FAQ¶
What does Venturi actually measure?¶
Venturi measures AI consumption events, their likely owners, the cost, energy, and carbon implications, and the evidence quality behind each of those conclusions. Every surface shows both the attribution outcome and the caveats attached to it.
What is Venturi’s AI System of Record?¶
Venturi is the system of record for AI: an authoritative, replayable record connecting AI activity, accountable ownership, workflows, systems, normalized cost and accepted business outcomes. Customers choose the domain for their decision. Workflow Economics Core supports economic decisions and is not a prerequisite for another domain.
Every measurement retains source coverage, confidence, freshness, correction history, and unresolved links. See the 42-view contract for the decision views and their evidence requirements.
How does Venturi measure AI inference ROI or E-FTE?¶
Venturi starts with Cost per Accepted Result, not raw token price or call cost. The denominator is the customer-defined accepted business outcome, and paid failures, retries, rejected work, and rework remain in the cost basis. Net Realizable Benefit separates cash savings from released capacity, while Monthly Attribution Reconstruction Effort measures recurring human effort spent reconstructing ownership, cost, and outcome relationships. E-FTE is reported only when the labor baseline, current effort, role basis, period, and non-overlapping work are evidenced; otherwise the result is a labeled scenario or unavailable. These views are visible only at their supported evidence state, never as automatic savings or ROI claims.
Do I need to send events, or is connecting my cloud enough?¶
Connecting your cloud is enough for most customers: it gives cost and usage attribution from your provider’s own billing and logs. Add events only if you want per-request resolution (tokens, cost, latency, identity per call).
Is there an SDK?¶
No Venturi-specific SDK is required. Integration is the read-only Terraform
connector plus, optionally, a plain HTTP ingestion API
or a drop-in proxy you point your existing provider
SDK at. To read attributed cost, energy, and carbon back out
programmatically, use the REST API. The typed TypeScript
(@venturi/sdk) and Python (venturi_sdk) SDKs are
distributed in your onboarding artifact bundle and private package channel, and
the bundle manifest pins each package’s filename, version, and SHA-256 digest.
Do not install them from public npm or PyPI; the public PyPI project named
venturi is unrelated.
Can Venturi change anything in my account?¶
No. Every onboarding module is read-only, and a platform gate blocks any module that requests a write from shipping. The customer-run check in Verify & confirm shows that the module grants no write permission, and automated credential validation confirms that probed write calls are denied before a connector activates. Active (Gate) never changes your account either: it sends a signed recommendation that your own enforcement point decides whether to apply. See Trust & security.
Does Venturi see my prompts or completions?¶
No prompt or completion text is persisted. Read access to a log source is not storage of its content: scope log grants to the log groups, log views, BigQuery datasets, Log Analytics workspaces, or Event Hubs namespaces that supply metadata. Where a granted source still carries content-bearing records, the collector applies a field allowlist in memory and discards prompt and completion fields before a record is materialized; they are never persisted, logged, retried, or exported. The event schema has no content field; only metadata (tokens, model, cost, identity, timing) is materialized.
Will the proxy slow down or break my traffic?¶
Not through a failing Venturi dependency. The decision-time interceptor is fail-open within a 50 ms P99 end-to-end budget, bounded by a cooperative timeout against a cumulative deadline: if the index lookup, the event stream, or any other Venturi dependency behind the proxy is slow or down, your request is forwarded to the provider unmodified. The proxy endpoint itself is on your request path, so a request that cannot reach it is not forwarded, and a provider key that is not registered to your tenant is rejected before anything is forwarded. Configure client-side failover to the provider’s direct base URL if a workload cannot accept that dependency; see Proxy endpoint reachability.
Where does Venturi run?¶
In a dedicated single-tenant data plane, in one of two deployment modes.
Self-hosted mode runs inside your cloud trust boundary, and you reach the
dashboard and API at your own instance host. SaaS mode runs in a
Venturi-operated tenant pinned to your contracted residency lane, and you reach
the dashboard at https:// and the API at
https://. See Deployment modes.
How do I rotate or revoke access?¶
Rotate with terraform apply (new external_id / federation subject), revoke with
terraform destroy, and record both actions in Integrations. On Google Cloud, validate the
replacement credential before revoking the previous credential, because
wif_ accepts both subjects during the change. An AWS external
ID, an Azure federated subject, and the Azure client-secret fallback are single
values, so rotate each as a coordinated cutover. Review each federated trust at
least every 12 months, and disable the connector before destroy. Details are in Rotate & offboard.
How long does onboarding take?¶
An authorized developer, Engineer, or Admin can reach the
engineering active-use milestone in approximately 15 to 30 minutes of hands-on time
by connecting at least 1 source or uploading a provider export file and seeing
the 1st AttributionRecord on the /invocations surface. Persona-complete
production onboarding is separate: Product, Security, Engineering, and Finance
must complete real in-product actions, including Finance acceptance of a
chargeback-eligible export. No human confirmation is a gate. See
Success evidence in the onboarding
guide.
What is attribution_confidence and why is it capped?¶
attribution_ is the descriptive API name for operational confidence,
coper; in an attribution record’s evidence card the same value appears as
interpretation.. It is capped at 0.95: Venturi never claims
certainty. The cap is enforced in the schema and scoring engine. The value sets
the band: Chargeback-ready at coper >= 0.80, Provisional at
0.50 <= coper < 0.80, and Estimated below 0.50.
How should I read observe-only, decision-support, and operationally trusted labels?¶
They describe how the evidence on a product surface can support decisions. Observe-only is informational; decision-support supports human-guided decisions with the stated caveats; operationally trusted means the workflow’s evidence quality supports tighter operational reliance, such as chargeback.
Is the trained Stage B attribution model live by default?¶
Only where a promoted model is in place. Stage B (trained edge-existence
inference) runs in the asynchronous processor for the edges the deterministic
Stage A rules cannot resolve, and a trained model serves a result only when
Venturi has promoted that model version through its human-gated promotion
process and its artifact has been staged into your processor and hash-verified
before it is bound. Otherwise, and whenever the model is unavailable or misses
its 20 ms timeout, the deterministic heuristic baseline answers. Each result
carries a result-source marker (trained model or heuristic baseline) and a
degradation_state, so a baseline result is never mistaken for a model-backed
one; see
Confidence & evidence.
The model-card index links the available model
documentation and records standalone model-card publication status. See
The attribution pipeline, in which RAIL (Resource
Attribution & Inference Layer) is Stage B.
What happens when identity, repository, or billing signals are incomplete?¶
Venturi degrades explicitly instead of fabricating certainty: confidence drops, caveats stay visible, and unknown or partial states are reported as unknown, never silently zeroed or guessed.
Can we use Venturi for chargeback on day one?¶
Only where the evidence supports it. Cost lines are chargeback-eligible at or
above the 0.80 confidence floor; lines below it stay Provisional
(0.50 <= coper < 0.80) or Estimated (coper < 0.50) and are never
auto-billed, so teams adopt chargeback progressively as coverage and confidence
grow.
How does Venturi surface hidden fan-out?¶
Dashboard warnings flag cases where one upstream action triggers an unusually large volume of downstream AI calls, prompting investigation before multiplied cost is misread as normal demand.
How do security and procurement teams validate the deployment posture?¶
Review Trust & security, the customer-run trust-boundary check in Verify & confirm, the Terraform artifacts, and the control-plane / data-plane separation described in How it works. The posture is designed to be reviewable, not implicit.
Which providers are supported?¶
Cloud-native inference (Amazon Bedrock, Vertex AI, Azure OpenAI) via the connector; direct OpenAI and Anthropic via the proxy or ingestion API. The provider picker in guided setup is the tenant-specific authority for enabled sources.
How does Venturi know how much energy and carbon my AI usage produces?¶
For each attributed invocation, Venturi uses a matching energy coefficient with its native unit, workload, and measurement scope. The public catalog preserves GPU-only AI Energy Score values in Wh/query, machine-level Power-Hungry measurements in Wh/inference, and modeled How-Hungry facility values in Wh/query for the stated prompt fixtures.
An estimate multiplies the coefficient by the count of matching workload
units. Output-token totals must not scale query or inference coefficients.
The legacy energy_ field is null. Carbon and water
require compatible source values or the inputs for a separately labeled
calculation. Facility values already include PUE; an additional facility
estimate must identify its overhead assumptions and apply PUE once.
The full formulas, source boundaries, and provenance are in the energy and carbon methodology; the customer-facing summary is on the Energy & carbon page.
Why do some models show no energy or carbon number?¶
A coefficient or required input can be unavailable even when the model appears in the catalog. Venturi preserves null, never zero in that case. A known energy value does not imply known carbon, water, or an eco-efficiency score. Each requires its own compatible inputs; a missing value must not be presented as zero consumption.
See the null-not-zero rule and the methodology.
How accurate are the carbon numbers for my region?¶
For a separately modeled carbon calculation, Venturi multiplies compatible energy in kWh by grid intensity in gCO2e/kWh. The grid factor follows a fixed order: an explicit value you supply, then a live feed, then your cloud region, then your country’s generation mix, and finally the world-average fallback of 475 gCO2e/kWh (always flagged as estimated). Configuring your deployment region replaces the global fallback with a more specific factor.
This calculation does not overwrite source-reported Power-Hungry carbon. The GPU-only AES catalog rows retain null carbon. Any separate estimate requires compatible energy and grid inputs and is labeled as modeled. The full resolution chain and fuel-factor table are in the methodology.
Can I set energy budgets, and is energy the same as cost?¶
You can set per-team and per-cost-center energy budgets (kWh per period), evaluated with the same Advisory (Shape) and Active (Gate) semantics as cost budgets. In Advisory (Shape), a crossed threshold produces alerts and recommendations. Under Active (Gate), enabled only for a named, operationally trusted workload, Venturi issues a signed deny recommendation that your own enforcement point decides whether to apply; the decision-time interceptor never blocks traffic. See budgets and alerts.
Energy is not cost. Energy is sustainability context; an energy multiplier must never be read as a cost multiplier. Reasoning modes, for instance, can be hundreds of times more energy-intensive than standard operation, which says nothing about price. Cost comes from the billing and pricing path; energy comes from the energy engine. The two are always reported separately.
How long is my data kept, and can it be erased?¶
Venturi stores only metadata (never prompt or completion content). Operational data is retained for 13 months by default, configurable from 30 days to 5 years, and the pseudonymized audit log is retained for 5 years. A data-subject erasure request is completed by crypto-shred within 30 days, with a deletion certificate. See data privacy & retention and data-subject rights.
How do I read attribution data?¶
Use the read REST API: GET /v1/attribution returns
attributed cost, energy, and carbon with an evidence card on every record. It is
backed by typed TypeScript and Python SDKs. Your base
URL depends on deployment mode: https:// in SaaS mode,
your own instance host when self-hosted. See
Deployment modes. (The /control/v1 surface on
api.venturi.systems is the public control plane, separate from tenant APIs.)
Where do I send feedback or feature requests?¶
Through the feedback portal at feedback.venturi.systems. For support questions, see Support.
Can Venturi show an individual employee’s AI usage?¶
No. Every cohort read path (adoption views, dashboards, exports, API responses, board packs, and webhooks) enforces a server-side minimum cohort size of 5: a cohort of fewer than five people rolls into its parent unit, never an individual, and it cannot be reconstructed by differencing overlapping cohorts. An Admin can raise the threshold but never lower it below 5. Venturi is built to report on teams and cost centers, not to surveil individuals. See tenant isolation and data privacy & retention.
Who do I contact?¶
Open a case from the in-product Support page (/support) or email
contact@venturi.systems. Report security
issues to security@venturi.systems. See
Support and support tiers.
Sign in to send feedback. Opens in a new tab.
Copy page details
Docs feedback: FAQ
Page: https://docs.venturi.systems/reference/faq/