchmarks, which is a common audit request that usually requires digging through old emails. Having ChatGPT draft that comparison in a structured way saves you an afternoon of formatting.
For your next step, try running the first prompt with your actual Q2 data and see how the draft compares to your prior memo. Expect to edit heavily the first time—the model is learning your style. After two or three cycles, you can create a standard operating procedure document that captures your preferred structure, tone, and data sources. Then, each quarter, you simply upload the new data and the prior narrative, and ChatGPT will produce a draft that is 80% ready for review. The remaining 20% is your professional judgment, which is exactly where you should be spending your time.
Published on 2 August 2026 on growwithgpt.com