What does this guide cover, and what does it not?
This guide is for finance team members using AI in their own work: FP&A analysts, finance business partners, management accountants and finance managers preparing reports, commentary and analysis. It is about drafting, explaining and checking, with tools such as ChatGPT, Claude, Gemini or Copilot in Excel.
It is not about AI in banking or financial services as an industry, such as fraud detection, credit scoring, algorithmic trading or robo-advice. Those are products built and regulated as systems, and they raise different questions. Accounting and bookkeeping work has its own guide, linked below. Nothing here is investment, tax or accounting advice.
Where does AI help finance teams most?
AI helps most where the work is explaining numbers you already have, rather than producing new ones. It is good at language and structure, and unreliable at arithmetic and at knowing why a number moved.
Commentary drafts
Turning a checked variance table into a first draft of management commentary, which you then correct and complete with the causes you have confirmed.
Formula help
Explaining an inherited formula in plain English, suggesting a cleaner structure or writing a lookup, then testing it against known values.
Document summaries
Condensing a long contract, lease or board paper into the terms that matter for a model, with clause references kept.
Review questions
Listing the questions a sceptical CFO would ask about a forecast, so you can prepare answers.
Explaining to non-finance colleagues
Rewriting a budget explanation for a department head who does not read P&L statements.
What does checked AI variance commentary look like?
Marlow Outdoor is a fictional retailer of camping equipment. Its October management pack shows revenue budgeted at 420 (in thousands) and actual at 388, and marketing spend budgeted at 60 and actual at 71. The analyst pasted the table into an assistant and asked for 'commentary on October performance'.
The draft said revenue was down 8.2 percent 'due to softer seasonal demand', that marketing overspend 'reflects investment in the autumn campaign, which should drive a stronger November', and called the overall position 'broadly on track'. Recalculating gave a revenue shortfall of 32, which is 7.6 percent of budget, not 8.2. Marketing was 11 over budget, 18.3 percent. The table said nothing about seasonality or the autumn campaign, and 'should drive a stronger November' is a forecast nobody had made.
The analyst rewrote the prompt: 'Using only this table, state each variance in absolute and percentage terms. Do not give causes or forecasts. List the questions I need to answer before I can explain each variance.' The new draft stated the numbers correctly and asked useful questions: did the marketing overspend relate to a timing difference in invoices, and was the revenue shortfall concentrated in one channel? The analyst answered those from the ledger and the ecommerce report, then added the confirmed causes in their own words.
The habit to take from this is splitting the job in two. Let the tool structure and phrase. Do the arithmetic and the explanation of cause yourself, from the source.
How do you use AI with spreadsheets safely?
Treat AI-written formulas like formulas from a new colleague: useful, and tested before trusted. Microsoft describes Copilot in Excel as able to edit worksheets, generate formulas, build charts and PivotTables, and work in edit, plan or chat modes, and says you remain in control of what it changes. Plan mode, which proposes an approach before editing, is the safer starting point for anything important.
Test every new formula on a few rows where you already know the answer, including an edge case such as a blank cell, a negative value or a date at the month end. Look for hard-coded numbers that should be references, ranges that stop short of the data, and lookups that silently return the first match. If a general assistant suggests a formula, paste it into a copy of the workbook, not the live model.
Want to see how this connects to a role?
Explore the planned team trial ↗How should you check AI financial analysis?
Check it as if a confident junior had prepared it without access to the ledger. NIST's Generative AI Profile describes confabulation as the production of confidently stated but erroneous or false content, and confident financial prose is exactly where that is easy to miss.
Recompute
Recalculate every total, variance and percentage from the source, in the spreadsheet rather than in your head.
Units and periods
Confirm thousands versus units, month versus year to date, and which budget version is being compared.
Sign conventions
Check that favourable and adverse are the right way round, especially for cost lines.
Cause versus change
Remove any explanation of why a number moved unless you have confirmed it from another source.
Forecast language
Strip out 'should', 'will' and 'expected to' unless the forecast exists and has been approved.
Trail
Keep the source table, your recalculation and the final text together so a reviewer can follow them.
Is there a ChatGPT for finance, and which AI is best for finance analysis?
There is no single finance version that is clearly best. Finance teams usually choose among three kinds of tool: general assistants with file analysis features, AI built into the spreadsheet such as Copilot in Excel, and AI features inside finance and planning software. The right choice depends on where your data already lives and what your company has approved, more than on which model tops a ranking this month.
Whichever you use, check the plan's terms on how uploaded files are stored and whether content is used to improve models. Uploading a management pack to a personal account is a different act from analysing it inside an approved business workspace.
What finance information should stay out of AI tools?
Keep out anything your company has not approved for the tool in question, and be strictest with unpublished results, payroll and individual salary data, deal and acquisition work, bank details and anything price-sensitive. For listed companies, inside information is governed by rules that do not change because a tool is convenient.
A practical habit is to practise and build structure on fictional or anonymised data, then apply the finished formula or template to the real file inside approved systems.
When should finance teams not use AI?
Do not use AI output as the source of a reported number, a forecast presented as fact or evidence for an audit. Do not use it to decide an accounting treatment, a tax position or an investment. It can help you draft the question you take to a qualified adviser, which is often the most useful thing it does.
It is also not worth it for small, familiar calculations. Typing a variance into a chat and checking the answer takes longer than writing the formula.
How can a finance team build this skill?
Use the Marlow Outdoor exercise, or one like it, as a team practice task: the same table, an AI draft, and a short list from each person of what they corrected. The corrections show who recalculates and who trusts the prose.
Tosheo.ai is being built with a planned finance path that includes tasks like this one, feedback against a visible rubric and a manager view of skill progress without chat logs. It is pre-build, so this is planned scope, not a live service, and the lessons are not professional finance training or advice.