An AI agent can produce a mathematically correct answer that is financially wrong. This is one of the less obvious problems that emerges when language models are connected to spreadsheets and financial calculation systems. The arithmetic may be perfectly valid, the output may appear reasonable, and yet the underlying result may be based on incompatible assumptions.
Consider a development feasibility model containing a rental area of 25,000 square metres and a rental rate of 250. An agent multiplies the two values and produces revenue of 6,250,000. At first glance, there is nothing wrong with the calculation. However, the result depends entirely on what the rental rate represents. Is it 250 AED per square metre per year, 250 AED per square foot per year, or 250 AED per square metre per month? The numerical input is identical in all three cases, but the financial interpretation is substantially different.
This illustrates a fundamental problem in financial automation: numbers cannot be interpreted independently of their context. A reliable financial modelling agent must understand not only the value of an input but also its unit of measurement, currency and time basis. These attributes should be treated as part of the underlying data structure rather than optional descriptions attached to a spreadsheet cell.
A number is not a complete financial input
Traditional spreadsheets communicate meaning through a combination of cell labels, formatting, formulas and surrounding assumptions. An analyst looking at a workbook may understand that a particular value represents annual rental income because the worksheet heading, row description and neighbouring calculations provide sufficient context.
An automated system extracting the same value into a separate calculation environment may not have access to all that information. Even when it can identify the numerical value and its cell reference, it still needs to establish what the number represents.
For example, consider a spreadsheet containing a rental area of 25,000, a rental rate of 250 and a calculated revenue of 6,250,000. The relationship appears straightforward, but the calculation is meaningful only if the area and rental rate use compatible measurement units and the resulting revenue has a clearly defined currency and time period.
A more reliable approach is to represent financial inputs as structured records. Instead of passing an isolated number to the calculation engine, the extraction process can preserve the value alongside its associated metadata.
An illustrative representation might look like this:
{
"name": "rental_rate",
"value": 250,
"unit": "AED/m2/year",
"currency": "AED",
"period": "annual",
"source": "approved_assumption"
}
This is a simplified example rather than a universal financial-data standard. Its purpose is to demonstrate how the meaning of an input can be preserved when information moves between a spreadsheet, an AI interpretation layer and a calculation engine.
The distinction becomes particularly important when the original workbook contains assumptions from multiple sources. A financial model may combine land acquisition information, construction estimates, rental benchmarks and financing assumptions. These inputs may have been prepared by different people using different conventions. Without an explicit metadata layer, the agent has to infer whether the values are compatible, which creates unnecessary uncertainty.
Unit consistency should be enforced before calculation
Dimensional consistency is a basic principle in engineering and scientific calculations, but it is equally important in financial modelling. Every calculation has an expected relationship between the dimensions of its inputs and its output.
Consider the standard rental revenue calculation:
Rental Revenue = Area × Rental Rate
When the area is expressed in square metres and the rental rate is expressed in AED per square metre per year, the units cancel correctly:
m² × AED/m²/year = AED/year
The resulting value represents annual revenue in AED. This is not simply a matter of presentation. The unit relationship confirms that the calculation is conceptually valid.
Now suppose the rental rate is expressed in AED per square foot per year while the area remains in square metres. Multiplying the two values directly produces a numerical result, but the resulting figure cannot be interpreted as valid annual revenue without first converting the area or rental rate.
An AI agent should therefore distinguish between a calculation that can be executed directly and one that requires an explicit unit conversion. It should identify the original measurement convention, apply a documented conversion factor and retain the information needed to trace how the final value was derived.
This becomes particularly relevant in international development projects, where land areas, built-up areas and construction costs may be presented using different measurement systems. A model that combines these assumptions without validating their units can produce internally consistent-looking outputs while containing errors that are difficult to identify during review.
Currency conversion is a calculation, not a formatting operation
Currency introduces another layer of complexity because financial models frequently combine assumptions expressed in different currencies. A development project may have land acquisition costs in one currency, construction expenditure in another and rental income benchmarked against an international market.
Replacing a currency label does not convert the underlying financial value. If a model contains revenue of 10 million AED and the reporting output requires USD, the calculation engine needs an exchange-rate assumption before it can produce the converted amount.
That assumption should contain more information than a numerical exchange rate. The system needs to understand the source currency, target currency, rate direction and effective date. It should also retain the source of the exchange-rate information so that the conversion can be reviewed later.
For example, an illustrative conversion record could be represented as follows:
{
"source_currency": "AED",
"target_currency": "USD",
"exchange_rate": 0.2723,
"rate_direction": "USD_per_AED",
"effective_date": "2026-11-01",
"source": "illustrative_assumption"
}
The direction of the exchange rate matters because a value expressed as USD per AED cannot be applied in the same way as a value expressed as AED per USD. A calculation engine should not rely on the language model to interpret this distinction informally. The conversion rule should be explicit and validated before the calculation is executed.
Currency identification should also use standardised currency codes rather than relying exclusively on symbols. ISO 4217 provides internationally recognised three-letter identifiers, including USD, AED and GBP. This reduces ambiguity when financial information is transferred between systems or combined into consolidated reporting.
Most importantly, the converted value should remain traceable to the original amount and the exchange-rate assumption used. A converted figure is a derived financial value, not an independent market observation. Preserving that distinction is essential when comparing projects or reviewing financial assumptions prepared at different dates.
Time basis creates a different category of financial error
Time is often more difficult to validate than units or currency because financial models contain several different types of time-dependent information. Rental income may be expressed monthly, operating expenses annually, construction expenditure quarterly and development revenue cumulatively. These values may all appear as ordinary numerical inputs, even though they represent fundamentally different financial concepts.
Consider a rental assumption of 100,000 AED. If the value represents monthly rent, the corresponding annual amount may be 1.2 million AED, assuming the rental income continues for twelve full months. If the same numerical value already represents annual rent, multiplying it by twelve would introduce a substantial error.
The problem is not that the arithmetic is complicated. It is that the calculation engine needs to know what the original number represents before deciding which operation is appropriate.
A reliable system should distinguish between periodic amounts, rates per period, cumulative amounts and one-off payments. It should also preserve the period to which each assumption applies. This becomes particularly important in development models where construction expenditure, revenue commencement and operating costs may begin at different points in the project lifecycle.
More advanced models introduce additional considerations, including day-count conventions, payment dates, fiscal-year boundaries and discounting periods. These details can influence cash flow timing and present-value calculations. Treating time basis as a simple label attached to a financial amount is therefore insufficient for more sophisticated modelling applications.
Excel workbook context must survive data extraction
Excel introduces another challenge because the meaning of a cell is often distributed across the workbook rather than contained within the cell itself. A numerical value may depend on its formula, number format, worksheet heading, named range or relationship with other cells.
Dates provide a useful example. Excel supports both the 1900 and 1904 date systems, which use different underlying serial values to represent calendar dates. The difference between the two systems is 1,462 days. An extraction process that interprets cell values without considering the workbook's date system can therefore misinterpret dates even when the underlying numerical data has been captured correctly.
Number formatting creates similar problems. A cell containing 0.15 might represent a percentage, a decimal assumption or an intermediate calculation. A value displayed with a currency symbol may still require additional information to establish the actual currency. Formula context is also important because a cell's value may be the result of several calculations rather than an original input.
For this reason, spreadsheet extraction should preserve more than the displayed values. The system should retain relevant formula information, cell references, number formats and surrounding labels wherever these contribute to the interpretation of the financial data.
This does not mean that every spreadsheet requires a complete reconstruction of its visual layout. It means that the extraction process must preserve enough workbook context to prevent the calculation layer from treating unrelated numerical values as interchangeable.
A validation layer between interpretation and calculation
One practical way to address these problems is to introduce a validation layer between spreadsheet extraction and calculation execution. The purpose of this layer is to ensure that the information identified by the AI agent is sufficiently complete and internally consistent before any financial calculations are performed.
The process begins with extraction, where the system captures values, formulas, cell references and relevant labels. The interpretation layer then classifies the information according to its financial meaning. A value might represent a rental rate, construction cost, percentage assumption, exchange rate or project date.
Once the inputs have been classified, the system attaches the required metadata and checks whether the proposed calculation is valid. If a conversion is necessary, it applies an explicit rule and records the transformation. Only after these checks have been completed should the calculation engine execute the relevant arithmetic.
The distinction between interpretation and calculation is important. A language model may help identify that a particular row probably represents an annual rental rate, but the calculation engine should not rely on that interpretation without validation. Deterministic calculation logic should enforce the required conditions and produce reproducible results.
The final stage is provenance preservation. The system should retain the original input, the metadata associated with it, any conversion applied and the resulting calculated value. This allows a reviewer to understand not only what the output is, but also how the system arrived at it.
What should happen when metadata is missing?
A validation system should not assume that every ambiguous input can be resolved automatically. Consider a spreadsheet containing a construction cost of 4,500 without any additional description. The value could represent AED per square metre, USD per square foot or a cost per unit. It could also be a total amount expressed in thousands or another reporting convention.
There is no reliable way to determine the correct interpretation from the number alone. An agent that selects one possibility without supporting evidence may produce a plausible result, but that does not make the calculation defensible.
A better approach is to classify the input according to whether it can be used directly, requires a documented conversion or must be blocked until additional information becomes available.
For example, an input with complete metadata and compatible units can be marked as valid. An input with sufficient information but incompatible measurement units can be marked as convertible, provided an approved conversion rule exists. An input with missing or ambiguous metadata should be blocked from calculations that depend on that information.
This approach creates a useful distinction between missing data and invalid data. Missing information may be resolved by obtaining clarification from the analyst or an authoritative source. Invalid information requires correction before the calculation can proceed.
It also introduces a more appropriate role for the AI agent. Instead of trying to produce an answer regardless of the quality of the available information, the agent becomes responsible for identifying uncertainty and communicating which assumptions need review.
The broader engineering lesson
Financial modelling agents operate across several different information layers. They interpret natural-language instructions, extract spreadsheet data, identify financial assumptions, apply measurement conventions and execute calculations that may depend on time and currency.
The difficulty lies in preserving the relationships between these layers. A numerical value extracted from a workbook is not necessarily ready for calculation, even when the extraction itself is accurate. Its meaning must be established, its metadata must be preserved and its compatibility with other inputs must be validated.
This is why explicit units, currencies and time bases should be treated as part of the data contract in financial AI applications. They should not be added as descriptive information only when a reporting output is prepared. They belong at the point where data enters the system and should remain attached throughout subsequent transformations.
For developers building financial modelling agents, the objective should not simply be to generate a financial number. It should be to generate a result whose assumptions, transformations and calculation logic can be independently examined.
A calculation is only as reliable as the meaning of the inputs used to produce it.
Top comments (0)