Gains:
- Ability to classify taxpayer and financial data within the framework of KVKK and Tax Procedure Law confidentiality obligation and determine the level of protection
- Ability to apply anonymization, masking and safe/institutional tool selection steps before giving data to artificial intelligence
- Ability to gain the habit of avoiding unsupervised driving by knowing the legal, criminal and professional consequences of privacy violations.
A professional sees the taxpayer's most intimate information: turnover, profit, bank movements, employee salaries, TR ID numbers, contracts, partnership structures. This information is protected by two separate legal layers: KVKK (Personal Data Protection Law, the law prohibiting the unauthorized processing of personal information) and the confidentiality obligation of the Tax Procedure Law (the obligation of the professional to keep confidential the information he/she learns due to his/her duty). In the age of artificial intelligence, this protection faces a new threat: sticking sensitive data in an uncontrolled tool. The principle of this unit is single: data is protected before entering the AI; No sensitive data enters the unsupervised vehicle.
Two-layer protection and why it should be taken seriously
- KVKK layer: All information (name, TR ID, salary, contact) belonging to real persons is personal data. Unauthorized processing, transfer and inadequate protection will result in administrative fines and compensation.
- VUK confidentiality layer: A professional cannot share the taxpayer's financial secrets with third parties. Violation carries both criminal and professional (disciplinary) consequences.
A trial balance or payroll pasted into a publicly available AI tool may be considered “transferred to a third party.” It is often unclear how the tool stores the data and whether it is used in training. This uncertainty alone is risk enough.
Attention: The idea that "no one will see it anyway" is not a legal defence. Once data enters an uncontrolled system, it is out of your control. A privacy breach is dangerous not because of the possibility of it happening, but because it is irreversible once it happens.
Before giving data to AI: protection steps
- Classify. Is the data you have personal, financial secret or public? Determine the protection level accordingly.
- Question necessity. Is it really necessary to give this data to the model, or can the task be done with anonymous/summary data?
- Anonymize/mask. Their names are K1, K2; companies Ş1, Ş2; Replace TR ID and account numbers with a mask. Generalize unique details that provide indirect diagnosis.
- Choose a safe vehicle. If possible, use a corporate/approved tool that guarantees data retention.
- Keep the trail. Record which data you gave to which tool and with what protection.
Data type
example
conservation approach
identity data
Name, TR ID
Mask/anonymize; never give raw
financial secret
Turnover, profit, bank
Summarize/anonymize; enterprise vehicle
relationship data
partner, supplier
Generalize; remove unique detail
public
declared balance sheet
can be processed normally
How is anonymization done in practice?
Anonymization is making the identity indistinguishable while keeping the data useful. For example, you will have a payroll analyzed: you put a sequence number instead of name and TR ID, the analysis of the salary distribution will not be disrupted, but no one can be identified. The critical point is the indirect diagnosis: expressions such as "the only female general manager" or "the only 92-year-old in a company of 35 people" give the person away even if you delete the name; These are also generalized.
three mini cases
Case 1 — Payroll leakage. An intern uploads a payroll of 3,500 people (name, ID card, salary) to a free web tool and says "summarize". Data runs the risk of being mixed into the vehicle's repository; Both KVKK and VUK confidentiality violations occur. If the customer finds out, it's over. The right way was to use a corporate-approved vehicle or to only give the salary distribution by masking your name/TR ID.
Case 2 — Indirect diagnosis. A consultant gives the model a list in which he deletes the names, but leaves the phrase "the company's only foreign national manager." This statement gives the person away; Anonymization is missing. Lesson: unique details that provide indirect diagnosis should also be generalized.
Case 3 — Good handling. Before feeding sensitive data to the model, an office goes through a standard anonymization step: using a template, it converts names, IDs, and brand names into codes, then has it analyzed. The quality of analysis does not decrease, no identity is leaked. Privacy and efficiency are protected together.
Weak prompt / Strong prompt
Weak prompt:
Summarize the following payroll: [Ahmet Yılmaz, TC 123..., salary 45,000; ...]
Raw personal data enters the vehicle uncontrolled; It is a direct violation.
Powerful prompt:
I'll give you a data set. BEFORE processing, apply the following anonymization and show me the anonymized version as well:- Make the person names K1, K2... - Remove the TR ID and account numbers completely.- Make the company/brand names T1, K2.... Generalize the unique details that could indirectly identify the person. Then work only on the anonymous version.[DATA (masked as much as possible): ...]
Note: It is safest not to enter raw personal data at all; If you enter, you must be masked and in a safe vehicle.
For data classification help:
Classify each of the following fields: is it personal data, is it a financial secret, is it public? Write the recommended protection level for each.Fields: [fieldnames, sample RETURN]
To generate an anonymization template:
Produce a standard anonymization checklist that I would use in a payroll/trial balance analysis: which fields should be masked, which should be omitted, what should I look for for indirect diagnosis?
For vehicle safety assessment:
What questions should I ask when evaluating whether an AI tool is suitable for sensitive taxpayer data (data retention, use in education, corporate contract, location)? Create a checklist.
Common mistakes
- Entering raw personal data. Never enter the vehicle without Name/TR ID verification.
- Skipping indirect diagnosis. The unique detail gives away the identity even if the name is deleted.
- Not asking about vehicle safety. Sensitive data is not given without knowing how the tool stores the data.
- Not questioning the necessity. If the task can be done with anonymous/summary data, raw data is not given at all.
- Not keeping track. If it is not recorded which data was given where, the breach cannot be traced.
In summary
Taxpayer data is protected in two layers by both KVKK and VUK confidentiality obligations, and the biggest new risk in the age of artificial intelligence is sticking sensitive data to an uncontrolled tool. Before data is entered into the model, it is classified, its necessity is questioned, identifiers such as name/TR ID are masked and details that provide indirect diagnosis are generalized; If possible, only a secure, data retention-guaranteed corporate tool is used. A privacy breach is an irreversible harm; The most secure data is data that is never entered.
Application task
Take a typical sensitive file (e.g. payroll or trial balance) in your hand. First classify their fields (personal/financial secret/public). Then apply an anonymization template: list which areas you will mask, which you will remove, and what you will look for for indirect identification. Finally, evaluate the security suitability of the artificial intelligence tool you use/will use with a checklist.
checklist
- [ ] I have classified the data as personal/financial secret/public.
- [ ] I questioned whether it was really necessary to provide the raw data.
- [ ] I masked/removed the name, TR ID and account numbers.
- [ ] I have generalized unique details that provide indirect diagnosis.
- [ ] I have only used secure/enterprise tools with data retention guarantees.
- [ ] I evaluated the data policy (storage, training) of the tool.
- [ ] I recorded which data I gave where as a trace.