Iter Advisors

LLMs in finance: compare ChatGPT, Claude and Gemini

Cases and sources reviewedBy ·

The right LLM for a finance team is the one that passes a reproducible business test in an authorised environment. Compare answer quality, traceability of figures and correction time on the same documents and the same instructions.

Dashboards and financial data analysis

English version published on 2 October 2026. Sources rechecked on the same date.

Which criteria should you compare across assistants?

ChatGPT, Claude and Gemini differ in more than writing style. The functions available depend on the plan, the administrator settings and the connected tools. This framework is a selection method, not a benchmark carried out by Iter.

Framework to complete for each tested environment
CriterionTestEvidence to retain
CalculationRecalculate a table containing credit notes, duplicates and a zero budgetFormulas or code and results reconciled to the answer key
DocumentsFind five clauses in a document setFile name, page and exact passage
ContextExplain a variance whose cause is not suppliedDistinction between fact and hypothesis
IntegrationRead only authorised foldersVerification of permissions and possible actions
CostMeasure test and correction timeLicences, consumption and review time
Three environments: documented functions and test questions
EnvironmentDocumented starting pointQuestion to verify
ChatGPTFile analysis and code-based calculationsCan you verify operations, exclusions and output files?
Claude inside a business applicationDocument analysis in the Hebbia caseDoes the chosen application preserve references to supporting documents?
Gemini in Data StudioAssistance with questions, calculated fields and presentationsAre these features available in your plan and data context?

These examples involve different products and integrations. They do not rank model quality. Compare the configurations actually available to your team.

ChatGPT documents file analysis using calculation tools. Ask every other candidate environment to demonstrate the same task on authorised company data, without assuming the quality will be equivalent.

How do you run a useful comparative test?

  1. Prepare an authorised set of files: variance table, closing procedure and supporting documents.
  2. Write an answer key with the expected figures, sources and unanswerable questions.
  3. Give every tool the same instructions and keep its full answers.
  4. Log errors, fabricated references, omissions and minutes spent correcting.
  5. Repeat the sensitive cases: one correct answer does not prove reliability.

An accurate result without a verifiable reference is not enough for an audit note. Reject an environment if the test requires excessive access rights. The ChatGPT guide exercises offer a reproducible starting point.

What does the Hebbia case show about financial documents?

Anthropic describes how Hebbia uses Claude to analyse financial and legal documents. Claude runs inside a platform that organises documents and tasks; the application goes well beyond a standalone conversation.

For an SME, the lesson to test is document traceability. Ask for a supporting file and passage behind every statement. The cache-related time saving reported by the publisher must not be read as a finance team’s productivity gain. Read the cases and their scope.

Which decisions must remain subject to human approval?

Budget assumptions, provisions, payment approvals and investor communications commit the company. An assistant can prepare a proposal; keep a named owner and a control trail.

For financial due diligence, use the model to spot missing documents or prepare questions. A model that fails to flag a risk does not prove there is none. The final review also covers what was never supplied.

What should you check before connecting financial data?

Document who can read what, which actions are possible, where data is kept and how access can be revoked. Read OpenAI’s business-data commitments and Anthropic’s training policy for the plan you actually choose. Do not carry a condition over from one product to another.

For the pilot, prefer read-only access and a restricted folder. A Fractional CFO can define the business need with IT and the data owners. The roadmap links selection to the deployment stages.

Sébastien Doat

Your finance contact

Sébastien Doat

Co-founder, fractional CFO and CFO

Frequently asked questions

Is there a best LLM for every finance department?

No. The choice depends on your set of files, the tools available, access, the quality expected and the cost of verification. Compare environments on identical tasks.

Can an LLM perform calculations?

An assistant can use code, a spreadsheet or other calculation tools. Check the operations executed and the assumptions; a number generated in prose is not evidence of a calculation.

How do you measure financial-summary quality?

Compare against an answer key: accurate figures, retrievable sources, no invented causes, material points covered and correction time. Also test for information missing from the file set.

Scope your project with a CFO

Share your needs, tools and difficulties. We can clarify the scope, deliverables and controls for support.

Discuss my project

Explore our Fractional CFO service

Sources & references

  1. Data analysis with ChatGPT — OpenAI Help Center, source rechecked on 2 October 2026.
  2. Hebbia: financial-document analysis with Claude — Anthropic / Claude, source rechecked on 2 October 2026.
  3. Data Studio Pro features (formerly Looker Studio) — Google Cloud, source rechecked on 2 October 2026.
  4. Business data protection — OpenAI, source rechecked on 2 October 2026.
  5. Data use for model training — Anthropic Privacy Center, source rechecked on 2 October 2026.