Enterprise Guide · 3 July 2026 · 5 min read
How to choose a safe AI assistant: a checklist for institutions
Before you compare features, compare data paths. A short checklist for evaluating any AI assistant a regulated organization is thinking of adopting.
Most AI evaluations start with a feature list — how well it summarizes, how many languages it speaks, how the interface looks. For a ministry, a bank, or a hospital, that is the second question. The first is where your data goes.
Here is a short checklist you can take into any vendor conversation, in roughly the order that matters.
1. Does my data leave the machine?
Ask it plainly: when I upload a document or ask a question, does that content leave my computer or my network? If the answer is yes, everything that follows — retention, jurisdiction, training — is now your problem to manage. If the answer is no, most of the risk disappears at the source.
2. Where would it run, and who operates it?
There is a real difference between software you install on your own hardware and a service someone else hosts. On-premise or on-device means you control the machine, the network, and the power switch. A hosted service means you are trusting another operator's controls and another country's laws.
3. What does the vendor learn about our usage?
Even licensing and analytics can leak information. Ask what the vendor's own servers receive: just a licence check, or a stream of usage data? The less that leaves, the smaller the surface you have to govern.
4. Does it work with the internet unplugged?
An assistant that keeps working offline is not just convenient during an outage — it is proof that the intelligence really lives on your side. If pulling the network cable breaks it, the work was never happening locally.
5. Is it usable by the people who will actually use it?
Only after the data questions are settled does the ordinary product evaluation begin: is it bilingual, is it fast enough on the hardware you have, will staff actually reach for it instead of a consumer chatbot. A safe tool nobody uses does not reduce risk — the shadow usage just moves elsewhere.
Sanad was designed to pass the first four questions by construction, so the evaluation can move quickly to the fifth.