Which AI chatbot is best for accountants?

There isn't one answer. All four handle drafting, explanation and spreadsheet help. For client data, choose the plan before the brand: ChatGPT Business or Enterprise, Claude Team or Enterprise, Copilot Chat with a work account, and Gemini in a qualifying Workspace edition each say they don't train on your content by default. Pick the one that fits the software your firm already uses.

Can I use Claude for accounting?

Yes, for the same tasks as other assistants: drafting, explaining standards, summarizing documents you're allowed to share and building spreadsheets. For client data, Anthropic says its commercial products (Team, Enterprise and the API) aren't used for training by default, while personal plans may be unless you opt out. Several accounting tools, including Uku, Financial Cents, Numeric and Planful, document connectors that let Claude read their data, and Aiwyn offers an official Claude connector in an early phase that uses anonymized data.

Is it safe to put client data into ChatGPT?

It depends on the plan and your obligations. OpenAI says it doesn't train on ChatGPT Business, Enterprise or API data by default, while data sharing is on by default for personal Free, Plus and Pro accounts unless switched off. Tax preparers also face federal rules on disclosing tax return information. Consider barring client data from personal accounts entirely.

Does Microsoft Copilot use my data to train its models?

For work and school accounts, Microsoft says Copilot Chat's prompts and responses aren't used to train foundation models under enterprise data protection. For the current personal Copilot app, Microsoft says prompts, responses and file contents aren't used to train foundation models.

Do humans read Gemini chats?

For the personal Gemini app, Google says a subset of chats is reviewed by human reviewers and warns not to enter confidential data when Keep Activity is on. For Workspace with Gemini licensees, Google says submissions aren't used to train models and are never reviewed by humans.

What is an MCP connector?

MCP (Model Context Protocol) is a standard way for an AI assistant to read from, and sometimes act in, another application. Accounting tools such as Numeric, Rillet, Validis and Uku publish MCP servers so an assistant can read their data, in some cases under the signed-in user's own permissions.

AI chatbots for accounting: what ChatGPT, Claude, Copilot and Gemini do with client data

The four big general-purpose assistants are all useful for drafting, explaining and spreadsheet work. For an accounting firm, the deciding question is usually what happens to client data you type or upload, and the answer depends far more on the plan than on the brand. This guide sets out each provider's own terms for personal and business plans, then lists the accounting tools that run on these models or connect to them.

Guide. Updated October 15, 2026. By AI Accounting Vendors. 9-minute read.

Key points

Plan beats brand: OpenAI, Anthropic, Microsoft and Google each say content on their business plans isn't used to train their models by default or without permission. Personal plans differ by provider and by setting.
Personal plans: ChatGPT and Claude personal plans can use chats for training unless you turn it off. Gemini's personal app is governed by the Keep Activity setting; when it's on, a subset of chats is read by human reviewers.
Retention differs: Even with training off, providers keep chats for a period: for example up to 30 days for deleted ChatGPT and Claude chats, and 72 hours for Gemini chats with Keep Activity off.
Connectors carry data: Several tools in the directory connect their data to these assistants. Once connected, the assistant plan's data terms apply to what it reads.

The four assistants compared on data terms

Training on your content, by plan type, per each provider's documentation (read October 15, 2026)
Assistant	Personal plans	Business plans	Retention notes
ChatGPT (OpenAI)	Free, Plus and Pro on a personal workspace: OpenAI says data sharing is enabled by default. Turn off "Improve the model for everyone" under Settings, Data Controls (OpenAI (opens in a new tab)).	Business, Enterprise, Edu, Healthcare, Teachers and the API: "We do not train our models on your data by default" (OpenAI (opens in a new tab)).	Deleted chats are scheduled for permanent deletion within 30 days, unless kept longer for security or legal reasons (OpenAI (opens in a new tab)).
Claude (Anthropic)	Free, Pro and Max: Anthropic may use inputs and outputs for training "unless you opt out through your account settings" (privacy policy (opens in a new tab)).	Team, Enterprise, the API and Claude Gov: not used for training by default (Anthropic (opens in a new tab)).	Personal: deleted chats are removed from back-end storage within 30 days; if you allow training, de-identified data may be kept up to 5 years in training pipelines (Anthropic (opens in a new tab)). API: deleted within 30 days by default.
Microsoft Copilot	Current Copilot app with a personal Microsoft account: "Prompts, responses, and your file contents ... aren't used to train foundation models" (Microsoft (opens in a new tab)).	Copilot Chat with a work or school (Entra) account: enterprise data protection applies, and prompts and responses aren't used to train foundation models (Microsoft (opens in a new tab)).	Look for the green shield in Copilot Chat, which Microsoft says shows enterprise data protection is applied.
Gemini (Google)	Gemini app with a personal Google account: Google uses activity to improve its services, including training generative AI models, controlled by the Keep Activity setting. A subset of chats is reviewed by humans (Google (opens in a new tab)).	Qualifying Google Workspace editions: content isn't human-reviewed or used for model training outside your domain without permission (Google (opens in a new tab)).	Personal: chats with Keep Activity off are kept 72 hours; reviewed chats up to three years (Google (opens in a new tab)).

Two caveats apply to all four. Providers can change these terms, so check the linked page before you rely on it. And a business plan only protects data that goes through the business plan: a staff member pasting a client file into a personal account on their phone is outside it.

ChatGPT

OpenAI's help article on model training applies to "ChatGPT Plus, ChatGPT Pro or ChatGPT Free plan on a personal workspace," where it says "data sharing is enabled for you by default." The opt-out is under your profile: Settings, Data Controls, then switch off "Improve the model for everyone." OpenAI says that once you opt out, new conversations won't be used to train its models. Temporary chats aren't used to improve OpenAI models while they remain temporary, though OpenAI may keep a copy for up to 30 days for safety (Temporary Chat FAQ (opens in a new tab)).

For ChatGPT Business, Enterprise, Edu, Healthcare, Teachers and the API Platform, OpenAI's enterprise privacy page (opens in a new tab) says it doesn't train on your data by default, that you own your inputs and outputs where allowed by law, and that data is encrypted at rest (AES-256) and in transit. Retention controls are listed for Enterprise, Healthcare and Edu.

More on uses and limits in ChatGPT for accounting.

Claude

Anthropic changed its consumer terms on August 28, 2025, for Claude Free, Pro and Max (announcement (opens in a new tab)). It trains new models on data from those accounts when the model-training setting is on, and keeps that data for up to five years; users who don't allow training stay on 30-day retention. The setting can be changed at any time in Privacy Settings. The current privacy policy (opens in a new tab) says Anthropic may use inputs and outputs for training "unless you opt out through your account settings." Anthropic says incognito chats aren't used to improve Claude even if model improvement is on (Anthropic (opens in a new tab)).

The consumer update doesn't apply to Claude for Work (Team and Enterprise), the API, Amazon Bedrock, Google Cloud Vertex, Claude Gov or Claude for Education. For commercial products, Anthropic says it won't use your inputs or outputs for training by default; feedback or bug reports you explicitly submit are the exception (Anthropic (opens in a new tab)). API inputs and outputs are deleted from Anthropic's back end within 30 days unless otherwise agreed (Anthropic (opens in a new tab)).

Several accounting tools offer official connectors for Claude; see Accounting tools that connect to these assistants (opens in a new tab) below.

Microsoft Copilot and Copilot Chat

Microsoft has two products with similar names. Microsoft Copilot is for personal use with a personal Microsoft account. Microsoft Copilot Chat is for work and school (Entra) accounts and, per Microsoft's Copilot FAQ (opens in a new tab), offers enterprise data protection: prompts and responses aren't used to train foundation models. Microsoft says enterprise data protection in Copilot Chat comes at no extra cost (privacy and protections (opens in a new tab)). Without a Microsoft 365 Copilot license, Copilot Chat can't access the user's shared enterprise data, individual data, or external data indexed via Microsoft Graph connectors.

For the current personal Copilot app, Microsoft says prompts, responses and file contents aren't used to train foundation models (Microsoft (opens in a new tab)). Microsoft keeps a separate article for the older version of the app (opens in a new tab), which describes opt-out toggles named "Training on conversation activity" and "Training on voice conversations."

For Copilot inside Excel, see AI for Excel.

Gemini

For the Gemini app on a personal Google account, Google's Gemini Apps Privacy Hub (opens in a new tab) says activity is used "to provide, develop, and improve its services (including training generative AI models)," controlled by the Keep Activity setting. It also says "a subset of chats are reviewed by human reviewers" and warns: "If that setting is on, don't enter data that's confidential or that you wouldn't want a reviewer to see or Google to use to improve its services." Temporary chats aren't used to train Google's models.

For Google Workspace, Google says your interactions with Gemini stay within your organization and content isn't used for model training outside your domain without permission (Privacy Hub (opens in a new tab)). For Workspace with Gemini licensees using the Gemini app, Google's business FAQ (opens in a new tab) says submissions aren't used to train models and are never reviewed by humans. The Workspace terms apply only with a qualifying edition, and Google notes that NotebookLM is now called Gemini Notebook.

Accounting tools that connect to these assistants

Several accounting tools in the directory let you bring their data into Claude, ChatGPT, Copilot or Gemini, usually through an MCP (Model Context Protocol) server or an official connector. That can be more useful than pasting exports, because the assistant reads current data, in some cases under the user's own permissions. It also means the assistant plan's data terms now cover client data, so connect only from business accounts.

Directory tools with an MCP server or AI-assistant connector, per vendor documentation
Tool	What its documentation says
AiwynPractice management	Official Claude connector (Phase 1, for exploration with anonymized data): Claude does the reasoning and Aiwyn's engine does the deterministic tax calculations
UkuPractice management	MCP server that lets Claude, ChatGPT or other MCP-capable assistants read firm data and take actions under the signed-in user's own permissions (Elite plan)
Financial CentsPractice management	MCP connector (Scale plan) lets outside assistants such as Claude, ChatGPT or Gemini query and act on firm data within user permissions
KickBookkeeping	MCP and CLI access so AI tools such as ChatGPT and Claude can read and update books; scheduled agent tasks across clients in early access
ZeniBookkeeping	Zeni MCP, which connects a company's books to outside AI assistants such as Claude and Codex
ValidisAudit	Validis MCP: a read-only, audit-logged connector that brings standardized ledger data into Claude, ChatGPT, Copilot, Excel and other MCP-compatible tools
FieldguideAudit	Fieldguide MCP connects live engagement data to outside AI assistants
NumericClose & reconciliation	MCP server that lets Claude, ChatGPT and other AI tools query Numeric data with the user's permissions
DoubleClose & reconciliation	An MCP server that connects a Double account to outside AI assistants
RilletClose & reconciliation	MCP access so outside tools such as Claude, Salesforce and Snowflake can read from and write to the ledger
CubeFP&A	MCP server that lets outside AI assistants query Cube data within user permissions
DatarailsFP&A	FinanceOS AI Connector, which lets outside AI tools work with governed Datarails data
PigmentFP&A	MCP server that gives outside AI tools such as ChatGPT and Claude governed access to Pigment data
PlanfulFP&A	Read-only MCP server lets users query Planful data from Claude under existing access controls

Accounting tools built on these models

Several accounting tools in the directory run on the same models. What matters is the contract between the tool and its model provider, which several vendors describe:

What directory vendors say about their AI model providers
Tool	What the vendor says
Canopy	Never uses firm data to train or test AI models unless the firm explicitly opts in, and its third-party providers (OpenAI, Anthropic) don't train their general models on customer data.
Double	Doesn't train models on client data, and its model providers (Anthropic, OpenAI and Google) are prohibited from training on Double's AI usage.
Juno	Won't use firm or client data to train its AI models; its AI partners (OpenAI, Anthropic, Google) don't store or train on client data.
Karbon	Doesn't use a firm's data to train AI models; its Azure OpenAI and OpenAI providers have represented they won't train on customer data.
Fieldguide	Customer data isn't used to train AI models; its providers (Microsoft Azure, AWS, OpenAI) don't train on submitted data and keep prompts at most 30 days for abuse monitoring.
TaxDome	AI runs on OpenAI and Google: data sent to OpenAI isn't kept after processing, while Google may hold some processed data for up to 30 days for abuse monitoring.
Docyt	Customer data is never used to train AI models; its AI setup uses Google Cloud Vertex AI with Gemini models and zero retention.
Vena	Uses Microsoft Azure OpenAI models.
Drivetrain	Uses its own models plus models from Anthropic and OpenAI.

Each listing links the vendor's own page for these statements. To ask a vendor the same questions, use the AI vendor questionnaire.

Ground rules for using chatbots in a firm

Before staff use a general AI assistant on client work

Provide business accounts (ChatGPT Business or Enterprise, Claude Team or Enterprise, Copilot Chat with Entra sign-in, or Gemini in a qualifying Workspace edition) and say personal accounts aren't for client data.
If personal accounts are allowed for anything, require the training setting to be off and say what must never be entered.
Keep taxpayer identifiers, bank details and full returns out of chat unless your policy and the plan's terms allow it.
Tax preparers: check the federal rules on disclosing and using tax return information (26 CFR 301.7216) before client return data goes into any third-party tool.
Firms covered by the FTC Safeguards Rule: treat AI providers as service providers you oversee.
Connect accounting tools to assistants only from business accounts, and review the permissions the connector uses.
Have a person check figures, citations and tax positions before anything reaches a client.

This is not legal advice. Your engagement letters, state rules and professional standards may set stricter limits.

AI chatbots for accounting: questions

Which AI chatbot is best for accountants?: There isn't one answer. All four handle drafting, explanation and spreadsheet help. For client data, choose the plan before the brand: ChatGPT Business or Enterprise, Claude Team or Enterprise, Copilot Chat with a work account, and Gemini in a qualifying Workspace edition each say they don't train on your content by default. Pick the one that fits the software your firm already uses.
Can I use Claude for accounting?: Yes, for the same tasks as other assistants: drafting, explaining standards, summarizing documents you're allowed to share and building spreadsheets. For client data, Anthropic says its commercial products (Team, Enterprise and the API) aren't used for training by default, while personal plans may be unless you opt out. Several accounting tools, including Uku, Financial Cents, Numeric and Planful, document connectors that let Claude read their data, and Aiwyn offers an official Claude connector in an early phase that uses anonymized data.
Is it safe to put client data into ChatGPT?: It depends on the plan and your obligations. OpenAI says it doesn't train on ChatGPT Business, Enterprise or API data by default, while data sharing is on by default for personal Free, Plus and Pro accounts unless switched off. Tax preparers also face federal rules on disclosing tax return information. Consider barring client data from personal accounts entirely.
Does Microsoft Copilot use my data to train its models?: For work and school accounts, Microsoft says Copilot Chat's prompts and responses aren't used to train foundation models under enterprise data protection. For the current personal Copilot app, Microsoft says prompts, responses and file contents aren't used to train foundation models.
Do humans read Gemini chats?: For the personal Gemini app, Google says a subset of chats is reviewed by human reviewers and warns not to enter confidential data when Keep Activity is on. For Workspace with Gemini licensees, Google says submissions aren't used to train models and are never reviewed by humans.
What is an MCP connector?: MCP (Model Context Protocol) is a standard way for an AI assistant to read from, and sometimes act in, another application. Accounting tools such as Numeric, Rillet, Validis and Uku publish MCP servers so an assistant can read their data, in some cases under the signed-in user's own permissions.

Sources

Official documents read on October 15, 2026. Vendor facts are sourced on each listing.