liveBook · Manning

Auditability

As explained in AI Governance

Auditability is the degree to which actions, decisions, and system behavior can be examined and verified from available records and evidence.

Auditability concerns how well actions, decisions, and system behavior can be examined from available records. In GenAI deployment, logging and auditability measure how much telemetry an organization can inspect and export; that visibility shifts with deployment posture. section 2.5

Table 2.2 Responsibility Allocation Across Deployment Postures
Risk dimension
SaaS Consumer
API Integrator
Model Hoster
Data residency & exposure
Vendor retains input/output logs and RAG data (unless zero-retention is enabled for logs)
Only the prompt and any supporting data sent to the model leave the organization; RAG data and logs stay local (if zero-retention configured)
All prompts, outputs, and supporting data remain within the organization’s controlled environment
Guardrail control (input / output)
Fixed, opaque vendor filters
Internal pre-/post-processing plus vendor safety filters (if any)
Fully custom guardrails on top of what is built into the model; must be built and maintained internally
Logging & auditability
Vendor and model-provider logs with limited organizational visibility. Log exports controlled by vendor
Full application-layer logs under organizational control; vendors may keep logs for abuse monitoring if zero-retention is not enabled
Complete stack logging under organizational control; full visibility into all requests, responses, and system events
Security responsibility
Primarily handled by vendor, except SSO and any admin-exposed configuration options
Organization secures user interface, middleware, and data pipelines; vendor secures model-hosting stack
Entire stack security (servers, model-serving software, middleware, and endpoints) fully owned and maintained by the organization
Regulatory accountability
Indirect: org relies on vendor attestations, certifications, and compliance reporting
Clearer data lineage and log access simplify regulatory mapping and evidence collection
Direct: organization holds full accountability for meeting and demonstrating all applicable requirements
Model-behavior liability
Vendor responsible for validating model safety, often with opaque evaluation criteria
Vendor evaluation plus organization-managed prompts, constraints, and content filters
Full responsibility for evaluating, monitoring, and mitigating unsafe or undesirable model outputs
Risk of misinterpretation by end users
Vendor-supplied onboarding guides; limited control over interface design
Provide targeted user guidance and tailor interface design to reduce user error
Full responsibility for user training and interface clarity
Post-deployment governance
Limited to vendor dashboards and monitoring tools; organizations depend on vendors for alerts and periodic reports
Organization maintains internal dashboards, monitoring, and playbooks; can supplement vendor alerts
Full internal control of monitoring, escalation processes, model documentation, and decision logging; all governance artifacts built and maintained in-house

Compare how deployment postures allocate logging visibility and control over log exports. section 2.5

How AI Governance covers it

How deep it goes
Frames logging visibility and export as a high-level risk dimension, deferring detailed security, privacy, and compliance analysis.
Examples
Uses vendor-controlled SaaS log exports and limited organizational visibility as a concrete case.
What you can do after
Evaluate deployment choices by comparing who can inspect and export telemetry.
How it connects
Connects auditability to SaaS, API, and model-hosting deployment postures.

Also explained in

Compare how 9 books explain auditability →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.