Plain-English Briefing
The AI can see your business data now. Start with permissions.
AI tools are moving from answering questions to working across company data. That can be useful. It also changes the question from “does the answer look right?” to “what was this thing allowed to see and do?”
The useful bit is also the risky bit
An AI that can only answer from a blank chat window is limited. An AI that can search your documents, inspect a spreadsheet or bring together information from several systems is much more useful.
That is where the new generation of business AI is heading. Recent product announcements describe tools that can work with connected company data and help investigate questions across it. The capability is not the problem. The permission model is the part worth examining before anybody connects it.
A simple rule: if a member of staff should not be able to read something or change something directly, an AI working on their behalf should not get that power indirectly.
Three questions before you connect anything
1. What can it read?
“Our company data” is not one permission. It is customer records, pricing, payroll, contracts, internal notes and the shared folder everybody forgot existed.
Ask for the actual list: which systems, which folders, which accounts and which fields. If the answer is “whatever the connected user can access”, that is a warning sign, not a technical explanation.
2. What can it change?
Reading a job record and changing a job record are different capabilities. Drafting an email and sending one are different capabilities. Finding an unpaid invoice and marking it paid are very different capabilities.
Start with read-only access. If the tool later needs to take an action, give it one narrow action at a time, with a human approval step where the consequence matters. A useful AI does not need to be an unrestricted employee.
3. Who can see what it did?
If an AI searches a customer file, updates a quote or sends a message, there should be a record of that event. You should be able to answer who initiated it, what information it used, what it changed and whether a person approved the result.
“The system keeps logs” is not enough. Ask where the logs are, how long they remain available and whether somebody in your business can actually review them.
Do not start with the most sensitive system
The sensible first project is usually a bounded, read-only task: finding information in a defined document set, classifying incoming enquiries or preparing a draft for somebody to check.
It is rarely the finance system, the whole mailbox or every customer record on day one. A small boundary makes it easier to test the answers, remove access and see whether the tool is helping before it becomes part of the plumbing.
Where private AI fits
Keeping an AI on your own hardware can reduce the number of places your data travels, and it can make the boundary easier to explain. It does not make permissions disappear. A local model with access to everything is still a model with access to everything.
The strongest setup is the boring one: a limited data set, a named purpose, read-only by default, explicit approvals for changes and logs somebody checks. Private infrastructure helps with control; it is not a substitute for control.
Worth doing this week
- Write down the first task you want AI to help with in one sentence.
- List the minimum files or records that task needs. Anything else stays disconnected.
- Test with read-only access and sample data before using live customer information.
- Ask the supplier to show you the activity log and the removal process.
If a supplier cannot answer those questions plainly, you are not ready to connect the system. That is a useful answer, and it has cost you nothing.
Where this came from
- OpenAI — an announcement about AI working with connected business data
- Anthropic — current product and safety announcements
More plain-English briefings on the Inferred blog. If you are deciding whether a private system is appropriate, the private AI overview explains the trade-offs without pretending every business needs one.