Most AI-risk anxiety is really a list of unanswered questions. Here are the ten that matter, phrased about your organization, loosely following NIST AI RMF's Govern · Map · Measure · Manage. Answer honestly; every gap comes with the control that closes it. Nothing leaves this page.
Why it matters: Without a written boundary, every team invents its own — and the first time you learn about a use case is when it goes wrong.
What closes it: Turtle policies are versioned, priority-ordered rules enforced on every tool call — a written policy the runtime actually executes. Governance →
Why it matters: Accountability that isn't assigned per action evaporates in the incident review.
What closes it: Approvals are single-use grants stamped with who approved what, when — including on-behalf-of delegation. Approvals & HITL →
Why it matters: If risky actions and routine ones travel the same path, your control is a hope, not a workflow.
What closes it: The policy engine gates matching calls until a human approves; unapproved writes are a no-op, whatever the prompt said. Approvals & HITL →
Why it matters: You can't assess a blast radius you can't enumerate.
What closes it: Access is granted per resource — agents, tables, knowledge bases, integrations — so the inventory is the configuration itself. Governance →
Why it matters: Unbounded reads mean any prompt injection is a data-exfiltration attempt with credentials.
What closes it: Server-side workspace isolation, per-workspace PII policies (mask · hash · remove · flag) and a prompt-injection shield over untrusted content. Trust center →
Why it matters: The model clause is now the first question legal asks — 'we're not sure' stalls the whole program.
What closes it: A public subprocessor and model-provider list, BYO keys so prompts run under your own provider agreement, and workspace data never trains models. Trust center →
Why it matters: Reconstructing what an AI did from application logs is archaeology; auditors want lookups.
What closes it: Every run and tool call recorded — sanitized inputs and outputs, PII flags, read/write classification — pinned to the config version it ran under. Audit & evidence →
Why it matters: Ungoverned spend is how pilots get cancelled — by the CFO, not the security team.
What closes it: Per-message token and USD cost tracking, rolled up to worker and workspace scorecards. Analytics →
Why it matters: A budget the system can't enforce is a suggestion.
What closes it: Hard daily and monthly caps with breach-date forecasting; the cap halts execution, not just alerts. Budgets →
Why it matters: If the kill switch lives with one builder, your incident response has a single point of failure.
What closes it: Circuit breakers auto-pause workers on cost or error thresholds, and admins can halt any worker from the console. Circuit breakers →
Isolation, runtime enforcement, audit evidence, framework mappings and subprocessors — the security review, answered in public.
Read itTen questions on process, data, governance and ambition — with a specific starting point at the end.
Take the assessmentWeek-by-week plan, pre-defined success metrics, explicit exit criteria — evaluation the way procurement wants it run.
See the framework