Skip to content

Known limitations

This page is Venturi’s known-limitations register: the single, maintained list of what the platform cannot capture, how it behaves when a component degrades, which attestations Venturi does not hold, and which controls are on the post-onboarding roadmap. It exists because stating ambiguity plainly is the system doing its job: Venturi would rather tell you “we cannot see this” than render a confident number that cannot be defended.

How to read this register

Each entry names the limitation, the behavior you will observe in the product, and the path that narrows or closes it. Nothing here is euphemized: an uncapturable pathway is called uncapturable, a fallback result is labeled a fallback, and an attestation Venturi does not hold is stated as not held. The customer-onboarding baseline, tenant-specific implementation status, and post-onboarding roadmap are defined in Docs authority & product state.

Uncapturable and detected-only pathways

The 100% pathway-coverage target means every observed event is assigned to a named pathway category (none is left as Unclassified: pending rule), and every category is inventoried as captured, detected-only, or uncapturable with a stated reason. It does not mean every event is attributable; see Coverage & unknowns.

The separate 0.95 attribution-coverage objective measures the share of capturable AI usage that resolves at or above the confidence floor. The contracted floor is 0.90 on the Standard support tier and 0.95 on Premium and Regulated (see Commitments by support tier). A complete pathway inventory can therefore coexist with unresolved spend.

Venturi attributes what it can observe. Some routes AI spend takes through your organization cannot be resolved at request granularity with the integrations you have connected, and some cannot be resolved at request granularity at all. These pathways are never silently absorbed: they surface as named output states and as entries in the discovery ledger, and they stay out of the chargeback denominator.

Observed behavior and actions that narrow each pathway’s limitation:

Pathway Limitation and behavior Action
Shadow AI Shadow AI: unsanctioned tools or personal accounts. Not attributable at request granularity; detectable only via CASB, DNS, or admin-API signals. Held as a detected-only line in the discovery ledger with first/last-seen dates and a labeled activity estimate, never an invented dollar figure, never billed Improve detection with CASB and DNS; attribution requires moving traffic onto an observable pathway
Client-side calls Client-side calls: browser or desktop calls that bypass your gateway. Visible, if at all, only in aggregate provider billing. Resolves to not_identifiable; inventoried under the 100% pathway-coverage target, never assigned to the nearest team Route calls through your gateway or send InvocationEvents
Embedded AI Opaque SaaS-embedded AI: vendor features that expose no per-call usage. The vendor’s plan does not expose AI-specific consumption. Shown as bounded or detected-only with the limiting vendor named; excluded from chargeback Vendor admin-API connectors where the vendor exposes usage; otherwise this is a genuine limit
Direct calls Uninstrumented direct provider calls. Calls that reach a provider by a path with no connector or event coverage. Spend appears as an unknown with the reason Uncaptured pathway in the Unknown Spend Analyzer Extend connector or gateway coverage; the coverage map names the would-close integration
Estate coverage Estate coverage. Where no external spend anchor (cloud billing AI line items, GL categories, invoices) exists, total-estate coverage cannot be verified. The coverage report states the figure is “measured against the instrumented estate only; total-estate completeness is unverified” Provide an independent anchor so capture coverage is measured against your real estate

Fallback and degradation behavior

Behavior What happens How it is labeled
Heuristic-baseline fallback If Stage B (trained edge-existence inference) cannot answer inside its 20 ms wall-clock timeout, or is unavailable, a deterministic, conservative heuristic answers instead so attribution processing keeps flowing. AI-traffic fail-open is enforced separately by the interceptor. The result carries an explicit result source marker (heuristic baseline, not trained model) plus a degraded flag, unchanged even when the heuristic is confident; the serving state shows degraded_serving
Fallback results and money A fallback attribution is excluded from the savings counterfactual and adoption cohort rollups, and scheduled chargeback exports hold it back with a stated reason until the same edge is re-observed under trained-model attribution Filterable by the result-source marker in dashboards, exports, and the API: see Confidence & evidence
Provisional and demoted attributions Calibration drift or model age can demote an edge from Chargeback-ready; demotion is automatic, re-promotion never is The evidence card records the demotion reason; demoted edges leave the chargeback base visibly
Confidence is capped Operational confidence (coper) carries a deliberate policy ceiling of 0.95; Venturi never reports certainty Only Chargeback-ready results (coper >= 0.80) are chargeback-eligible; Provisional (0.50 <= coper < 0.80) and Estimated (coper < 0.50) results are labeled and never auto-billed

Attestations and roadmap controls

Tenant-specific implementation status for every security and identity control is recorded as described in Docs authority & product state. The entries below are restated here because a reviewer reading only this register must still get the complete picture:

  • SOC 2 status. SOC 2 Type I readiness/report target on our security roadmap with Type II targeted within 12 months; it is not yet held. Readiness work and audit evidence are distinct from an issued SOC 2 report, and no page on this site implies otherwise. See Compliance → SOC 2.
  • Uniform identity enforcement is on the post-onboarding roadmap. Enforcing role and attribute authorization, SSO-only sign-in, SCIM lifecycle changes, and step-up authentication uniformly across every route, worker, and export is a roadmap item. The status of each of these controls for the routes, workers, and exports your tenant uses is recorded in its implementation status.
  • Record-level tenant isolation and the support break-glass gate are baseline controls, not roadmap items. Both are verified for your tenant at activation and recorded in its implementation status, and both fail closed until they are verified: a cross-tenant request is rejected regardless of role, and support receives no tenant-data access until the break-glass gate is verified.
  • The inference_geo inference-residency control is on the post-onboarding roadmap. Tenant data and support grants are bound to your residency region, and residency checks fail closed.
  • Some commitments are tier-scoped. The 99.95% serving-plane availability step-up, the 15-minute Sev-1 case acknowledgment target (around the clock), and the 5-minute index-freshness SLA apply under the Premium and Regulated support tiers; Standard carries the 99.9% commitment and business-hours case acknowledgment. Incident notices follow the same cadence on every tier. See Commitments by support tier.
  • ISO 27001, ISO 42001, and FedRAMP are not held and are stated as roadmap or out-of-scope rather than implied. See the control framework crosswalk.

Documentation boundaries

  • Discovery-only connectors do not attribute requests. Cloud accounts (AWS, Google Cloud, Azure), the ingestion API and drop-in proxy, and the source-system connector catalog have setup guides. Discovery-only connectors, including CASB, DNS/firewall, and optional eBPF passive capture, produce discovery evidence only and do not imply request-level attribution or chargeback readiness.
  • Model documentation. The AI-transparency page documents model behavior and review controls. The model-card index records the separate publication status: a standalone public Stage B model card is not published.
Give feedback on this page

Sign in to send feedback. Opens in a new tab.

Copy page details 
Docs feedback: Known limitations
Page: https://docs.venturi.systems/known-limitations/