Gains:
- Ability to summarize income statement, balance sheet and cash flow statement with AI and turn it into meaningful insight
- Ability to translate raw financial data into management language with a structured prompt that includes role, context and output format
- Ability to perform check steps to verify each key figure in the AI summary against the source
You have an income statement, balance sheet or cash flow statement in front of you. The numbers are correct, the equation holds, but the general manager turns to you in the meeting and asks "so what does this mean?" he asks. This is where artificial intelligence (AI) creates value most quickly for an accounting and finance professional: translating the raw picture into a story that the decision maker can understand.
In this unit, we will learn to use AI like a "senior analyst assistant". But let's set a framework from the beginning: AI is not a calculator. The Big Language Model (LLM, a type of artificial intelligence that produces text like a human) makes sense of numbers, interprets them, and puts them into narrative; but it doesn't do guaranteed arithmetic. So at the heart of this unit will be two things: writing good prompts (the instructions you give to the model) and validating each output against the source.
How Does AI “Read” the Financial Statement?
The model sees the table as a plain text array, not as a grid like in Excel. The concept of "cell B4" or "that column" does not exist for him; It reads whatever you write. Therefore, the more organized and labeled you provide the data, the more accurate the interpretation will be.
Give him three things clearly every time:
- Data — item names and numbers, with period labels (such as 2025 / 2024).
- Context — company's industry, currency, period, special circumstances if any (one-off revenue, exchange rate impact).
- Target audience — who will the brief be addressed to? The language for the board of directors, the bank, the tax office and the team is completely different.
Tip: Instead of saying "Interpret this table", say "Interpret this table in 5 points to a general manager without a finance background." As the target audience becomes clearer, the depth and language of the output falls into place.
Step by Step: From Raw Painting to Insight
A solid comment flow consists of these five steps:
- Paste the data by labeling it. Let each row have its name and period. Correct the mixed copied data first.
- Define the role and output format. “You are a corporate finance analyst” + “answer with these headings.”
- Set restrictions. "Only use the numbers in the data, write 'no data' for those that are not."
- Ask for formula transparency. Show the calculation in parentheses for each rate/percentage.
- Validate and link to action. Verify at least one account manually, then add a "what to do" section.
Weak Prompt / Strong Prompt
You get two different results with the same data:
Weak prompt:Interpret this income statement. [table]
This often produces a superficial answer at the level of "incomes have increased, so have expenses"; It won't be of any use to you in the meeting.
Strong prompt:Your role: a corporate finance analyst.Task: Analyze the following 2025 and 2024 comparative income statement.Context: The company is a food manufacturing company, the currency is TL.Output format:1) Three-sentence executive summary2) The 3 most notable changes (amount + percentage + probable cause)3) 2 critical questions management should askConstraint: Use only numbers in the <data> block. Do not generate any numbers that are not in the table; If missing, write "no data". Show the calculation in parentheses for each percentage.<data>Net Sales: 2025: 48,200,000 | 2024: 39,500,000Cost of Sales: 2025: 31,300,000 | 2024: 24.100.000Operating Expenses: 2025: 9.800.000 | 2024: 8.200.000Net Profit: 2025: 4.900.000 | 2024: 5,100,000</data>
This prompt; notices that net sales are up about 22% but net profit is down and asks “has the cost increase eaten up the sales growth?” It raises the question. This is exactly the insight management wants to hear.
The Role of AI in Three Fundamental Tables
Table
AI's strength
You need to verify
income statement
Margin analysis, period comparison, trend narrative
Accuracy of percentages and totals
balance sheet
Interpreting liquidity/liability ratios, item relationships
The ratio formula is applied correctly
cash flow statement
Explaining the distinction between activity/investment/financing, cash burn interpretation
Items are classified into the correct section
Mini Cases
Case 1 — Cost eroding profit. In a textile company, net sales increased from 39.5M to 48.2M (an increase of approximately 22%), but net profit decreased from 5.1M to 4.9M. AI flagged that cost of sales increased by 30% (24.1M → 31.3M) and gross margin decreased from 39% to 35%. Comment: "Turnover is growing but profitability is weakening; raw material costs must be scrutinized." With this single insight, the analyst opened an agenda to the purchasing team.
Case 2 — Misleading growth. The "Other Income" item in the income statement of a service company suddenly increased by 2M. AI noted that this could come from a one-time asset sale and should not be confused with operating profit. By explaining this distinction to the board, the CFO prevented the false perception that "our business is doing great."
Case 3 — Manually caught error. An intern checked AI's summary "operating expense increased by 5%": (9.8 − 8.2) / 8.2 ≈ 19.5%. The model gave the wrong percentage. If there was no verification, 5% would be included in the report. This case shows why the "confirm every percentage at least once" rule is non-negotiable.
"Don't Produce a Number That Doesn't Exist" Rule
The most dangerous mistake in financial interpretation is when the model fabricates a convincing number — this is called hallucination (when the model confidently produces information that does not actually exist). You can reduce this significantly with two sentences:
Constraint: Only use digits in the <data> block. When calculating any ratio or total, indicate in parentheses the formula you used (e.g. gross margin = gross profit / net sales). Mark what is not in the data as "cannot be calculated". Separate the comment you are not sure about with the "assumption" tag.
A control prompt also works:
Check the following financial summary against the source table. Match each numerical assertion to the table; List each expression that does not match or has no equivalent in the table as "NOT VERIFIED".<summary>...</summary><table>...</table>
Attention: Do not use the percentages given by the AI with your eyes closed. When it says “Net sales increased by 22%,” verify at least one item head-on: (48.2 − 39.5) / 39.5 ≈ 22%. If they match, confidence increases; If it does not match, the entire output is suspect and should be requested from the beginning.
Linking Comment to Action
A good financial summary goes beyond “what happened” to “what to do.” Add this to the prompt:
Add a "Suggestion" heading as the last section: for each significant finding, suggest a concrete action that the finance team can take (e.g. request a breakdown of the relevant cost item). Recommendations should be based solely on data and should not contain speculation.
Common mistakes
- Pasting the table without context. Without a period label, currency, and industry, the model makes incorrect assumptions.
- Putting percentages in the report without verifying them. The model can be wrong in arithmetic; The most dangerous mistake is the silently wrong number.
- Not specifying the target audience. The same summary does not suit both the board and the bank; Language and depth vary depending on the target.
- Expecting a "definitive result" from the model. AI provides interpretations and possible causes; The certainty of the cause can only be confirmed by data.
- Thinking of one-off items as normal income. It is misleading to unquestioningly present the jump in "other income" as growth.
In summary
- AI is a powerful assistant that translates the raw financial picture into the language of the decision maker; But since it sees the table as text, you must provide the data in a labeled, periodic and orderly manner.
- Powerful prompts that clearly state the role, context, target audience and output format turn superficial comment into accurate insight.
- The constraints “only use digits in the data” and “show formula” significantly reduce the risk of hallucinations.
- Verify each key percentage and total against the source; Manually confirming at least one calculation tests the reliability of the entire output.
- A good summary moves from “what happened” to “what to do”; Include a concrete, data-based recommendation section in the prompt.
Application task
Take your (or a sample) comparative income statement. Fill the "Strong prompt" template above with your own data and run it. Then: (1) manually calculate each percentage given by the model once, (2) flag the incorrect one, (3) have the summary checked at the control prompt. Finally, prepare a three-item "executive summary" and a "suggestion" section and prepare it to a quality that you can send to a colleague.
checklist
- [ ] I provided the data by labeling it with item names, period and currency.
- [ ] I specified the role, context, target audience and output format in the prompt.
- [ ] I added the "Only use digits in the data" constraint.
- [ ] I wanted it to show the formula for each percentage/ratio.
- [ ] I manually verified at least one account and corrected the one that didn't match.
- [ ] I completed the summary with a "what to do" suggestion section.