AI in finance, accounting and analysis

Where things stand (as of 2026)

AI tools are genuinely useful for drafting commentary, summarising long filings, restructuring messy data, and speeding up first-pass triage of reconciliations or expense reports. Several accounting and finance platforms now have AI features built directly into the workflow - flagging anomalies, drafting narrative, suggesting categorisations.

What they are still not reliable at: arithmetic on anything non-trivial, especially across large or messy datasets, and producing citations or figures that are genuinely accurate rather than plausible-looking. Large language models predict likely text, not calculate; if a tool is doing genuine calculation, that's usually a separate function or plugin bolted on, not the model itself, and the boundary between the two isn't always visible to you. Firms have already had to correct AI-drafted reports where figures were subtly wrong but read fluently. This area moves fast - a tool's arithmetic reliability this year is not a guarantee for next year, so re-test rather than assume.

The mental model

Think of the AI as a fast, well-read junior who has read enormous amounts of financial writing and can produce fluent, structurally sound drafts - but who cannot actually run a calculator reliably, has no idea what your specific client's numbers should look like, and will not tell you when they're guessing unless you insist on it. Every number they write down needs to be traced back to a source you control.

Scenarios that work

Where it fails, and what it costs you

Arithmetic errors that look correct. Language models generate numbers by pattern, not calculation. A variance percentage or a running total can be wrong while still being internally "plausible" - which makes it harder to spot than an obviously nonsensical error. If that number ends up in a client report or regulatory filing, you own the error, not the tool.

Fabricated or misattributed citations. Ask for a source on an accounting standard, tax rule, or market statistic, and a model can produce a confident-sounding reference that doesn't exist or doesn't say what it claims. This is the same failure mode that has caused professionals in other fields to submit fabricated citations to courts - the finance equivalent is a client report or valuation memo built on a source that isn't real.

Model risk and silent drift. Firms that build workflows around a specific AI model can find behaviour changes when the underlying model is updated, without warning. A prompt that reliably produced good variance commentary last quarter might behave differently this quarter. Treat any AI-dependent workflow as something to periodically re-validate, the same way you'd revalidate a spreadsheet macro after a software update.

Confidentiality and client data. Pasting client financial data into a consumer-grade AI tool without a data processing agreement can breach client confidentiality obligations and, depending on jurisdiction, data protection law. Some tools use input data to further train models unless you've specifically opted out or are on an enterprise contract that excludes it.

No audit trail by default. Regulators and auditors expect to be able to see how a number was derived. "An AI drafted it" is not a control. If AI is part of your process, document what it was used for, what was checked, and by whom - the same discipline you'd apply to any other draft-then-review step.

Never the source of truth. The single most important rule: AI output is a draft or a triage aid. The general ledger, the reconciled bank statement, the audited financial statements - these remain the source of truth. If an AI-generated number and your system disagree, the system wins until proven otherwise, not the AI.

How to check the answer

1. Recalculate every number the AI produced using your own spreadsheet formulas or accounting system - do not accept a total, percentage, or ratio on faith.
2. Trace every cited figure or claim back to its original source document and page or line reference.
3. Cross-check totals: does the sum of the parts the AI listed actually equal the total it quoted?
4. Ask a colleague who didn't see the AI draft to sanity-check the conclusion independently.
5. Confirm no client-identifying or commercially sensitive data was entered into a tool without proper authorisation.
6. Record, briefly, what the AI was used for and what you checked - treat it like any other piece of workpaper documentation.

How to read the docs and find the truth

Go to the primary source, not a summary of it. For accounting standards, that means the IFRS Foundation or your local standard-setter (e.g. FASB in the US, the FRC in the UK) - not a blog post or an AI-generated explainer. For tax and regulatory questions, use your tax authority's own published guidance and, where relevant, the actual legislation.

For the AI tool itself, read the vendor's model card or system card and its data-handling terms, not just marketing copy. Look for: what happens to data you input (is it used for training?), what the tool's known limitations are (most vendors now disclose that arithmetic and citation accuracy are limited), and the date of the document - AI vendor documentation changes often, and a page you read three months ago may already be out of date. A genuine source will have a visible publication or last-updated date and will link back to primary regulatory text rather than paraphrasing it.

Your one-page playbook

ROLE: You are a [drafting assistant / triage assistant] for [task].
INPUT: Here is the data or document you should use: [paste or attach].
BOUNDARY: Only use the data provided. Do not use outside knowledge for figures.
Do not perform final calculations I will rely on without me checking your working.
FORMAT: [table / three paragraphs / bullet list], tone: [neutral and factual].
UNCERTAINTY: If you are not confident in a figure or source, say so explicitly
rather than estimating.
OUTPUT IS A DRAFT: This will be reviewed and verified against source data
before use. Do not present it as final.

Check yourself

1. If an AI tool produces a variance percentage, what is your process for confirming it's correct before it goes in a client-facing document?
2. What data-handling rule applies at your firm before you paste a client's financial data into any AI tool?
3. Name one thing that could go wrong if you treated an AI-drafted reconciliation as complete without independent review.
4. Where would you go to check whether an accounting standard has actually changed, rather than trusting an AI's explanation of it?