AI becomes more useful when it understands what you're working on. Give it access to your calendar, and it can help you manage appointments. Give it access to your email, and it can help you find information and prepare replies.
But every new connection raises a question.
How much access does AI really need to help you?
I've been thinking about this as a trade-off between usefulness and privacy. I still think that's a helpful way to look at it, but there's more to the picture.
We don't always need to give AI more access to get better results. Often, we need to give it the right access.
Why access makes AI more useful
Imagine asking an AI assistant to help you prepare for a client meeting.
Without any background, it can suggest a meeting agenda or give you general advice.
If it can read the relevant emails and documents, it can do something more useful. It might identify what you've already agreed, find questions that still need answers, and prepare a summary of the client's situation.
That's a real improvement.
But does it need access to every email you've ever received? Does it need your entire company file system? Does it need permission to send messages on your behalf?
Probably not.
The useful information might be contained in one email thread and three documents.
The goal is to give AI the information it needs, without giving it access to everything else.
Privacy comes before access
Before we think about how much information to share, we need to know where that information is going.
This is where I use the Golden Rule of AI Data Privacy:
Never enter confidential, client, personal or private business information into Non-Private AI.
I divide AI systems into two groups for this purpose.
Private AI has suitable privacy protections and has been approved by your organisation for the work you're doing.
Non-Private AI has not been approved to handle your organisation's private information.
For example, your company might provide an approved business AI workspace. That may be suitable for confidential work within the rules your organisation has established.
A personal AI account or an unfamiliar web tool should not be assumed to offer the same protections.
The important distinction isn't whether you're paying for the tool. It's whether the system has been checked and approved for the information you're using.
I've written more about the differences between consumer, business and local AI.
What happens when AI can take action?
There's another part of this discussion that I think deserves more attention.
We used to spend a lot of time talking about what information we could give AI.
Now we also need to consider what AI is allowed to do with that information.
Think about an AI assistant connected to your email.
There are several things it might be allowed to do.
- Read: Find messages and summarise them.
- Draft: Prepare replies for you to review.
- Send: Send those replies without asking you first.
These are very different levels of permission.
Being able to read an email doesn't mean the AI should automatically have permission to send one.
And being allowed to send an email doesn't mean it should be able to delete your messages or change your account settings.
This is particularly important with AI agents, which can use software tools and carry out tasks on your behalf.
An agent might book appointments, update customer records, create documents or send invoices.
Those are useful capabilities. But when an AI can change something in the real world, mistakes can have consequences beyond an incorrect answer on your screen.
What can go wrong?
There are two kinds of problems worth separating.
The first is sharing information where it shouldn't go.
In 2023, Samsung restricted staff use of generative AI after employees uploaded sensitive company code into ChatGPT. The concern was that confidential company information had been sent to an external service without the required safeguards.
Read Bloomberg's report on the Samsung incident.
The second problem is giving a system more permission than it needs.
Imagine you've connected an AI assistant to your email. You ask it to find messages about a client project and prepare a summary.
If the assistant can only read messages, the main risks concern the information it can access and where that information goes.
If it can also send emails, a mistake might result in confidential information being sent to the wrong person.
That second example is hypothetical, but the risk is real.
And we don't need to assume the AI provider is careless or insecure for either kind of problem to matter.
Even a well-protected system can produce an incorrect result or take an action you didn't intend.
Give AI the permissions it needs, not every permission available
Before moving into AI consulting, I worked in software development and computer security, including on the Solaris Networking Security team at Sun Microsystems.
One principle from that work still shapes how I approach AI systems.
Give a system only the access it needs to do its job.
In security, this is often called the principle of least privilege.
The name sounds technical, but the idea is straightforward.
If an AI assistant only needs to read your calendar, why give it permission to change appointments?
If it needs to prepare a draft email, why let it send messages without approval?
If it only needs documents from one client project, why connect it to every folder in the business?
Limiting access doesn't necessarily make AI less useful. It can make the system easier to understand, test and trust.
Privacy and usefulness aren't always opposites
It's tempting to think there's a fixed choice between powerful AI and private AI.
I don't think that's the right way to look at it.
An approved business AI system may allow you to work with confidential information while protecting that information under suitable agreements and controls.
A properly secured local AI system may let you process information without sending it to an external model provider.
And an AI assistant with limited permissions may give you almost all the benefit of one with much broader access.
There are still trade-offs. Some privacy controls cost money, require setup or limit which tools you can connect.
But we shouldn't assume that the most useful system is the one with the most access.
A well-designed AI system can be both useful and carefully limited.
Who is responsible for making these decisions?
I think it's important to separate what the organisation needs to do from what an individual AI user needs to do.
Your organisation should decide which AI systems are approved, what information those systems may handle, and what they are allowed to access or change.
That may involve checking provider contracts, privacy protections, data retention, security settings and connected tools.
The Vendor Evaluation Card offers a starting point for those checks.
Your responsibility as an employee is more straightforward.
Use the approved systems, follow the rules for the information you're handling, and don't give AI access or permission that hasn't been authorised.
If you're self-employed, you'll need to make these decisions yourself, with specialist advice when necessary.
Four questions before connecting AI to your work
When you're considering a new AI tool or connection, I suggest four questions.
1. Is this an approved Private AI system?
If you're handling confidential information, start with the Golden Rule.
2. What information does it actually need?
Could you give it a few selected documents rather than your entire drive or mailbox?
3. What can it do with that information?
Can it only read and suggest, or can it send, change and delete things?
4. What happens if it gets something wrong?
Can you review important actions before they happen? Can you see what the system did and undo a mistake?
For more important tasks, I would normally start by letting AI prepare a recommendation or draft. Once the workflow has been tested and its limits are understood, we can decide whether some actions should happen automatically.
Getting the balance right
I don't want privacy concerns to stop people experimenting with AI. There is too much useful work we can do with these tools.
But I also don't think we should connect AI to every available system simply because we can.
At the start, AI may help you with general questions and drafts. As you become more confident, it can work with approved business information, use connected tools and help manage repeatable tasks.
As the system becomes more capable, we need to be clearer about its boundaries.
For me, the useful questions are:
- Where does my information go?
- What can the AI access?
- What can it change?
Those three questions cover much of the practical trade-off between usefulness, privacy and control.
The aim is to get the benefit of AI while keeping authority over the information and systems we're responsible for.
And the starting point remains the same:
Never enter confidential information into Non-Private AI.
If you'd like to put that into practice, start with the Golden Rule of AI Data Privacy. If you're setting up AI for a team, you may also find my guide to creating an AI policy useful.