Before connecting a business app to ChatGPT, Claude, Gemini or Copilot, check four things: what data the connection can read, what actions it can take on your behalf, whose account it runs under, and what happens when the AI gets something wrong. A connection to your drive, CRM or accounting system turns an assistant that writes text into one that can see company records and sometimes change them. That is useful, and it deserves the same review you would give a new employee’s system access.

What changed when AI assistants started connecting to apps?

For the first few years, business use of ChatGPT meant copying text in and copying answers out. Now the major assistants connect directly to other software.

In October 2025 OpenAI introduced apps in ChatGPT along with an Apps SDK for developers. Services such as Canva, Figma, Booking.com and Spotify appeared inside conversations with their own interactive interfaces, and ChatGPT can suggest an app when it seems relevant. This followed an earlier attempt, the GPT Store, which asked users to leave their conversation and find a separate custom GPT. Putting apps inside the conversation is a much more natural design.

Video: OpenAI DevDay 2025: Opening Keynote with Sam Altman from OpenAI.

The Apps SDK is built on the Model Context Protocol (MCP), an open standard first published by Anthropic for connecting AI models to data sources and tools. MCP is now supported across much of the industry, including Claude, ChatGPT and a growing number of developer tools. The practical effect is that one connector can often work with several assistants.

Which Japanese business tools can connect to AI?

Several tools common in Japanese SMEs now have official ways to connect:

  • kintone. Cybozu publishes an official kintone MCP server that lets an MCP-compatible assistant read app settings and read, add, update or delete records.
  • freee. freee publishes an official freee MCP server covering its accounting, HR, invoicing and other APIs.
  • Google Workspace and Microsoft 365. Gemini and Copilot work inside their own suites, and ChatGPT and Claude offer connectors to Google Drive, Gmail, Outlook and SharePoint on many plans.
  • Automation platforms. Zapier and Make can sit between an assistant and tools that lack a direct connector. See Zapier vs Make for Japanese businesses.

Read that list carefully. An AI that can delete kintone records or create freee transactions is a different risk from one that drafts emails.

What should you check before connecting an app?

What can it read?

Most connectors act with the permissions of the account that authorized them. If the person who connects Google Drive can see every folder, including HR and payroll, the assistant can too. Before connecting, check what that account can access and whether it should. This is often the moment a company discovers its shared drive permissions were never set up properly.

What can it change?

Separate read access from write access. Reading records to answer questions is relatively low risk. Creating, updating or deleting records, sending emails or posting messages is not. Where a connector allows it, start read-only. Where the assistant can act, check that it asks for confirmation before each action, and test what happens when you decline.

Whose account is it?

A connection set up on an employee’s personal ChatGPT or Claude account belongs to that employee. When they leave, you may not be able to see or revoke it. Connect business tools only from company-managed business or enterprise plans, where an administrator controls which connectors are allowed and can remove access. The same rule applies to API keys: issue them from a company account and record who holds them.

Where does the data go?

When the assistant reads a document to answer a question, that content is processed by the AI provider and, for third-party apps, possibly by the app developer. OpenAI tells developers to collect only the minimum data they need, but you should still read what each app receives. If personal information is involved, Japan’s privacy law (APPI) applies, and the Japan SaaS compliance guide covers the checks for cross-border processing.

What happens when it is wrong?

An assistant will sometimes pick the wrong record, misread a field or misunderstand an instruction. Decide in advance which actions need human approval, whether there is a log of what the assistant did, and how you would reverse a bad change. For accounting data, anything the AI creates should be reviewed before it reaches your 税理士. I walk through how to audit an automation that runs unattended in what 851 unattended bot runs taught me about automation.

Who chooses the provider?

When ChatGPT suggests an app, it chooses which service to show. In Japan that might mean choosing between delivery or booking platforms, and how assistants rank competing services is not transparent. For personal tasks that is a convenience question. For business purchasing, keep the choice of supplier with a person.

How should a small business roll this out?

Start with one read-only connection. Connecting your company drive or knowledge base so the assistant can answer questions from real documents is the most common useful first step.

Clean the source first. An assistant searching a drive full of outdated price lists will quote outdated prices. Archive old versions and mark official documents.

Write a short connection policy. Which apps may be connected, from which accounts, with what permissions, and who approves new ones. One page is enough.

Add write actions slowly. When a read-only setup is trusted, consider actions such as creating draft records or tasks, with approval required.

Review access every quarter. Remove connectors nobody uses and check that permissions still match people’s roles.

This is the same discipline good companies already apply to shared drives and SaaS accounts. AI connections simply raise the cost of skipping it. I make the broader case in why AI adoption stalls in Japanese SMEs, and the Japan SME tech stack guide covers the tools underneath.

If you want AI connected to your business tools safely, with permissions reviewed and a rollout your team can follow, I can plan and implement it with you. If you are not sure your accounts and data are ready, start with a Diagnostics review.

Further reading: what ChatGPT is actually good for in small business work · how to evaluate an AI vendor before you buy · what 851 unattended bot runs taught me about automation · Japan’s copyright rules for businesses using AI