Claude, ChatGPT or Gemini for office work: compare them on your actual tasks

Define an acceptable result before testing

Give each assistant the same fictional inputs and output request. For a report, supply three documents and require an evidence matrix. For a spreadsheet, use data with known answers. For slides, define the audience, message and editable format.

Write the criteria before seeing outputs. Otherwise an attractive response can influence the rules used to judge it. A beautiful report should not pass if it invents a decision or misstates an amount.

Check capabilities in your actual account

Verify supported formats, file creation, necessary integrations and organizational settings. Consult the documentation for Claude files, ChatGPT and Gemini uploads. An announced feature does not prove that it is available in your account or accurate on your particular file.

Use a fictional dataset for exploratory testing. It can reproduce layout and calculation problems without exposing client records. Organizational restrictions on tools and data come before personal preferences.

Create tests that reveal meaningful differences

For documents, introduce a deliberate date conflict: an initial schedule says November 12, an amendment proposes November 19, and a meeting record says the new date is not approved. The correct answer should preserve uncertainty rather than silently choose a date.

For spreadsheets, use 12 units at €80, 8 at €120 and 20 at €20. Total revenue is €2,320 and the quantity-weighted average is €58. Add a subtotal row to test whether it is counted twice. Our Excel verification guide explains these traps.

For presentations, request five slides for an operational audience. Check that titles express an idea, figures come from the supplied inputs and the exported file remains editable. Test the exported artifact, not just a preview.

Score quality and record correction work

Use a simple scale for each criterion: 0 incorrect, 1 partial, 2 compliant. Evaluate fidelity, traceability, arithmetic, format and ease of revision. Define blocking errors separately: a high total score cannot compensate for a false critical commitment.

Measure preparation, generation and review time separately. Repeat tasks with the same initial instruction. Save the raw outputs, date, plan and model when displayed. This will not produce a perfectly controlled experiment, but it makes your conclusion easier to interpret.

Compare subscription cost, any additional usage charges and correction time. Do not turn a fast result on three examples into a general productivity promise.

Using only the supplied documents, produce a report with facts, precise sources, contradictions and open questions. Invent no amounts or decisions. Follow the requested format, explain calculations and identify file-reading limits. I will evaluate the output using criteria established before the trials.

Choose a workflow and revisit it

Your conclusion may favor one product for reports and another for presentations, if your organization permits both. It may also be that no tool meets the criteria for a particular task. Retain a manual workflow when verification demands more effort than direct preparation.

Reevaluate after a material change in the product, inputs or business need. Keeping your test set lets you check the change against evidence rather than slogans. Download the assistant comparison grid.

Frequently asked questions

Which assistant is best? The one that meets the criteria for your task under your constraints, after verification. This guide does not claim a winner for all use cases.

Does a more expensive plan guarantee better answers? No. It may change available capabilities or quotas, but the output still needs evaluation.

Try the supplied fictional document set

Download the comparison pack with three documents and an answer reference. The source documents are in French: an initial plan, revised quote and pending decision. They are fictional and contain no customer data.

Ask for a table of facts, status and source. The reference retains a planned budget of EUR 1,200, a quote of EUR 980 excluding VAT and a total of EUR 1,176 including VAT. The proposed date of 19 November is not confirmed, and Laura's approval remains pending. corrige.json lists the reference evidence and critical errors.

Keep the same inputs and instruction for each assistant. Record the actual model version, tools, settings and rework. Use the comparison workbook for expected facts, correct facts, critical errors and cost. Inputs are blank: no model test or winner has been invented.

Start with one attempt, check it against the documents, then repeat on another case. Success on this small set does not establish general reliability. Try the evidence checking challenge before comparing tools.

Sources

• →Anthropic: upload files to Claude.
• →OpenAI: use ChatGPT.
• →Google: upload and analyze files with Gemini.