Discover how artificial intelligence in finance handles numbers. Learn about AI hallucinations, key use cases, and how finance teams stay in control.
Key takeaways
- Generative AI tools are probabilistic and predict numbers based on statistical patterns, rather than executing traditional mathematical formulas.
- AI is prone to numerical hallucinations, particularly with complex tables, and can state incorrect figures with absolute confidence.
- Hybrid setups use code execution to run calculations in sandboxes, which improves accuracy but can still suffer from data-extraction errors.
- UK regulations and professional bodies maintain that human professionals retain ultimate accountability for all financial reporting and filings.
The adoption of AI in finance refers to using technologies like machine learning and generative AI to automate complex workflows, flag errors, and prepare financial files. While these tools offer incredible speed, their ability to calculate numbers accurately is still highly debated. Modern finance teams are finding that while AI cannot replace human verification, it is an invaluable partner for manual tasks like workbook structuring and first-pass analysis.
How is AI in finance being used today?
Is AI ready to run your department? According to a joint survey by the Bank of England and the Financial Conduct Authority (November 2024), 75% of regulated UK financial firms actively use AI, with adoption expected to reach 85% by late 2027. This marks a massive jump from 58% in 2022.
Yet, there is a clear boundary. The same survey revealed that 55% of AI use cases in UK financial services involve some degree of automated decision-making, but only 2% are fully autonomous. AI is still treated as an assistant, not an authority.
While 73% of UK CFOs are optimistic that AI will materially improve business performance (Deloitte UK CFO Survey, July 2026), they are not handing over the controls. Research by the Financial Reporting Council (FRC) in July 2026 showed that while corporate preparers widely adopt generative AI for drafting narrative text, they strongly avoid utilising it for formal financial statements due to data quality, legal liabilities, and investor preference for human verification.
The core problem with AI in accounting and calculations
For years, the joke about an AI accountant or a large language model (LLM) was simple: it could write a strategy memo, explain IFRS rules in a calm voice, then get a basic calculation wrong with the confidence of a broken workbook.
The annoying truth is that "AI calculating" is not the same as a calculator calculating. Standard bookkeeping systems are deterministic, requiring 100% mathematical consistency across every journal entry and balance sheet.
In contrast, LLMs are probabilistic models that predict what should come next based on patterns they have learned. They do not possess an innate mathematical engine. Instead, they break down numbers into character fragments (known as tokenization), treating digits as text-completion tasks. When you ask an LLM to add two numbers, it will probably answer correctly because it has seen that pattern many times before, not because it ran the calculation. On the wrong day, with the wrong prompt, it can look you in the eye, answer incorrectly, and act like the calculation tied out perfectly.
Why AI hallucinations happen with numbers
When an AI model generates highly plausible but mathematically incorrect information, it is called an AI hallucination.
Pure mathematical calculation is the easiest task for general LLMs to fail. Specialised benchmarks, such as Omni-MATH, show that raw text-based models struggle with complex maths, showing an average accuracy of just 0.38 out of 1. This results in numerical hallucinations in up to 41% of finance-specific queries involving tabular data.
What makes this particularly dangerous for finance teams is how these errors are presented. Research from MIT in 2025 revealed that AI models are 34% more likely to use highly confident, authoritative vocabulary (such as "definitely" or "certainly") when generating incorrect numbers compared to when they are correct.
This is why throwing raw, unstructured financial data at a probabilistic LLM is inherently high-risk. UK and international firms encountering AI-related errors incurred an average financial loss of $4.4 million per affected organisation (EY Global AI Sentiment Survey, April 2026).
Finance tasks AI can support without owning the calculation
So no, the answer is not "never use AI for finance data." That would be lazy. The secret to implementing AI in accounting safely is knowing where it shines and where it breaks. Instead of expecting the AI to act as a final calculator, finance departments are using AI as a structure engine.
(1) First-pass analysis and exception triage
AI excels at looking for anomalies and explaining variances in unstructured text. Instead of doing the core arithmetic, you can ask AI tools to:
- Identify outlier transactions in a massive general ledger export
- Draft explanations for variance reports by reading internal commentaries
- Categorise invoices or sort transactions for pre-accounting
- Write first drafts of complex Excel formulas or code scripts
- Turn a messy CSV export into something a human can actually inspect
Using AI for these repetitive, high-volume tasks allows finance professionals to free up time for strategic analysis.
(2) Preparing workbooks vs final validation
The safer route is usually this: let AI prepare the workbook, let Excel do the calculation, and let finance validate the logic.
Microsoft explicitly advised financial professionals against using its Copilot function in Excel for direct numerical calculations, instead recommending it for semantic or drafting tasks, and advising users to rely on native Excel formulas for deterministic calculations.
Rules for using AI safely in number-driven finance work
To successfully navigate how to use AI in finance, departments must establish clear rules of engagement.
First, never treat AI as the final number of record without a review. Where possible, take the safer route by asking AI to prepare a workbook rather than return a final figure. Then open the file and check the row count, source data, formulas, exclusions, and signs.
To bypass text-based calculation errors, the financial sector is rapidly transitioning to hybrid setups that decouple text synthesis from calculation. These frameworks, known as tool-integrated reasoning or Code Interpreter, prompt the LLM to write deterministic code, such as Python, to execute calculations in a secure sandbox and return the exact execution output. On mathematical benchmarks, this code execution setup boosts calculation accuracy from around 50% to 99.5%.
But even with code execution, you are not off the hook. The model can still fail due to ecological errors, where the maths in the generated script is perfectly correct, but the output is false because the AI applied the formula to incorrect variables, skipped rows, misaligned adjacent columns, or extracted the wrong cells from a spreadsheet. If the AI finds 50 out of 100 relevant transactions and sums them perfectly, the calculation is right. The finance answer is still not. For critical checks, run the same calculation twice in separate clean chat windows and compare the results. Matching outputs can increase confidence, but they still do not replace careful human review.
In general, you can run a simple trust test for AI-prepared documents:
- Did it use code execution to calculate the numbers?
- Can it show you the exact source rows it used?
- Are the formulas visible and auditable in the spreadsheet?
- Are all assumptions stated clearly, and can the AI explain them when questioned?
- Can you rerun the query in a new chat and get a similar answer?
Finance professionals must also maintain ultimate accountability. UK professional bodies, including the ICAEW, CIOT, and ATT, published updated guidance under the Professional Conduct in Relation to Taxation (PCRT) framework in January 2026. This guidance clarifies that practitioners must actively challenge AI outputs and maintain 100% personal responsibility for all client filings.
The UK’s FCA and the Bank of England’s Joint AI Framework also focus heavily on the systemic risks of third-party AI supplier concentration and establish that AI oversight must align with the UK’s Senior Managers and Certification Regime (SM&CR). Simply put: if the AI makes an error, the responsibility stays entirely with the human manager.
What we learned teaching AI to process finance data correctly
We ran into this ourselves while building our SaaS and AI Spend Surface skill. What looked like a straightforward prompt became much harder inside a chat window: getting AI to process every row, apply the same logic consistently, and return an output structured enough for finance to review required repeated testing, scripts, and clear calculation rules. We turned that work into a finished skill. If you want to read more about what we learned while building it and try the skill yourself in your own approved AI environment, you can find the full article and skill here.
A balanced path forward
While the rise of artificial intelligence in finance represents a massive shift in corporate operations, the ultimate responsibility for data accuracy remains human. By combining the speed of AI for preparation and first-pass analysis with the deterministic precision of spreadsheet formulas and senior finance validation, departments can successfully capture the benefits of automation without compromising financial integrity.