Gains:
- Anonymization and evaluation of the necessity before giving the customer's sensitive and financial data to artificial intelligence tools
- Ability to distinguish differences between corporate and public AI tools in privacy, data retention, and use in education
- Ability to manage data breach, retention period and third party sharing risks within the framework of KVKK
Insurance is one of the professions that work with the most sensitive data. An insurance file; It may include identity information, address, income, bank account, health history, damage records, license plate, title deed, and even family information. Some of this data is subject to the highest protection as "personal data of special nature" in the KVKK (Personal Data Protection Law); Health, biometric and similar data are typical examples of this. Uncontrolled sharing of this data when using artificial intelligence tools (especially cloud-based ones) is a serious risk of breach and loss of trust. Once leaked, data cannot be retrieved. In this unit, you will learn how to anonymize customer data (removing personally identifiable information) before giving it to artificial intelligence, the privacy differences between corporate and public tools, and how to manage the risks of data breach, storage and third-party sharing within the framework of KVKK. This unit is not legal advice; Explains general principles, consult compliance/legal department in specific case.
Why privacy is critical: data types
In insurance, it is necessary to separate data according to the level of protection:
- Identity data: Name, surname, TR ID number, date of birth, contact. It directly identifies the person.
- Financial data: Income, bank account, payment history. Personal and sensitive.
- Special quality data: Health, biometric data. Highest protection; Stricter conditions (explicit consent or statutory exception) are required for processing.
- File data: Policy number, license plate, title deed, damage detail. When combined with others, it can reveal identity.
The biggest danger is the identifiability that occurs when you bring this data together. While "45 years old woman" is anonymous, "45 years old woman + a certain license plate + a certain address" can reveal identity.
Attention: Saying "delete" after giving a data to the artificial intelligence tool does not actually delete it. Some tools store the input and even use it to train the model. Rule: never send sensitive data; Anonymization comes before sending.
Anonymization: how to do it
Anonymization is the removal of identifiable information and replacement by factual equivalent; The goal is to make the person unrecognizable while maintaining output quality. Good anonymization:
- Removes name, TR ID, telephone, address, policy number, license plate and title deed information.
- He replaces them with an analytically meaningful equivalent: "45 years old, male, insurance policy, Central Anatolia, one-sided collision".
- It avoids rare/distinctive combinations: a very specific profession + a very small place, which alone can open up identity.
- Retains medical/technical meaning: such as "complementary health, planned orthopedic surgery".
Tip: After anonymizing, ask yourself: "If a stranger read this text, could he find the person?" If the answer is yes, generalize further. Especially rare combinations (small district + specific event) are dangerous.
Enterprise vs public tools
Not all AI tools have the same level of privacy. They differ in three dimensions:
- Data retention: Does it retain the input and for how long?
- Use in training: Does it use the input to train the model?
- Location and contract: Where is the data processed, is there a corporate data processing agreement?
Institutional (the institution's contractual, data processing assurances) tools often offer limits on not using and storing data in education. Public, free tools can store input and use it in training. Rule: sensitive data is processed only in institution-approved, contracted means and in as anonymised a form as possible. Entering customer data into an unapproved tool makes the data processor unauditable.
Step by step: Working securely with confidential data
- Classify the data. Identity / financial / special quality / file data.
- Necessity test. Is this data really necessary for this task? If not, don't use it.
- Anonymize. Remove identifiers, substitute factual equivalents, generalize rare combinations.
- Select vehicle. Only agency-approved, contracted vehicles; No sensitive data available to public tool.
- Work with minimum data. Share the minimum information needed for the task.
- Manage storage and sharing. Where do you store the output, with whom do you share it; Comply with KVKK retention period and third party rules. Leave your mark.
three mini cases
Case 1 — Protecting sensitive data. A claims specialist wanted to ask the AI about a complex health insurance file. Instead of writing the name, ID and diagnosis of the insured, he created an anonymous context such as "52 years of age, female, complementary health policy, planned heart valve surgery, disputed amount". The quality of the output has not decreased; Private data did not go to any cloud. Health data is subject to the highest protection; this habit prevented the violation.
Case 2 — Rare combination trap. “38 years old, female, only pharmacist in [very small county], professional liability policy,” one expert wrote. Although this technically did not contain a name, it directly revealed his identity as he was the only pharmacist in that district. The expert noticed this and generalized it as "small settlement, healthcare profession". Anonymization is not just deleting the name, it is destroying recognizability.
Case 3 — Wrong tool selection. For speed, an employee would upload the customer claim breakdown to a free, public tool. Team policy comes into effect: customer data is processed only in the institution-approved, contracted tool. The employee first anonymized the data and then ran it in the approved tool. If an unapproved tool were used, the risk of data storage and use in education would be uncontrollable.
Four copyable prompts
1) Anonymization assistant:
Remove all personal and identifying data from the following text: name, surname, ID, date of birth, telephone, address, policy number, license plate, title deed, IBAN. Replace them with a factual equivalent with analytical meaning (e.g. "45 years old, male, insurance policy, one-sided collision"). Generalize rare/distinctive combinations (small place + special occupation). Retain medical/technical meaning.Text: [paste]
2) Recognizability check:
Check the following anonymized text for recognition: [text]Ask: Can a stranger find the person with this text? Is there a rare combination (small settlement + specific profession/event), unique date or amount? Highlight each risky point and suggest how to generalize more safely.
3) Data requirement and classification:
Classify and requirement test the data required for the following task:Task: [description] The data I have: [list]For each data: type (identity/financial/special/file), is it required for this task (yes/no), if necessary, how to use it anonymously.Mark unnecessary data as "do not use".
4) Vehicle selection checklist:
Create a checklist before using an AI tool with insurance data. Include questions: Is the vehicle approved by the institution? Is there a data processing agreement? Does it store/use the input in training? Where is the data processed? For which data type is it suitable/unsuitable? Is anonymization enough?
Weak prompt / Strong prompt
Weak: "Evaluate the file of client Ahmet Yılmaz (TC 123..., medical report attached)."
Problem: Identity and sensitive (health) data are shared directly; great risk of KVKK violation.
Strong: "Consider the following anonymized context: age 52, female, supplemental health policy, planned surgery, amount disputed. No identifying information."
Why it's good: Analytical meaning is preserved, identity and sensitive data are not revealed.
comparison chart
Data type
Protection level
Use in artificial intelligence
Name/TC/contact
high
Remove, put equivalent
Financial (income/IBAN)
high
Remove/generalize
Health/special qualifications
highest
Never raw; anonymous + approved vehicle
Policy number/plate/title deed
medium-high
Remove; alone is risky
Aggregate/statistics
low
Available (if no contact)
Common mistakes
- Pasting raw personal/health data. The most frequent and most severe violation.
- Thinking that just deleting the name is enough. Rare combinations may still reveal the identity.
- Relying on saying “delete later.” The tool may be storing/using data in training.
- Unapproved/public driving. Uncontrolled data processing.
- Not managing storage and sharing. Unlimited storage of output, uncontrolled sharing with third parties.
In summary
Insurance data is very sensitive; Special data such as health are subject to the highest protection under KVKK. Classify, challenge and anonymize data before giving it to AI: remove identifiers, introduce factual equivalents, generalize rare combinations. Process sensitive data only in institution-approved, contracted tools and to a minimum extent. Manage storage period and third party sharing within the framework of KVKK and leave a trace. Consult compliance/legal in the specific case; Leaked data cannot be retrieved.
Application task
Get a real damage/reference text (for testing purposes). Anonymise with prompt #1, then check recognition with prompt #2: are there still rare combinations in the text that could reveal the person? Generalize each risk you find. Also evaluate the AI tool you use with checklist number 4: is it suitable for sensitive data?
checklist
- [ ] I classified the data according to their types.
- [ ] I applied the necessity test; I did not use unnecessary data.
- [ ] I removed the identifiers and replaced it with a factual equivalent.
- [ ] I generalized rare/distinctive combinations.
- [ ] I did a recognizability check.
- [ ] I only used company-approved, contracted vehicles.
- [ ] I managed storage and third party sharing in accordance with KVKK; I left a mark.