Gains:
- Ability to calculate base, exemption, deduction and tax amount step by step with artificial intelligence and make the logic chain transparently visible
- Ability to catch silent calculation errors by independently recalculating (by hand or with a table) each arithmetic result produced by artificial intelligence
- Being able to gain the habit of not trusting old values in the memory of artificial intelligence, knowing that the rate, exception limit and periodic parameters are constantly changing.
Tax calculation is the task of finding the tax to be paid by correctly determining the amount on which tax will be charged (the base amount, that is, the basic amount on which the tax is calculated) and applying the relevant rate. It sounds simple, but it's actually a chain of exclusions, discounts, periodic rates, and sequential steps. Artificial intelligence is very useful in establishing and making visible this chain step by step; But here we encounter the model's greatest weakness: large language models silently make mistakes in arithmetic and present the wrong result with complete confidence. The immutable rule of this unit is therefore clear: every number generated by the artificial intelligence is recalculated independently.
Anatomy of the base: making the steps visible
A good tax calculation is never a single number; It is a chain of steps. Use AI to build this chain transparently, as each step becomes individually verifiable.
A typical capital gains tax chain works like this:
- Trading profit / gross earnings: Starting point.
- Additions: Expenses not allowed by law (KKEG) are added.
- Deductions and exclusions: Items permitted by legislation are deducted.
- Previous year's loss offset: If any, carried forward losses are deducted.
- Base: Taxable amount is formed.
- Rate application: Tax is calculated at the current rate.
- Offsets: Prepaid advance tax, withholding tax, etc. is deducted.
- Payable/refund: Final result.
Tip: Tell the AI "don't give the result as a single number; show each step with its amount on a separate line." It is easier to catch the error when the steps are discrete; In an aggregate result the error is hidden.
Why recalculation is non-negotiable
A language model might write "1,240,000 × 23% = 285,200" and that would look correct. But sometimes the model produces a close but incorrect number instead of "285,200" and presents it without hesitation. Even a 15,000 TL calculation error in taxes will result in tax loss and penalties. So every multiplication, ratio and sum is checked either with a standalone calculator/table or by rebuilding the model in a different way.
Two practical techniques:
- Reverse check: Verify by going back from the result. Tax/rate = base?
- Standalone tool: Repeat the same calculation with formula in spreadsheet (Excel); The two results should match exactly.
Periodic parameters trap
Tax rates, exemption limits, deduction amounts and tariffs change almost every year, sometimes during the year. The model's information is frozen in the training date and may mistake last year's rate as "current". This silently corrupts the entire account. Therefore:
- No rates or limits are retrieved from the model's memory.
- Each periodic parameter is taken from the official source (Presidential Decree, communiqué, GIB) for that year and you give it to the model.
- You clearly write the instruction to the model "use the rates I gave you, DO NOT use the rate you know".
Parameter type
Where to buy
How to export to model
tax rate
Official regulation of that year
It is clearly written in the claim
Exclusion/discount limit
Current notification/decision
It is clearly written in the claim
Income tax tariff
Official tariff of that year
Slices are given in full
Depreciation rate
VUK depreciation list
Is taken from the list
three mini cases
Case 1 — Silent multiplication error. An advisor has the artificial intelligence calculate a corporate income. The model orders all the steps correctly, but gets the final multiplication wrong: it writes 426,300 instead of the real 462,300. The figure could have been overlooked because it seemed reasonable. When the consultant recalculates in Excel, he finds the difference of 36,000 TL. Lesson: even if the steps are correct, the arithmetic is checked separately.
Case 2 — Old tariff. An intern uses the tariff given by the model to calculate income tax; The tariff is from the previous year. Even though the base is the same, the tax will be 18,500 TL incorrect because the brackets have changed. The correct way was to give the model the official tariff of that year and say "just use this".
Case 3 — Good handling. A CPA has the artificial intelligence set up the steps in a complex exception calculation, asks for each step on a separate line, and gives the rates himself. Then it repeats the entire calculation with the formula in Excel; The two results match exactly. Artificial intelligence has increased the speed of thinking, and the verification remains with the CPA. A 40-minute job is reduced to 10 minutes, no errors will occur.
Weak prompt / Strong prompt
Weak prompt:
Calculate the corporate tax of this company, you know the rate.
The model uses the (possibly old) rate from its memory and returns a single result; Both ratio and arithmetic are risky.
Powerful prompt:
Your role: tax calculation assistant.Task: Calculate corporate profits tax step by step with the following data.Rules:- Only use the rate (X%) I give you; Do not use the rate you know. - Show each step with its amount on a separate line (commercial profit, CCEG, discount, base, tax, offset, payable). - State that the results must be "recalculated manually / with Excel". Data: [commercial profit, CCEG, discounts, previous year's loss, provisional tax, rate: X% (taken from official source for this year)]
The second prompt fixes the rate, separates the steps, and reminds of the verification obligation.
Reverse check prompt for recalculation:
Verify the following calculation in reverse:- Find the base again by going back from the given result.- Is the resulting base the same as the base given at the beginning?- If there is a difference, show in which step it occurred.[CALCULATION STEPS: ...]
To request an individual table formula:
I will set up this account in a spreadsheet. Write the cell formula for each step (e.g. base = trading_profit + CUEG - discount - past_loss). Just give the formulas and I will produce the result in the table.
Ratio/parameter confirmation reminder:
List the following parameters I use in this account and add a note for each "which year should be confirmed from the official source": rate, exception limit, discount amount, tariff slices.
Common mistakes
- Relying on arithmetic. The model silently errs on multiplication/addition; each result is recalculated.
- Getting the ratio from the model. Periodic parameters are taken from the official source and given to the model.
- Wanting collective results. An output that does not separate the steps hides the error.
- Not specifying the year. If it is not stated which period it belongs to, the model uses the old parameter.
- Being satisfied with one method. The critical calculation should be cross-checked by having it performed on both the model and the independent tool (Excel).
In summary
Tax accounting is a chain of steps, and AI is powerful in making this chain visible; but it silently makes mistakes in arithmetic and wears out periodic parameters. Therefore, each number is recalculated independently, each rate and limit is taken from the official source for that year and given to the model, and no periodic values are accepted from the model's memory. Separating steps, reverse checking, and cross verifying with an independent table are the three safety pillars of this unit.
Application task
Select a real or sample tax account (corporate, provisional or income tax). Get a step-by-step calculation from artificial intelligence with a powerful prompt template; You give the rate. Then independently repeat the same calculation with the formula in Excel and compare the two results. If there is a difference, determine at which step it is. Confirm every periodic parameter you use from the official source and make a note of it.
checklist
- [ ] I had the account set up step by step, each amount on a separate line.
- [ ] I took all the rates and limits from the official source and gave them to the model.
- [ ] I wrote the instruction "use the ratio you know" to the model.
- [ ] I recalculated each result independently in Excel/by hand.
- [ ] I verified the base backwards with the reverse check.
- [ ] I have clearly stated the period/year information in each account.
- [ ] I detected and fixed the differences and recorded them in the verification trace.