ChatGPT vs Claude vs Copilot vs Gemini vs Perplexity for accountants: how to choose

How this comparison was made

What matters more than the model

Vendors release new models every few months. In our view, for everyday accounting tasks, how you use a tool matters more than which leading model it runs. Three things decide more:

1. The plan and its data terms. Business plans from all five vendors say your data isn't used to train their models by default. Personal plans don't all say that. What your firm has approved, on which plan, decides what you may put in. See whether it's safe to put client data into ChatGPT.
2. Where your work lives. An assistant inside Excel and Outlook saves copying and pasting, and copying is where client data leaks and figures get mistyped.
3. How you check the output. Every one of these tools can produce a confident wrong answer. A firm that checks well gets more from an average tool than a firm that trusts a strong one.

The five tools at a glance

Prices change more often than anything else, so check each vendor's current plans: ChatGPT, Claude, Microsoft 365 Copilot, Google Workspace and Perplexity.

Each tool in brief

ChatGPT

OpenAI says it doesn't train its models on inputs and outputs from ChatGPT Business, Enterprise and Edu or its API by default (OpenAI). On personal plans, training is on unless you turn it off (OpenAI Help Center). OpenAI also offers ChatGPT for Excel and Google Sheets as an add-in (OpenAI Help Center). If staff already use ChatGPT at home, the risk is that they use the same personal account for work. That's the gap a firm policy needs to close.

Claude

Anthropic says it doesn't use data shared through its commercial products to train its models (Anthropic). Claude for Excel is available on its Pro, Max, Team and Enterprise plans and shares context with its Word, PowerPoint and Outlook add-ins. Anthropic's own documentation says it isn't recommended for audit-critical calculations without verification, and notes that the add-in doesn't inherit an organization's custom data retention settings (Anthropic). Check both points against your firm's policy.

Microsoft Copilot

As used by organizations, Microsoft Copilot and Copilot Chat are covered by Microsoft's Data Protection Addendum, and Microsoft says prompts and responses aren't used to train its foundation models (Microsoft). It works inside Excel, Outlook and Teams. The Copilot guide covers what it does in each, and what to check.

Gemini

Google says Workspace doesn't use customer data, including prompts, to train models without the customer's prior permission or instruction, and Gemini works in the side panel of Gmail, Docs, Sheets, Drive and Meet (Google). Those commitments are for Workspace accounts, not personal Google accounts.

Perplexity

Perplexity's core product is research: answers drawn from the web, shown with their sources. For its Enterprise plans, Perplexity says your organization's data is never used to train or fine-tune its models, and that its agreements with the model providers it uses prevent training on its data too (Perplexity). Sources shown next to an answer still need opening: see how to check AI's tax research.

Which fits your situation

• Your firm runs Microsoft 365 and IT manages it. Start with Copilot, because it's already inside the apps and under the agreement you have. Other assistants' Excel add-ins are an option if your IT team approves them.
• Your firm runs Google Workspace. Start with Gemini, for the same reasons.
• You need research with sources you can follow. Perplexity or any assistant with web search, plus the citation checks in the tax research guide. For tax and accounting authority, a research tool built for professionals may suit better (see below).
• You work mainly in spreadsheets. Look at the Excel options: Copilot, ChatGPT for Excel, Claude for Excel. Run the test below on one of your own (anonymized) workbooks.
• You're a solo practitioner or a very small firm. A single business plan from any of these vendors, used consistently and checked carefully, beats a mix of personal accounts.

Test them yourself in 30 minutes

Benchmarks published by vendors don't test accounting work. Your own test will tell you more. Use made-up data or data with every identifying detail removed, and give each tool exactly the same task and files.

1. Pick one real task you do every month, such as coding a bank export or explaining account movements.
2. Plant two problems in the practice file: a line nobody could code from its description, and a figure that stops the total tying.
3. Use the same prompt in each tool, with the rules from the prompts guide.
4. Score each answer on the checks below, then try one correction and see whether it fixes the problem without breaking anything else.

Run the same test again when a vendor releases a new model or your firm considers a change. It takes half an hour and tells you more than any ranking.

Specialist tax and accounting research tools

Several publishers of tax and accounting research now build AI into their products, answering from their own libraries of authority. They can be a better fit for research than a general assistant. They still make mistakes: a 2024 Stanford study found that AI research tools built for lawyers gave incorrect information more than 17% of the time (Stanford HAI). The same checks apply whichever tool you use.

Keeping this guide current

Plans, features and terms change often. We check every fact in this guide each quarter and update the date at the top. If you spot something out of date before then, tell us at tom@secondskillset.com.