A permission-aware AI assistant is one that reads only the emails, files and records the person asking is already allowed to see. The assistant should never have wider access than its user; AI use should be switchable per person, anything it does on its own should be logged, and there should be an organisation-wide usage limit. That way AI speeds up work without opening a new route for data to leak.
Key takeaways
- The rule is simple: an AI assistant must not read or summarise any data the person asking cannot see.
- In IBM's 2025 research, 97% of organisations that suffered a breach of an AI model or application lacked proper AI access controls.
- Treat AI as a separate permission: who may use it, in which mailboxes and folders, and how much.
- Tie permissions to departments, so AI access disappears automatically when someone leaves or changes team.
- Actions the AI takes on its own should be logged separately from human actions.
Why do AI assistants put company data at risk?
In conventional software, an employee simply cannot open a folder they have no rights to. An AI assistant that indexes "all company data", however, can return a paragraph from that very folder in answer to a question. The cause is rarely malice; it is badly designed access, where the assistant reads through a broad service account and the user's own rights are never checked at query time.
The second risk is tools the company never approved. Verizon's 2026 Data Breach Investigations Report (DBIR) found that 45% of employees are now regular AI users on corporate devices, and 67% of users reach these services from corporate devices using non-corporate accounts. Contracts, price lists and customer emails end up pasted into accounts the business has never seen; our guide to ChatGPT and company data security covers this in depth.
The cost is real. According to IBM's Cost of a Data Breach Report 2025, 13% of organisations reported breaches of AI models or applications, and 97% of those lacked proper AI access controls. The OWASP Top 10 for LLM Applications 2025 likewise lists sensitive information disclosure (LLM02) and excessive agency (LLM06) among its top risks.
How does permission-aware AI actually work?
A permission-aware assistant asks "may this person see this document?" before it writes an answer. In retrieval-augmented generation (RAG) systems, which answer from your documents, that check has to happen while documents are being retrieved; filtering after the answer is written is too late. Our article on RAG and enterprise LLMs explains the plumbing.
The table below compares three common set-ups.
| Criterion | Broad-access assistant | Separate knowledge base | Permission-aware AI |
|---|---|---|---|
| Data it reads | Everything the service account can see | Documents uploaded by hand | Mailboxes and folders the asker is entitled to |
| Permission changes | Not reflected | Documents must be removed manually | Follow the source system immediately |
| Per-user AI right | Usually none | Tool-level only | Per user and per area |
| Logging | Query log, little context | Limited | AI actions recorded under their own identity |
| When someone leaves | Account must be closed separately | Shared access may linger | AI access ends with the identity |
Keeping permissions in one place is what makes this work. If file rights live in one system, mailbox rights in another and AI rights in a third tool, getting all three right at once is down to luck. Our guides to file sharing permissions and mailbox permissions and delegation cover those layers in detail.
How do you set up permission-aware AI? Seven steps
These steps apply to businesses of any size, whichever product you choose.
- Map your data. List the sources the assistant could reach (mailboxes, file areas, customer records) and classify each by sensitivity.
- Fix human access first. If everyone can see every folder, so can the AI. Build a department-based structure on the principle of least privilege.
- Make AI a separate permission. "May read this mailbox" and "may use AI chat in this mailbox" should be different rights; consider keeping AI off by default in HR, legal and finance areas.
- Insist on query-time checks. Ask the vendor whether the user's rights are checked before an answer is generated, then test it with a restricted account.
- Build in logging and traceability. Automatic actions such as classification or quarantine should be logged under a distinct identity, and it should be clear which document an answer relied on.
- Set usage limits. Monthly caps per organisation and per person keep both cost and unusual activity under control.
- Write the policy and train people. Spell out which data may be processed by which tool in a company AI acceptable use policy; for personal data, see our guide to AI and KVKK.
What should you ask a vendor?
"Your data is safe" is not an answer. Ask for written replies to these questions: under which identity does the AI read data? When someone is removed from a mailbox, how soon does the assistant stop reading it? Can AI be switched off organisation-wide or per area?
The questions grow once AI can act. If an AI agent can send emails or create orders, important steps should wait for human approval, and source data should be connected read-only to limit prompt injection. Also ask whether security checks depend on the AI: spam and spoofed-sender checks should not stop when an AI allowance runs out.
How does Smart360 make permission-aware AI easier?
Smart360, developed by Digital Bridge, is a family of business applications (SmartMail, SmartFiles, SmartOperator and SmartPass) managed from a single panel. In Smart360, AI works within each person's permissions: chat reads only the mailboxes and files that person can access. AI is a separate permission in both SmartMail and SmartFiles, granted person by person, and in SmartFiles it can be switched off for the whole organisation.
Take a concrete scenario. At an accountancy practice, the bookkeeping team uses "Chat with folder" to ask questions about tax return documents in client folders, while AI has never been enabled in the HR area. When a team member is moved out of the accounting department, the access they had through that department is removed across every product automatically, and the change appears on their open screens straight away. The next day, the same person's questions can no longer draw answers from those folders.
On the logging side, actions the AI takes on its own appear in the SmartMail audit log as "System", not as a person. The monthly AI allowance is visible in the panel; when the cap set for the organisation or a person is reached, AI operations pause until the next period while access to email and files continues. SPF, DKIM, DMARC and look-alike domain checks are run by the software itself, so they keep working even when the AI allowance is used up.
The products do not reach AI services directly but through Smart360's central layer, and each organisation's data sits in a separate database for each product. For details of the AI services in SmartMail and SmartFiles, see the Smart360 artificial intelligence page; for the family and its single panel, see the Smart360 product page.
How can Digital Bridge help?
Digital Bridge makes sure the access structure is sound before AI is switched on. Our team can plan your department and permission structure with you, along with which areas should have AI enabled. If you want to find access gaps in your current systems, we can carry out an assessment as part of our cyber security consultancy.
If you are considering a custom AI solution connected to your own systems, the same permission principles need to be written into the contract and the architecture from day one. Get in touch through our contact page; all related articles are collected on our AI topic page.
Next step: roll out AI together with permissions
To use AI safely, first be clear about who can reach what, then add AI on top of that access as a permission of its own. Smart360 brings both layers together in one panel: users, departments, product access and AI usage are all managed in the same place.
To choose a package and start straight away, use the Smart360 purchase page; if you would like to explore the product family first, visit our Smart360 page.