AI you can put in front of a security review
In an operational setting, AI is only deployable if you can state what it is permitted to do, prove what it did, and cap what it costs. Everything below is a control that exists, not a principle we intend to honour.
Controls
What governs an AI request
These are properties of the AI Intel execution layer, applied to every AI surface in the platform.
Bounded before execution, not filtered after
A request passes role and tenant entitlement, then a per-user budget, then grounding, before it reaches a provider. An action outside its boundary is refused up front rather than generated and then suppressed.
Grounded in your data, or it says so
Answers are assembled from your live operational context — your sites, incidents and personnel. Where a question cannot be grounded, the answer says so rather than filling the gap from general knowledge.
Attributable to a person
Every AI action is attributable to the user who initiated it and recorded for review. There is no anonymous system actor that can act on a tenant's data without a trace.
Tool authorisation by risk class
What an AI surface may do, as opposed to say, is governed by risk class. A conversational surface cannot become an unaudited command channel.
Spend is capped, not reconciled afterwards
Per-user budgets are enforced before a request executes. An organisation cannot be surprised by consumption it never authorised.
Your provider, where your plan allows
Organisations on the plans that include it can supply their own AI provider credentials, so inference runs under a contract they hold directly.
Limits
What Virtik AI does not do
Stated as plainly as the capabilities, because the boundary is the part that matters in review.
- Virtik does not use Customer Data to train generalized or shared AI models offered to other customers without the customer's express contractual authorization. Virtik does not control or represent the data-retention or training practices of AI providers or customer-connected accounts.
- AI output is decision support. It does not dispatch, escalate or close an incident on its own.
- Model providers are third parties; which ones are available depends on your plan and configuration.
- No AI capability is claimed here that is not enforced by the execution layer described above.