Gains:
- Being able to distinguish where in the actuarial workflow (data preparation, code generation, report drafting, scenario interpretation) artificial intelligence saves real time and why decisions such as premium/provision/capital are left to the competent actuary, depending on the task risk level.
- Ability to apply a discipline that verifies each AI output through the steps of linking it to procedural source, independent recalculation, and testing it with actuarial reasoning.
- Anonymizing policy, damage and health data within the scope of KVKK/privacy and gaining the habit of choosing safe vehicles that do not use data in model training
In an insurance or pension company, hundreds of future numbers are produced every month: how much premium should we pay for this group of policies? how much of past damage has not yet been reported; What is the value of the pensions we will pay in 30 years today? If a bad year comes, will the company survive? The profession that answers all of these questions is actuary. An actuary is a specialist who measures, prices and reports insurance and retirement risks using probability, statistics and financial mathematics; Every figure it produces directly affects a company's financial balance, an insured's rights and an auditor's trust. This is where artificial intelligence comes into play.
Artificial intelligence (AI, or AI for short—computer systems that can generate text, write code, recognize patterns, and summarize data like humans) reduces the repetitive and time-consuming parts of actuarial work to minutes: writing data cleaning code, translating an account into Python, drafting an actuary report, explaining a complex distribution in plain language. However, the same AI can also make up a confident number such as "your damage frequency is 8 percent" without ever seeing your data. The first unit of this module is not a software introduction; Its purpose is to clarify where to put AI in your actuarial business and where not to put it at all.
Let's lay out the basic principle from the beginning: Artificial intelligence is an assistant, not an actuary. Responsibility and final approval of finance-critical decisions such as premiums, provisions (reserves), capital and liabilities belong to the competent, appointed and authorized signatory actuary. An unverified AI output is as risky as an unsigned actuary report.
Layers of actuarial business and the place of AI
Separating actuarial work into three layers makes it easier to place AI correctly. The technical/computational layer is where the number is generated: data processing, model building, premium and reserve accounting. The reasoning layer is where the number is interpreted: which assumption to choose, which improvement factor to manually correct, how confident we are in that estimate. The responsibility layer is where the signature is made: the report goes to the regulator, the decision is presented to management, the actuary is the legal owner of the results. AI is very powerful in the first layer, only auxiliary in the second, and not at all in the third.
Let's define a few basic terms from the beginning. Premium is the price paid by the insured in exchange for risk. Damage (claim) is the damage suffered by the insured and paid by the company. Provision/reserve is money set aside today for claims payable in the future. Liability is the present value of the company's future debt to policyholders. Exposure is the unit that measures the magnitude of risk (e.g. a vehicle insured for one year = 1 vehicle-year). We will explain these concepts one by one in the following units; For now, know this: AI gives you the blueprint, code, and analysis for all of these concepts, but you produce the final issue and signature.
The following table summarizes the role and risk level of AI by mission:
Quest
Role of AI
Risk level
Who approves
Data cleaning / coding
code generator
low
Actuary / data analyst
Report and summary draft
sketch generator
low
actuary
Distribution/method description
tutorial, summative
low
actuary
Damage frequency-severity model
Model builder, draft
medium
senior actuary
IBNR / reserve account
account draft
high
Responsible actuary
Pricing / tariff
Model sketch
high
Pricing actuary + compliance
Capital (SCR) / liability
auxiliary input
very high
Appointed actuary + management
Keep in mind one line in this table: as risk rises, the role of AI shrinks and actuarial approval grows.
Why "verification" is the heart of this business
Artificial intelligence language models seem confident in their answer, but they may not be sure. In technical language, this is called hallucination: it is the model's fabrication of non-existent information in a fluent sentence, just as if it were true. For an actuary, this is a deadly trap. The model may produce a number for you such as "the loss/premium ratio of this portfolio is 72 percent"; However, he has never seen your data and this number is completely made up. Or he might say the name of a mortality table as "TRH-2020"; However, there is no such table. Since he says both with equal fluency, the only thing that separates right from wrong is your actuarial knowledge and verification habit.
The verification discipline consists of three steps:
- Link to method and source: Rely on your own data and official sources for numbers, not the AI's memory. Take the AI compendium as a starting point for formulas, distributions and regulatory articles; then confirm from standard actuarial sources (textbook, legislative text, institution procedure). Use AI to interpret data, not to remember it.
- Independent recalculate: Independently check every numerical result, coefficient, and ratio the AI produces. Reproduce a present value, a growth factor, a total yourself or with a verified spreadsheet.
- Test with actuarial judgment: Test with an expert eye whether the output contradicts the facts on the ground (portfolio behavior, inflation, legislation, past experience). "Why did this factor come out as 1.45 and not 1.05?" Always ask the question.
Attention: Signing a reserve figure or premium schedule produced by AI without verifying it is like publishing an unsigned actuary report. Just because the output is smooth and orderly is not accurate.
Privacy: policy and damage data are private data
Policy and claims data includes identity, contact, health and financial information. Some of these are special categories of personal data within the scope of KVKK (Personal Data Protection Law) in Türkiye and GDPR in Europe; In particular, diagnostic information in health and life insurance is subject to the highest protection. It is a serious violation to paste the insured's name, TR ID number, policy number and diagnosis on a public AI vehicle. The same sensitivity applies to portfolio data that are trade secrets (premium tables, loss distributions, reinsurance conditions).
The rule is simple: anonymize data and don't share unnecessary. "54-year-old male, history of heart disease" instead of "Ahmet Yılmaz, 54, policy 2024-88213, myocardial infarction"; Share grouped/summarized tables instead of individual records. Most actuarial analysis is done with already aggregated data (age group, coverage type, year of development); This both protects confidentiality and simplifies analysis. If possible, choose corporate tools that have a data processing agreement and do not use your data in model training.
Tip: In actuarial work the AI is often given a summary/aggregate table rather than a raw record. A table like "table 3: age group × coverage × average damage" has no personal data, but everything needed for analysis.
three mini cases
Case 1 — Safe and efficient use. An assistant actuary would spend 2 days manually cleaning 40,000 lines of claims data. He explained the structure of the data (column names, example 5 anonymous rows) to the AI and asked for a Python cleanup code. AI produced a draft of code in 15 minutes; The actuary read the code line by line, fixed two logic errors (negative damages had to be examined separately rather than deleted), and ran it on his own data. Duration: half day instead of 2 days. The AI wrote the code, the actuary checked it and ran it.
Case 2 — Unverified number trap. An employee asked YZ, "What is the combined ratio of our insurance branch?" The AI confidently gave a number of "about 104 percent" even though it had no access to any data. The employee put this in his management presentation; the actual rate was 97 percent. The difference led to a profitable branch being portrayed as a loss and an incorrect price increase decision. Error: Waiting for an indicator from the AI without giving data to it.
Case 3 — Breach of confidentiality. An actuary loaded an Excel containing life insurance applicants' full name, ID and health declaration into a public AI tool and said "flag the risky ones." The data went to an external server; A KVKK review and reputation risk arose. The correct way was to remove the name and ID and share only age, gender, coverage amount and anonymous risk factors.
Weak prompt / Strong prompt
Weak prompt:
Tell us the loss frequency of our portfolio this year and the appropriate premium.
This claim is flawed: the AI is given no data, so it can only answer the question of "frequency" and "premium" with made-up numbers. What branch, what period, what exposure is clear.
Powerful prompt:
Your role: analysis assistant assisting an actuary.Task: Help me interpret the ANONYMOUS aggregate table below.Data (insurance, 2024, no province/name):- Exposure: 12,400 vehicle-years- Number of claims: 1,612- Total claims paid: 48,900,000 TLI want:1) Find the frequency of damage (number/exposure) and average calculate the damage severity, show the formula.2) Calculate the risk premium (frequency × severity).3) Write the intermediate steps so that I can verify each number manually.Do not make things up; Just use the numbers I gave you. If there is anything missing, ask.
This request is strong because it provides data anonymously, requires formulas, requires intermediate steps, and prohibits fabrication.
Common mistakes
- Asking for numbers without giving data. The AI does not have your portfolio in its memory; Every number he gives is fake. First provide anonymous data, then ask for an account.
- Signing the output without verifying it. A fluid and orderly chart does not mean it is accurate. Reproduce each number independently.
- Pasting personal data in its raw form. Name, ID number, policy number and diagnosis are never shared. Anonymize and use aggregates.
- Letting AI choose the method and not questioning it. "Which method should I use?" The answer to the question is a suggestion; the final choice rests with actuarial judgment and legislation.
- Mistaking AI for a source of legislation. The model can make up legislation articles and table names. Confirm from the official text.
In summary
Artificial intelligence is a powerful assistant that speeds up the technical and repetitive parts of actuarial work; writes code, produces drafts, explains. But the responsibility for premium, provision, capital and liability decisions lies with the competent actuary. Each output is validated through three steps: bind to method/source, independent recalculate, test with actuarial reasoning. Data is always anonymized and unnecessary is not shared. An unverified AI output is as risky as an unsigned report.
Application task
Choose a branch in which you work (or assume). Prepare a completely anonymous and aggregated mini table for this branch (exposure, number of claims, total damage). Adapt the "Powerful prompt" above to your own table and give it to an AI tool. Recalculate the return frequency, severity, and risk premium manually (with a calculator) and compare with the AI's result. If you find a difference, note why.
checklist
- [ ] Is the data I give to the AI anonymous and aggregated? Is there any personal/private information?
- [ ] Have I recalculated each numeric output independently?
- [ ] Did I take the method/assumption choice as a suggestion and decide it myself?
- [ ] Have I confirmed the legislation and table references from the official source?
- [ ] Do I know whether the tool I use uses data in model training?
- [ ] Is this output a draft, or did I accidentally treat it as final?