From policy on paper to policy in the traffic¶
The problem. Your AI policy says personal data stays out of models, calls stay in the EU and records are kept for six months. In practice, guardrails are configured somewhere, applied by each app or not, and "where do our prompts go?" is answered from a wiki page. When the auditor asks for proof, someone spends weeks gathering screenshots and exports, and the proof covers what was configured, not what happened.
With Vidai. Every AI call your apps, agents and people make goes through the gateway. On each one, the gateway checks who is calling, what they may use, what the call contains and where it may go, then records what it decided. Your rules act before a model sees the data, and every decision leaves a record. When the auditor asks, you produce the evidence yourself, for any period, in minutes.
sequenceDiagram
participant App as App or agent
participant GW as Gateway
participant M as Model provider
participant CR as Control room
App->>GW: AI call with a Vidai key
GW->>GW: Who? Which models? What's in it? Where may it go?
alt A rule refuses or blocks
GW-->>App: Refused, with the reason
else Allowed (perhaps masked or pseudonymised)
GW->>M: The call, as your rules allow
M-->>GW: The answer
GW-->>App: The answer
end
GW->>CR: The call's record: what each rule did
What changes for you¶
| Before | With Vidai |
|---|---|
| Each app team builds (or skips) its own guardrails | One set of guardrails applies to every call, from every app |
| "Where do prompts go?" answered from a wiki | Region rules and approved models decide it, and a refused call says why |
| A regulation is a PDF and a spreadsheet of controls | You choose the regulations you follow; Vidai sets the retention they ask for, proposes the policies that serve them and shows each control's coverage |
| Proof is gathered by hand before each audit | An auditor report and an evidence pack for any period, which the auditor checks on their own machine |
| A person asks for their data, and nobody knows where it is | A data subject request finds their calls, makes a copy or erases the content, and says what happened to each call |
Try it in a minute: open Daily › Request logs, choose a refused call, and read Why this happened: every rule that acted on it, in the order the gateway weighed them, each linked to the rule itself.
What you need from IT, once¶
- Apps send their AI calls through the gateway. Changing the base address and the key is the only change an app makes. Calls that don't go through the gateway aren't governed, and Vidai's evidence says so.
- Your providers are connected and placed in the regions they run in, so region rules know where data goes.
- You have the right role. On Enterprise, an organisation admin makes you a Compliance officer; reading what calls actually said is a separate role, Content reader. See The Compliance officer role.
- Tamper evidence is switched on (Enterprise) and the gateway's signing key confirmed, so calls are recorded in a way an auditor can verify. See Switch on tamper evidence.
What you then run on your own¶
- Follow a regulation and take its recommended policies, watching first: Follow a regulation.
- Move a policy from observing to enforcing once you've seen what it would have done.
- Set how long records are kept, decide a clash between regulations, and put records on legal hold.
- Record serious incidents, run access reviews, answer data subject requests.
- Produce the auditor report and the evidence pack.