Building the space between prohibition and neglect

Control over enterprise AI usage currently swings between two extremes: an outright ban, or effective neglect. A ban costs efficiency and competitiveness, and because teams simply move to personal devices, it tends to make usage less visible rather than less frequent. Neglect surfaces later, all at once, as leaked information, poor decisions, and an inability to audit.

What is needed is the operation in between: allow the usage, decide how far it may go, keep the reasoning, and revoke it when conditions change. That operation is the product.

The unit of control is not an approval but a conditional execution right. Who requested it, against what, within which scope, and until when — together with why it was permitted. That set is recorded as structured data.

We also avoid reducing decisions to allow or deny. Permit with sensitive content redacted, permit once further evidence is supplied, permit within a narrowed scope. Enterprises do not want a mechanism that stops work; they want one that keeps work running on the safe side.

We are currently building Phase 0 — making usage visible. The goal is to record where, by whom, what is being sent to which AI service, and to make a risk register possible.

From there the system extends to human-issued conditional execution rights, then to optimising rules from accumulated decisions, and then to communication between AI agents. The data structures used at each stage are common from the outset, so that later stages do not require rebuilding the earlier ones.

Because the product itself incorporates generative AI, we verify the following each time as an internal standard: the magnitude of harm from an incorrect output, whether human review is required, how outputs are evaluated, defences against prompt injection, whether cost and latency hold up, and the preservation of logs, auditability and reproducibility.

We do not believe we can sell a standard we are unable to meet ourselves.