You can't govern what you can't see.
Tools, plugins and agents arrive before security knows. In our sample, 5 of 6 AI apps were unsanctioned.
Smalldata AI Control Plane
One gateway between your company and every AI model. It authorises, redacts, reroutes or blocks each request in real time, and writes a signed record of every decision. No endpoint agents, no code changes, under 50 ms.
Tools, plugins and agents arrive before security knows. In our sample, 5 of 6 AI apps were unsanctioned.
PII, contracts and source code get pasted into consumer AI. DLP was built for files, not prompts.
Broad database and tool access means one prompt injection can trigger a full table scan.
Blocking drives it underground. Governing it inline doesn't.
Every AI tool, model, plugin and agent, on one map.
Passive inline discovery of web and API AI traffic. Sanctioned versus unsanctioned, users and departments, token volume, and risk severity per app.
Sanctioned Unsanctioned
Every discovered AI destination flows through one inline point.
Decide, inline, what reaches a model.
Redact PII and secrets before they leave. Block or redirect unapproved tools to sanctioned tenants. Sandbox tools and MCP for agents. Set department quotas and model routing.
Policy
IF payload contains IBAN | codice fiscale | customer email → REDACT
IF destination = consumer ChatGPT → REROUTE to enterprise tenant
An evidence-grade record of every decision.
A cryptographically signed, tamper-evident log of prompt origin, sanitised payload, tool execution and token cost, mapped to GDPR and the EU AI Act.
Audit record
every tool, model and plugin, with users, departments and token volume.
PII, financials and secrets redacted inline before they leave.
block or redirect unapproved tools and SaaS plugins at the perimeter.
per-tool policies, schema-level boundaries, and signed traces across every hop.
per-department attribution, quotas, model routing and failover.
an immutable, searchable record of every decision, ready for auditors.
Test a control, such as “block all non-EU-hosted models”, against real traffic before you enforce it.
From a 14-day sample assessment (illustrative)
See and control AI traffic with no endpoint agents and no code changes. Under 50 ms, fail-closed.
An AI system inventory, regulatory mapping, and signed logs you can put in front of an auditor.
Route usage to sanctioned tools instead of banning it, so people keep the productivity gain.
Per-department cost and token attribution, and fine exposure quantified.
A short form. We reply within two working days to agree a time.
A network-level gateway in passive mode. No endpoint agents and no application changes. We'll confirm the exact routing (proxy, DNS or egress) with your network team on the scoping call.
In the EU. Payloads are obfuscated for PII by default in the workbench. A DPA is signed before kick-off.
If you deploy AI systems used in the EU, very likely. The report maps your inventory to risk tiers.
You can switch the same gateway to active protection, with critical leak prevention in 24 hours and full posture in 14 days. Or you can stop there: the report is yours.
All major model providers and self-hosted models; the gateway is provider-neutral.
Yes. Sanctioned tools are recognised and allowed under policy, and their usage is logged.
Under 50 ms of gateway overhead, with fail-closed defaults.