Gains:
- Ability to distinguish where AI saves real time in the finance workflow and where decisions and responsibility should remain with humans
- Ability to apply a three-layered discipline that verifies each financial outcome by order of magnitude, independent reproduction and source control
- Get into the habit of anonymizing context and editing prompts to leverage AI without sharing sensitive financial data
When you look at a finance professional's day, the picture is similar in most teams: reading long spreadsheets, calculating ratios, examining budget-to-actual variance, creating formulas in Excel, preparing presentations for management, email and meetings. So the time devoted to the real "finance judgment" part, that is, what a number means and what decision it requires, gets overwhelmed by repetitive work. This is where artificial intelligence (AI for short; software that works on text and numbers with a large language model) comes into play. AI doesn't make the decision for you; It prepares you for the decision, produces a draft, speeds up the calculation and puts processed information in front of you. Throughout this module, we will position AI not as an “automatic analyst” but as a disciplined assistant whose output is verified every time.
In this first unit, we clarify three things: at what stages of the finance workflow does AI add real value, what decisions must remain strictly human, and what the verification and privacy discipline you must adhere to when doing so. Without this roof installed correctly, techniques on subsequent units can become dangerous; because finance is a field where mistakes are measured in money.
Concepts: Hallucination: AI's convincing fabrication of a number, source, or item that does not actually exist. Context: The input you give to the AI (table, assumption, question). Verification: Checking the output in an independent way. These three concepts are the backbone of the entire module.
In Which Businesses Is AI Accelerator, In Which Businesses Is It Risky?
Financial affairs fall on a two-pronged spectrum in terms of their results. At one end are reversible, low-risk prep work; At the other end, there are irreversible decisions that result in money, reputation and legal liability. The value of AI varies depending on where you stand on this spectrum.
business type
AI contribution
The role of the professional
Table/report summary
Extraction from long table
Compare items to source
Ratio/account preliminary study
Formula setting, first number
Definition, unit and order control
Excel formula / macro
Skeleton and logic generation
Validation with test data
Report/presentation draft
Propose narrative and structure
Linking each figure to the source
Scenario/prediction fiction
Proposing a hypothetical framework
Reasonableness and backtest
Investment/credit decision
Analysis material
Final decision and signature
The rule is simple: the risk of an AI output equals the damage it will incur if that output makes an error. Misspelling a chart title is harmless; Miscalculating a credit limit can result in a loss of millions. So the first question to ask before using the output is: "What happens if this is wrong and who will notice?"
Attention: AI produces fluent and confident text. Fluency is no guarantee of accuracy. A language model can "make up" convincing figures, item names, and legislation even when it has no real data. In finance, this does not remain on paper; It turns into a wrong decision and money.
Decisions That Should Stay with People
Some decisions should never be fully automated; carries technical, legal and ethical risks:
- Decisions and approvals: Final approval of investment, loan allocation, financing, dividend and pricing decisions.
- Declaration and signature: Statements such as "these tables are correct" or "this amount complies with the legislation" require a human signature.
- Advice to the customer: Personalized investment advice is an area subject to legislation that requires authority and responsibility.
- Decision with confidential data: Transactions made with results and personal data that are not publicly disclosed.
Warning: Even if AI produces a result like "45 million TL credit limit is suitable", it is unacceptable to apply it without an expert validating it with the collateral, cash flow and risk profile. Any output that leads to a decision must be independently verified and approved by a competent person before being implemented.
Verification Discipline: Three-Layer Control
Apply a three-layer control to use AI output with the eyes of the editor and auditor, rather than blindly. This is the basic reflex we will repeat throughout the module.
- Unit and rank (sanity check): Is the unit of the result correct? Is the size reasonable? If a company with a turnover of 40 million produces a net profit of 900 million, there is a mistake somewhere.
- Independent reproduction: Reproduce the calculation manually, with a calculator, or with a short Excel/Python run. If two different methods give the same result, trust increases.
- Source verification: Every item, rate definition, legislative article and market data provided by AI must be verified verbatim from the official source.
Verification prompt (makes it easier to audit the output): "Clearly list ALL assumptions, formulas, and input figures you used in performing the following analysis. Show each intermediate step on a separate line. Mark each market/regulatory value you provide with the 'source required' label. Do not make up any numbers you do not know for sure; write 'not sure' if you are not sure."
Weak Prompt / Strong Prompt
WEAK: "Interpret the financials of this company." (Result: no context, general sentences; it is unclear which figure, which period, which basis of comparison.) STRONG: "Interpret the following 2023 and 2024 income statement items. Calculate the change in turnover, gross margin and operating profit as a percentage, compare the two years and list 3 striking findings with justification. Use only the figures I have given; if there is missing data, 'no data' [pencils here]"
The difference is in context. Powerful prompt; It includes the period, the basis of comparison, the desired metric, and the "use only the given data" constraint. This one-sentence discipline greatly reduces the risk of hallucinations.
Mini Cases
Case 1 — Fake growth rate. An analyst asks AI "what is the industry average growth?" he asks. AI firmly says "14.3%". The analyst asks for the source before putting it in the report; AI cannot give a clear source. A brief research shows that real industry growth is around 6%. Thanks to verification, the hallucination was caught before it entered the report.
Case 2 — Order error. While AI makes a calculation for EBITDA (earnings before interest, tax and depreciation) of 12 million TL, it shows the result as "1.2 billion". With a rank check, the analyst sees that this exceeds turnover by 100 times; unit (thousand TL / TL confusion) error is found and corrected.
Case 3 — Risk of privacy breach. An expert is about to paste yet-to-be-released quarterly results into a public tool. It remembers the corporate policy, anonymizes the data (makes the company name "Company A", gives the amounts pro rata) and uses the corporate tool. This way, analysis is done, but sensitive data does not get out.
Principle of Working with Sensitive Financial Data
The most sensitive aspect of finance is data. Non-public results, customer information and personal data; It carries risks in terms of privacy, KVKK and insider trading. Basic principle: anonymize data before sharing, only ask for structure if possible.
Anonymised prompt pattern: "I want to interpret the income statement of a retail company. Instead of real numbers, I give proportions: assuming turnover is 100, cost of goods sold is 62, operating expenses are 24, net profit is 8. Compare this structure with a reasonable retail margin for the industry and tell me the points to consider."
Common mistakes
- Using the output without validating it. “AI said” is not a justification; Each issue requires independent control.
- Asking questions without context. When the period, unit and basis of comparison are not given, the interpretation becomes general and misleading.
- Sharing confidential data without thinking. Undisclosed financial data, personal data and customer information should not be released without anonymization.
- Confusing precise language with accuracy. The more confident the AI speaks, the more careful you should be; Confident tone is not evidence.
- Delegating the decision to AI. Investment, loan and pricing decisions remain with the individual; AI only produces materials.
In summary
AI speeds up the repetitive and time-consuming parts of finance work: summarizing, accounting, formulating, drafting. However, the decision and responsibility remain with the person. Each output must pass three layers of control (rank, independent reproduction, source). Writing prompts with context and anonymizing sensitive data are two key habits that we will repeat in each unit of this module. When you use AI with discipline, you gain speed, when you use it without discipline, you lose money.
Application task
Choose a financial task from your own business or a fictitious company (e.g., two annual income statement interpretations). First write a weak prompt and get the output. Then apply the powerful prompt pattern from this unit: add the period, metric, comparison basis, and the "use only the given data" constraint. Put the two printouts side by side and write the difference. Then independently verify at least two figures in the strong printout (hand account or source) and note what you found in which verification.
checklist
- [ ] I added period, unit and comparison basis to my prompt.
- [ ] I wrote the "Only use the given data, don't make it up" constraint.
- [ ] I checked the order of at least two digits of the output.
- [ ] I have reproduced at least one figure independently.
- [ ] I marked the legislation/market values as "resource required".
- [ ] If there was sensitive data, I anonymized it or used the enterprise tool.
- [ ] I confirmed that the final decision remains with the person.