Gains:
- Being able to distinguish where in the tax and accounting business chain (research, calculation, analysis, declaration, correspondence) artificial intelligence saves real time and where the signature responsibility and judgment is left to the human, according to the task risk level.
- Ability to apply a discipline that verifies each AI output by connecting it to the current official source, recalculating it and passing it through legislation logic.
- Ability to establish a working order that protects taxpayer data within the scope of KVKK and Tax Procedure Law confidentiality obligations, chooses safe vehicles and observes ethical boundaries.
Finance and tax work is to measure, declare and accurately represent the earnings, transactions and assets of a person or institution in accordance with the applicable rules and to the state. Professionals who do this job include CPA (Certified Public Accountant, a professional who has the authority to keep books and prepare declarations), CPA (Sworn-in Financial Advisor, a senior professional who has the authority to certify and audit), tax inspector, financial analyst and accounting manager. The essence of this profession is calculation, but it is something deeper: signing with responsibility. Every declaration signed, every opinion given is a legal commitment and behind it lies the professional responsibility of a professional.
Artificial intelligence (especially large language models, that is, text generating systems such as ChatGPT, Claude, Gemini) creates a serious acceleration in this profession. He summarizes lengthy legislation, lists the steps of a complex calculation, drafts a petition in minutes. But if used incorrectly, it can also cause harm just as quickly: a fabricated article of law, a rate from last year, or a silent calculation error can lead to direct tax loss (missing or late tax accrual) and penalties. Throughout this module, we will position the AI like a senior assistant: a quick research assistant, a tireless drafter, a careful forechecker. But the signer is never a professional. The signature, that is, the responsibility of the declaration and the opinion, always belongs to the human being.
Where does artificial intelligence come in handy in a professional's day?
A typical tax/accounting job goes through roughly the following steps: regulatory research, data and document collection, calculation, financial statement analysis, declaration, official correspondence, and audit preparation. Artificial intelligence touches every link of this chain, but it does not have the same authority in every link.
- Research and scanning: Summarizing a long notification, placing a subject in relevant legal articles, explaining the logic of a process. Artificial intelligence is very powerful here — but every reference it makes needs confirmation.
- Calculation support: Sorting out the steps of a tax base, reminding of a forgotten discount, reestablishing an account. It is useful here, but each number must be recalculated independently.
- Analysis: Interpreting the ratios of a balance sheet, comparing two periods, making a trend visible. It is powerful, but misleading without industry and period context.
- Correspondence and draft: A petition, a letter of agreement, a defense text. Generates draft; The professional verifies the legal basis and duration.
- Signature and decision: Declaration, certification report, official opinion. This is entirely the domain of humans and, when necessary, of competent experts.
Seeing this distinction in a table makes things clear.
Tax/accounting task
The role of artificial intelligence
Whose word is final?
Communiqué/legislation summarization
High: fast distillation
Professional (confirms source)
Set up account steps
Medium: draft + preflight
Professional (re-accounts)
Financial statement interpretation
Medium: configuration
Professional (adds context)
Petition/correspondence draft
High: text production
Professional (verifies the basis)
Declaration signature
None: pre-control only
Professional (legal responsible)
CPA certification, regulated decision
Draft only
Competent expert + human
Tip: View AI as a “thinking and drafting accelerator,” not an “answer machine.” The best professionals use it to quickly grasp a topic and produce a first draft; Not so that he can decide for himself.
Why verification is non-negotiable
The most dangerous feature of large language models is this: they appear confident even when they are not. They may produce a non-existent statute, a fictitious ruling number, an incorrect exemption amount, or an incorrect percentage; This is called hallucination, fabricated information that the model presents as real. In tax, this is fatal because the output becomes a direct declaration, a petition or an opinion and has legal consequences.
Therefore, the immutable rule of this module is this: No item, no amount, no account that comes out of artificial intelligence enters an official output without being verified. Verification is three steps:
- Link to the source: Open each legislative reference (law, notification, ruling) from the official source yourself, see the text, and confirm that it is in force.
- Recalculate: Repeat each base, rate, amount independently, manually or in a spreadsheet.
- Logical filter: "Is this result reasonable according to the logic of the legislation and the reality of the company?" ask.
Security and compliance - critical areas: it's up to the person to decide
Tax and accounting is an area of compliance-critical and legal responsibility from start to finish. Artificial intelligence output does not replace professional signature, declaration responsibility and official audit. A corporate tax declaration, a VAT return report, a transfer pricing file... The final say in all of these; It is subject to the approval of the relevant CPA, CPA or tax lawyer. The AI prepares the draft, the professional verifies it and signs it. “The AI said so” is not a defense; The responsibility always lies with the professional and the taxpayer.
Privacy: taxpayer data is protected by law
Professionals see the taxpayer's most sensitive data: turnover, profit, salaries, bank transactions, contracts. Pasting this data into any artificial intelligence tool is a violation of both the KVKK (Personal Data Protection Law) and the Tax Procedure Law's obligation of confidentiality (the obligation of a professional to keep taxpayer information confidential). Corporate tools that guarantee data retention should be used; If it cannot be used, the data should be anonymized (identifiers such as names, TR ID and title are masked and given to the model).
Caution: A payroll, trial balance, or contract you paste into a publicly available AI tool may be considered "shared with a third party." If in doubt, anonymize the data or do not upload it at all.
three mini cases
Case 1 — Fabricated ruling. A financial advisor asks the AI an exception question; The model says "According to the Revenue Administration's ruling numbered 2021/47, this exception applies" and explains the number in detail. The consultant puts this into consideration. During the investigation, it turns out that such a feature does not exist; The basis of the opinion collapses and the taxpayer faces a 180,000 TL assessment. The correct way was to confirm each ruling number exactly from the Revenue Administration system.
Case 2 — Old rate. An intern asks the AI about an income tax bracket; The model gives the tariff of the previous year. The calculation made with this tariff will result in 24,000 TL less tax. If the declaration is submitted in this form, tax loss and penalty will occur. The right way: to get the official version of the tariff for that year from the Official Gazette/GIB.
Case 3 — Good handling. A CPA gives a 60-page section of the VAT General Application Communiqué to the artificial intelligence and says, "Based only on this text, extract the return process steps and necessary documents item by item, and write 'not in the text' for what is not in the text." 2 hours of reading is reduced to 15 minutes. Then he compares each step with the text of the notification. Saves time, doesn't make mistakes. This is the targeted balance.
Weak prompt / Strong prompt
A prompt is the instruction you give to the AI; Its quality directly determines the quality of the output.
Weak prompt:
Provide information about VAT exemption.
This demand produces a text without context, without period, and without verification; The model fills in the gaps with fitting.
Powerful prompt:
Your role: meticulous tax research assistant working for a financial advisor. Task: According to the text of the notification I pasted below, draw out the application conditions of the export exemption, item by item. Rules: - Rely only on the text I have given; Write "not in the text" what is not in the text. - Do not make up any article, notification or ruling number. - Mark every point you are not sure of as "needs to be verified". Format: numbered steps + list of "sources to be confirmed" at the end. [COMMUNIQUE TEXT: ...]
The second prompt gives the role, task, resource, boundary, and format; narrows the field of hallucination.
Four principles of using artificial intelligence safely
Keep the following four principles on your desk like a card; Each unit is built on these.
1) GIVE CONTEXT: role + task + resource + period/year + target + limits.2) CONNECT TO SOURCE: every item and amount must be based on a traceable official source.3) RECALCULATE: never entrust the arithmetic to the model, check.4) HUMAN SIGNATURES: declaration, opinion and approval of a professional/expert in the regulated field is required.
A role description template you can use when starting each new conversation improves the quality of the output from the start:
You are a meticulous tax/accounting assistant working for a financial advisor. Your immutable rules: - Rely only on the sources I give you and the year I specify. - Do not fabricate any law articles, notifications, ruling numbers or rates. - Mark anything you are not sure of as "must be verified". - Give the arithmetic result but add a note "must be recalculated". Follow these rules throughout this conversation. If you understand, say "I'm ready" and wait for the task.
A simple template to keep the verification trail saves your work:
Output: [what I produced: calculation / article / petition]Source: [which law/notification/decree, which year, which URL]Recalculation: [was it done, was the result valid]Approval: [who, when]Status: [confirmed / not verified / waiting for expert approval]
Common mistakes
- Mistaking the model for authority. Fluent and confident language is no guarantee of accuracy; The hallucination is also fluid.
- Requesting an account without specifying the period. Rates vary by year; The output without the year is most likely incorrect.
- Accepting the reference without opening it. Article and ruling numbers cannot be used without being seen from the official source.
- Uploading confidential data to an uncontrolled vehicle. Violation of KVKK and VUK confidentiality results in criminal and professional consequences.
- Not keeping a verification trail. “Where did this amount/item come from?” If you cannot answer the question, the output is not ready for publication.
In summary
Artificial intelligence is a powerful assistant that accelerates every step in finance and tax, from research to drafting, from calculation to analysis. But its authority varies depending on the risk level of the task: in screening it is high, in accounting it is at pre-flight level, in statements and opinions it is only draft. The constant rule is verification; No item, amount or account is entered into the official printout without confirmation. Taxpayer data is protected by KVKK and VUK confidentiality obligations; The final responsibility and signature always belong to the professional. Throughout this module we will learn each technique along with this security framework.
Application task
Choose your own (real or sample) taxpayer file. Divide your typical workflow for this file into five steps (research, calculation, analysis, correspondence, statement), similar to the table above. For each step, fill in the AI's role (high/medium/none) and the "whose say is final" column. Then create a sample record using the verification trail template in one step: which output, which source, was it recalculated, who approved it, what is its status.
checklist
- [ ] I positioned AI as an "assistant", not as a "signing professional".
- [ ] I determined the risk level of each mission and who had the final say.
- [ ] I planned three-step verification (source, recalculation, logic) for each critical output.
- [ ] I have specified the period/year in each prompt.
- [ ] I evaluated the taxpayer data in terms of KVKK and VUK confidentiality; I anonymized it if necessary.
- [ ] I wrote my prompts with role, resource, boundary and format.
- [ ] I have committed to using the verification trail template.