SecureLLM.aiHome › What should we ask before we approve it?
Before you approve
Before you approve an AI assistant on your documents: the questions to ask
The short answer
You don't need to be technical to approve it well. You need five questions, and a sense of what a good answer sounds like.
Security is where these projects stall: in one 2026 survey of 750 IT and security leaders, 87% had delayed a rollout over data security and how information is managed. Most of those questions have clear answers. The risk is approving before anyone has asked them.
Think of hiring a very fast research assistant
Before they start, you'd decide which keys they get. You'd want to know which rooms they can enter, what they'd do with a strange note left in a file, and whether anyone keeps a log.
The five questions, and what a good answer sounds like
1. Could it leak?
Could it show someone information from a file they aren't allowed to open, including in another department?
A good answer: it copies each document's access rules along with its text, and checks who is asking before it searches anything. And we've reviewed who can open what today.
2. Could someone trick it?
With instructions hidden inside a document, or a planted file that steers its answers?
A good answer: it treats what it reads as information, never as orders, and it can't do anything risky without a person approving.
3. Can we prove who saw what?
If something goes wrong, can we show who asked what and what they were shown? When we must delete something, is it gone everywhere?
A good answer: it names the record that's kept, who can read it, how long it's kept, and how deletion reaches every copy.
4. What changes if it can act?
Will it only answer, or can it also send emails, update records or start tasks?
A good answer: it acts with the access of the person who asked, never more, and anything it can't undo waits for a person.
5. How do we check it before signing?
How will we know all of this works, rather than taking it on trust?
A good answer: a test where people from two departments ask the same question and each gets only what they may see. For higher stakes, someone independent tries to break it before launch.
What goes wrong
- Approving on a demo. A demo uses one person's access. Leaks show up when two people with different access ask the same thing.
- Skipping the permissions review. The assistant finds files that were shared too widely, in seconds. Microsoft publishes a whole deployment blueprint for oversharing for this reason.
- Treating documents as safe input. A document can carry hidden instructions.
Why security teams care
The OWASP security foundation's 2025 list of risks for AI applications puts hidden instructions ("prompt injection") first, and leaks of sensitive information second, up from sixth (Infosecurity Magazine).
The questions, ready for your meeting
- Does it check who is asking before it searches anything? Does every passage carry its document's access rules?
- Have we reviewed who can open what, before we connect it?
- What does it do with instructions it finds inside a document?
- What record is kept of who asked what and what they were shown, who can read it, and for how long?
- When we delete something, how do we know it's gone everywhere?
- If it can act, whose access does it use, and what needs a person's approval?
- Can we see it tested with two people from different departments before we sign?
Other names you may hear
- Oversharing
- Files shared more widely than anyone realized.
- Permission-aware retrieval
- Checking who may see what as part of the search. Why it must come first.
- Prompt injection
- Hidden instructions inside a document: a note slipped into a briefing pack.
- Data poisoning
- A planted, misleading document: a forged page in the binder.
- Audit log
- A record of who asked what and what they were shown: the visitor book.
- Red-teaming
- Hiring someone to try the locks before launch.