How far should you let AI into your business?: AI agents with access to email and information

It is becoming increasingly common to use artificial intelligence tools that are directly linked to a company’s email, calendar, CRM or document management system.

This makes it possible to do things that, until recently, seemed hard to imagine: asking the AI to find a specific email amongst thousands, to draft a reply taking previous conversations into account, to consult documentation stored on the company’s servers or in the cloud, or to search for and cross-reference information from different sources.

And they can do this because we are giving them access, via our user account, to the very same systems and information that we ourselves already have access to. In practice, We’re letting an AI in with our keys: you can look up whatever we can look up.

That is precisely why making an off-the-cuff enquiry to ChatGPT is not the same as integrating it into the company’s systems.

Before doing so, it is worth going over a few points.

To understand this better, in this article we’re going to focus on a specific case: ChatGPT and its integration with a company’s Microsoft environment, such as Outlook, the calendar or SharePoint. We’ll look at the differences between its various versions, what guarantees OpenAI offers when it’s used in a business environment, and what happens when we grant it access to our systems. The same analysis can be applied to other AI tools.

1. Are we using a business tool?

Using a personal account on an AI tool is not the same as using a solution designed specifically for businesses.

For example, OpenAI states the following for ChatGPT Business: Data from the workspace is not used to train your models and that the information is encrypted both during transmission and whilst in storage.

It also offers specific controls for organisations and features a SOC 2 Type 2 independent audit, a security standard that verifies that a company has in place and effectively implements controls to protect information and properly manage aspects such as access, system availability and data confidentiality.

Therefore, before linking corporate data, the first question should be a simple one: which version of the tool are we using, and what conditions apply to our data?

It is not enough simply to know what the tool can do. We must also understand how it works when we connect it to our systems: what its terms and conditions say about company information, what guarantees the supplier offers, how it protects data, and what controls we will have over its access and use. This is a review that should be carried out before we grant it access to our systems, not afterwards.

2. Connecting Outlook or SharePoint does not necessarily mean opening up the entire company to AI

Business integrations work on the basis of permissions.

In the case of ChatGPT connected to Microsoft 365, access is granted via Microsoft’s authorisation mechanisms. In simple terms, ChatGPT does not receive a master key from the company: a specific connection is authorised, and Microsoft controls which user connects and what information they are permitted to access. Technically, this is facilitated by services such as Microsoft Graph and, in certain cases, may also require authorisation from the organisation’s administrator via Microsoft Sign In.

This enables the company to maintain a level of control over integration: which applications can be connected, who can do so, and what permissions are granted.

Furthermore, ChatGPT respects the permissions that already exist in Microsoft systems.

For example, if a user cannot access a particular SharePoint folder, connecting ChatGPT should not grant them access to that folder.

This is important, but it also highlights another problem: If internal permissions are configured incorrectly, the AI may exacerbate the problem.

Before integrating an AI tool, therefore, it is a good time to check who has access to each item.

3. Can you only view information, or can you also take action?

Allowing an AI to search for an email is not the same as allowing it to send one.

Nor is it the same thing to view a document as it is to edit, move or delete it.

We need to approach this from two angles.

The first is technical. The tool must be configured so that each user and each integration has only the permissions required to carry out their work. We should not allow an action simply because it is technically possible. It is, moreover, the principle of minimising access and privileges that underpins the European data protection and cybersecurity regulations: to restrict access and capabilities to what is necessary.

The second, equally important aspect is training the user. A well-configured tool can reduce many risks, but it is no substitute for sound rules of use.

Employees need to know how to ask the AI the right questions and, above all, when to stop and check. For example, they can ask it to draft a reply to an email, but they should check it before sending it;

• check who the recipient is and who is copied in;
• review the attached documents;
• or expressly instruct them not to take any action until they have received your confirmation.

That is why integrating these tools is not just a matter of configuring permissions. They must be accompanied by a clear internal policy and practical training for those who will be using them.

The rule could be summarised as follows: minimum authorisations from a technological perspective and sufficient oversight by staff. The greater the scope for action we grant to AI, the stricter both controls must be.

4. It is the integration that needs to be audited, not just the tool

A tool may offer good security guarantees and yet still be poorly integrated into our organisation.

Therefore, once the supplier has been analysed, we need to drill down a level and carry out a technical check to see what we are connecting and how we are doing it:

• which systems the AI will be able to access,
• on what grounds,
• what permissions it receives,
• what information you can retrieve,
• what actions it can perform
• and what records remain of their activities.

And the review must also be carried out from the other side. We need to check our own systems:

• if the Microsoft 365 permissions are set up correctly,
• if there are any joint accounts,
• if we retain unnecessary accesses,
• whether we can quickly revoke an integration
• or what would happen if an incorrect action were to alter or delete information.

ChatGPT Business, as we have seen, offers significant safeguards for business use. However, ultimate security does not depend solely on ChatGPT. It also depends on how we integrate it, the permissions we grant it, and how our own systems are configured.

Ultimately, simply auditing the supplier is not enough. We need to assess the intersection between AI and our company.

5. There is some information that perhaps ought not to be available

Just because we can technically connect a tool to our entire SharePoint environment does not mean we should do so.

Human resources, medical information, particularly confidential documents, business secrets, information subject to professional secrecy or certain financial repositories may require additional restrictions.

The question should not simply be “Is ChatGPT safe?”, but also “What information do we want ChatGPT to be able to access?”.

Security also depends on how we configure it. Enterprise AI solutions are increasingly incorporating safeguards for security, privacy and control.

But Simply purchasing a secure tool does not automatically make any use we make of it secure.

Before connecting an AI tool to its email, calendar, CRM or document management system, the company should carry out a simple check:

• to identify which data you will be able to access,
• what permissions it will have,
• what actions he will be able to take
• and what information should be left out.

Because, in this new landscape, a significant part of the risk no longer lies solely in what artificial intelligence knows.

It all depends on the keys we choose to give it.

More on AI tools and compliance

• The Law of Artificial Intelligence in the EU: a comprehensive framework for the regulation of AI
• Your AI crosses borders. And what about your legal compliance?
• Data Act 2025: new rules for sharing IoT and cloud data in Europe

This article is for information purposes only and does not constitute legal advice. A case-by-case assessment is required to evaluate a specific case.