Skip to content

2026 04 26 ai lowcode observability telemetry governance

Jump to bottom
github-actions[bot] edited this page Apr 30, 2026 · 2 revisions

What observability and telemetry model is required to govern Artificial Intelligence (AI) and low-code systems at scale?

Research Question

What observability and telemetry model is required to govern AI and low-code systems at scale, specifically, what must be logged, at what frequency, and at what level of granularity, including prompt and response logging, decision traceability, linkage between user intent and system actions, cross-system correlation, and the ability to reconstruct events for audit, debugging, and compliance purposes?

Scope

In scope:

  • What must be logged for AI systems: prompt content, response content, model version, token counts, confidence scores, retrieved context (for Retrieval-Augmented Generation (RAG) systems), tool calls made by agents, and decisions taken
  • What must be logged for low-code automations: trigger events, action sequences, data inputs and outputs, Application Programming Interface (API) calls made, errors and exceptions, and human override events
  • Logging granularity and frequency: the trade-offs between full-fidelity logging (expensive, privacy-sensitive) and summary logging (cheaper, less auditable)
  • Decision traceability: how to link an automated decision back to the data inputs, model version, prompt template, and user intent that produced it, sufficient for post-incident reconstruction and compliance audit
  • Cross-system correlation: how to correlate events across multiple systems when an AI agent or automation workflow spans multiple platforms (correlation IDs, distributed tracing)
  • Attribution linkage: connecting system-level actions back to the user or process that initiated them, consistent with the identity model (Q2) and the enforcement architecture (Q3)
  • Retention and access: how long different categories of log data must be retained, who can access them, and under what conditions (regulatory retention requirements, right to erasure obligations)
  • Privacy constraints on logging: what cannot be logged (or must be pseudonymised) under data protection obligations

Out of scope:

  • Infrastructure-level observability (application performance monitoring, infrastructure metrics) beyond what is needed for governance purposes
  • Real-time alerting system design (focus is on the log data model, not the alerting infrastructure)
  • Identity model design (covered by Q2)
  • Enforcement architecture (covered by Q3)

Constraints:

  • Logging requirements must be grounded in both operational need (debugging, incident response) and regulatory obligation (audit, compliance evidence generation)
  • Must address the privacy tension: logging prompt content may capture personal data; the resolution of this tension must be evidence-based
  • This item requires Q2 (identity) and Q3 (enforcement architecture) as inputs because logging must be identity-attributed and placed at enforcement points

Context

  • [inference; source: https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-lowcode-governance-enforcement-architecture.html; https://davidamitchell.github.io/Research/research/2026-04-26-permission-safe-rag-enterprise-information-architecture.html] Without a defined observability and telemetry model, AI and low-code governance remains asserted rather than provable, because enterprises cannot reconstruct what system acted, with which permissions, on what inputs, and under which control decisions.
  • [fact; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate] The reviewed vendor stacks already split telemetry across different streams, such as trace data, payload logging, lifecycle audit logs, and runtime execution logs, which means a single log source is unlikely to satisfy governance on its own.
  • [inference; source: https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-lowcode-governance-enforcement-architecture.html] This item depends on Q2 and Q3 because telemetry only becomes governance evidence when every event can be bound to an attributable actor and to a concrete enforcement point.

Cross-references:

  • Q2: 2026-04-26-ai-agent-identity-access-management-enterprise (prerequisite)
  • Q3: 2026-04-26-ai-lowcode-governance-enforcement-architecture (prerequisite)
  • Q6: 2026-04-26-data-governance-ai-lowcode-enterprise-enforcement (prerequisite)
  • Q9: 2026-04-26-human-in-the-loop-ai-automated-workflows
  • Q15: 2026-04-26-ai-lowcode-regulatory-compliance-alignment
  • Q16: 2026-04-26-ai-agent-control-plane-architecture-enterprise

Approach

  1. Log data taxonomy: Define the categories of log data required for AI and low-code governance: interaction logs (prompts, responses, actions), decision logs (what decision was made, on what basis), attribution logs (who or what initiated the action), system state logs (model version, configuration), and exception or override logs (human overrides, circuit breaker activations, incidents).
  2. Granularity trade-off analysis: Assess the trade-offs between full-fidelity logging and summary logging for each log category, including cost, storage, privacy risk, and audit completeness. Identify which categories require full fidelity for compliance purposes and which can use summary representations.
  3. Traceability chain design: Define the data elements required to enable complete event reconstruction: what fields must be present in each log record for a reconstruction to be complete for audit and compliance purposes. Assess distributed tracing standards, including World Wide Web Consortium (W3C) Trace Context and OpenTelemetry, for applicability to AI agent action chains.
  4. Privacy constraint mapping: Identify which log data categories may contain personal data (prompt content, retrieved document content, decision rationale) and what data protection obligations apply, including retention limits, pseudonymisation requirements, and right-to-erasure conflicts with audit retention obligations.
  5. Cross-system correlation: Assess how cross-system correlation can be achieved when an AI agent or automation workflow spans multiple platforms, what correlation identifiers need to be propagated, and what tooling supports cross-platform log correlation in enterprise environments (Azure Monitor, Amazon Web Services (AWS) CloudTrail, OpenTelemetry collectors).
  6. Retention and access policy: Based on regulatory obligations (General Data Protection Regulation (GDPR) retention limits, Australian Prudential Regulation Authority (APRA) Prudential Standard (CPS) 230 audit-trail obligations, and Digital Operational Resilience Act (DORA) incident logging), define retention periods by log category and access-control requirements.
  7. Synthesis: Produce an observability and telemetry reference model, including log taxonomy, retention policy, correlation model, and privacy-constraint matrix, suitable for use as an enterprise governance artefact.

Sources

Related

Research Skill Output

(Full output from running the research skill, retained verbatim in the completed item. Sections 0 to 5 are the investigation, and section 6 seeds the Findings section below.)

§0 Initialise

  • [fact; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://www.w3.org/TR/trace-context/; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate] Research question restated: what telemetry model lets an enterprise reconstruct every material AI and low-code action by linking the initiating user or process, the acting machine component, the prompt or workflow state, the downstream tools or connectors used, the result, and the control decisions applied?
  • [fact; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata] Scope confirmed: the investigation covers AI prompt, response, retrieval, tool-call, token, and model telemetry; low-code trigger, action, run, connector, and override telemetry; cross-system correlation; retention and access; and privacy constraints on logging content.
  • [fact; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://handbook.apra.gov.au/standard/cps-230; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774] Constraints confirmed: the answer must satisfy both operational reconstruction needs and regulatory limits, especially storage limitation, data minimisation, and the need for auditable operational-risk evidence.
  • [fact; source: https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-lowcode-governance-enforcement-architecture.html; https://davidamitchell.github.io/Research/research/2026-04-26-permission-safe-rag-enterprise-information-architecture.html] Prior work cross-reference: prior completed items already established that meaningful governance depends on attributable machine identity, explicit enforcement points, and permission-safe retrieval architecture, so this item narrows the problem to the telemetry needed to prove those controls operated in practice.
  • [fact; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/purview/audit-log-retention-policies] Output format confirmed: knowledge, specifically an observability and telemetry reference model with field taxonomy, retention tiers, correlation rules, and privacy controls.

§1 Question Decomposition

  • Root question: What telemetry model makes AI and low-code governance reconstructable, attributable, privacy-bounded, and regulator-defensible?
  • A. Baseline telemetry taxonomy
    • A1. Which metadata fields are mandatory for every governed event, regardless of platform?
    • A2. Which AI-specific fields are required for model, prompt, retrieval, tool, and output reconstruction?
    • A3. Which low-code-specific fields are required for trigger, action, connector, and run reconstruction?
  • B. Granularity
    • B1. Which categories require immutable metadata for every event?
    • B2. Which categories justify full-content capture, and when must that capture be selective rather than universal?
  • C. Correlation and attribution
    • C1. Which correlation identifiers must flow across platforms and vendors?
    • C2. How should user identity, machine identity, and session lineage be preserved together?
  • D. Retention and access
    • D1. What retention rules are directly prescribed by the reviewed sources?
    • D2. Where do the sources require a purpose-based retention policy rather than a single universal period?
    • D3. Who should be able to query which log categories?
  • E. Privacy
    • E1. Which telemetry fields routinely contain personal or otherwise sensitive data?
    • E2. What do the sources imply about redaction, minimisation, and erasure handling?
  • F. Synthesis
    • F1. What reference model fits both AI and low-code systems?
    • F2. What minimum set of logs should be mandatory before an enterprise claims governed deployment?

§2 Investigation

  • A. Baseline log-management and trace-correlation requirements

  • [fact; source: https://csrc.nist.gov/pubs/sp/800/92/final] NIST SP 800-92 says log management should cover log generation, transmission, storage, analysis, and disposal, and should protect the confidentiality, integrity, and availability of logs.

  • [fact; source: https://csrc.nist.gov/pubs/sp/800/92/final] NIST SP 800-92 describes authentication logs as typically recording origin, username, success or failure, and date and time, which makes identity, outcome, and timestamp minimum baseline fields rather than optional enrichments.

  • [fact; source: https://csrc.nist.gov/pubs/sp/800/92/final] NIST SP 800-92 recommends keeping log-source clocks synchronized to a common time source, because cross-source reconstruction fails when timestamps cannot be normalized reliably.

  • [fact; source: https://www.w3.org/TR/trace-context/; https://opentelemetry.io/docs/concepts/signals/traces/] W3C Trace Context standardizes traceparent and tracestate propagation for distributed tracing, and OpenTelemetry describes traces as correlated spans with trace identifiers, parent identifiers, timestamps, attributes, events, links, and status.

  • [fact; source: https://www.w3.org/TR/trace-context/] W3C Trace Context says tracing vendors must not use traceparent or tracestate for personally identifiable or otherwise sensitive information, which means correlation headers must carry identifiers for linkage, not business payloads.

  • [inference; source: https://csrc.nist.gov/pubs/sp/800/92/final; https://www.w3.org/TR/trace-context/; https://opentelemetry.io/docs/concepts/signals/traces/] The reference model therefore needs two separate layers: immutable event metadata on every material step, and content-bearing fields held outside correlation headers and governed by stricter access and retention controls.

  • B. AI-specific telemetry requirements

  • [fact; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/] OpenTelemetry's generative AI attribute registry includes fields for conversation ID, agent ID, input messages, output messages, prompt name, request and response model, retrieved documents, retrieval query text, tool definitions, tool-call arguments, tool-call results, finish reasons, and input and output token counts.

  • [fact; source: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/] OpenTelemetry generative AI events state that user inputs, system instructions, tool definitions, and output messages are opt-in event attributes rather than mandatory always-on capture.

  • [fact; source: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-metrics/] OpenTelemetry defines gen_ai.client.token.usage as the recommended metric for input and output token consumption when counts are readily available, and requires billable tokens when both used and billable values are exposed.

  • [fact; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://learn.microsoft.com/en-us/azure/foundry/observability/concepts/trace-agent-concept] Microsoft Foundry tracing captures latency, exceptions, prompt content, retrieval operations, inputs, outputs, tool usage, retries, token consumption, and cost signals, and stores traces in Application Insights using OpenTelemetry semantic conventions.

  • [fact; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup] Microsoft Foundry warns that tracing can capture sensitive information such as user inputs, model outputs, tool arguments, and tool results, and recommends redacting personal data and applying production-grade access controls and retention policies.

  • [fact; source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock/latest/APIReference/API_PutModelInvocationLoggingConfiguration.html] Amazon Bedrock model invocation logging can capture full request data, response data, and metadata for supported model-runtime calls, and its configuration can separately enable logging for text, image, embedding, audio, and video modalities to CloudWatch Logs or Amazon S3.

  • [fact; source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] Bedrock states that model invocation logging is disabled by default, supports only bedrock-runtime endpoint calls, and logs until the logging configuration is deleted, which shows that vendors expose full-content capture as a configurable control surface rather than as a universal default.

  • [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] A governed AI event therefore needs, at minimum, trace and conversation identifiers, prompt or template identity, model requested and model served, retrieval set identifiers, tool calls and results, token counts, latency, status, and policy outcome; full prompt and response bodies should be selectively captured when risk, incident handling, evaluation, or legal evidence needs justify it.

  • C. Low-code telemetry requirements

  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-overview; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate] Microsoft Purview audit for Power Platform covers who did what, settings and configuration changes, environment details, lifecycle events, and permission changes for Power Automate flows.

  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate] Microsoft explicitly says Purview logs capture flow lifecycle events and permission changes, but do not include individual flow runs, action executions, or connector calls at runtime.

  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] Power Automate telemetry in Application Insights maps cloud flow runs to the Requests table and triggers or actions to the Dependencies table, enabling monitoring of runs, trigger execution, action execution, and failures.

  • [fact; source: https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata] Dataverse FlowRun records include start time, end time, run duration, status, trigger type, error code, error message, owner, workflow name, workflow ID, parent run ID, and time-to-live for each flow execution.

  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors; https://learn.microsoft.com/en-us/azure/azure-monitor/reference/tables/powerplatformconnectoractivity] Power Platform connector logs capture API creation and deletion, connection events, permission changes, client Internet Protocol (IP) address, user identity, operation, request correlation ID, resource ID, result status, and additional environment metadata, and Azure Monitor exposes a dedicated PowerPlatformConnectorActivity table for those audit events.

  • [inference; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors] Low-code governance cannot rely on one log stream: it needs one administrative audit layer for maker and permission changes, one runtime layer for flow runs and action failures, and one connector layer for cross-system calls and correlation.

  • D. Attribution and cross-system lineage

  • [fact; source: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html] AWS CloudTrail records actions taken by a user, role, or AWS service and preserves who or what took which action, what resources were acted upon, and when the event occurred.

  • [fact; source: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html] CloudTrail's userIdentity element shows the identity type, how temporary credentials were obtained, and the issuer context for assumed-role sessions.

  • [fact; source: https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html] AWS source identity can be required when a role is assumed, is recorded in CloudTrail, persists across role chaining, and cannot be changed during the role session.

  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors; https://learn.microsoft.com/en-us/azure/azure-monitor/reference/tables/powerplatformconnectoractivity] Power Platform connector logs expose both actor identity and request correlation identifiers, which provides a pattern for linking user or service actors to outbound connector activity.

  • [fact; source: https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-lowcode-governance-enforcement-architecture.html] Prior completed repository work already established that enterprise AI governance needs distinct machine identities and concrete enforcement points, so telemetry must bind initiator identity, acting workload identity, and enforcement outcome in the same traceable chain.

  • [inference; source: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html; https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors; https://www.w3.org/TR/trace-context/] End-to-end attribution requires four linked identifiers on each material event: the initiating human or process identity, the acting machine identity, the distributed trace identifier, and the local event or request identifier for the specific platform.

  • E. Privacy, retention, and access controls

  • [fact; source: https://gdpr-info.eu/art-5-gdpr/] GDPR Article 5 requires personal data to be adequate, relevant, and limited to what is necessary, kept no longer than necessary for the processing purpose, and protected with appropriate integrity and confidentiality controls.

  • [fact; source: https://gdpr-info.eu/art-17-gdpr/] GDPR Article 17 gives data subjects a right to erasure, but preserves exceptions when processing is necessary for compliance with a legal obligation or for the establishment, exercise, or defence of legal claims.

  • [fact; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/] Microsoft Foundry and OpenTelemetry both treat full prompt, response, and tool payload capture as sensitive enough to require redaction or opt-in handling rather than unconditional default storage.

  • [fact; source: https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata] Dataverse stores cloud flow run history for 28 days by default and lets administrators shorten, disable, or customize time-to-live values, showing that runtime telemetry retention is an explicit governance setting.

  • [fact; source: https://learn.microsoft.com/en-us/purview/audit-log-retention-policies] Microsoft Purview defaults to 180-day retention for audit records outside the specified premium workloads, one-year retention for specified premium workloads, and supports custom retention policies up to 10 years.

  • [fact; source: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html] CloudTrail Lake supports retention up to 3,653 days, about 10 years, for event data stores, while event history exposes 90 days of management events by default.

  • [fact; source: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32022R2554] The reviewed DORA texts require logging procedures to define which events are logged, the retention period, and the measures used to secure and handle log data, and require financial entities to record and track information and communication technology related incidents consistently.

  • [fact; source: https://handbook.apra.gov.au/standard/cps-230; https://handbook.apra.gov.au/ppg/cpg-230] APRA CPS 230 and CPG 230 require effective internal controls, monitoring, analysis, reporting, escalation, and resilience evidence, but the reviewed texts do not prescribe one universal retention period for all operational logs.

  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://learn.microsoft.com/en-us/purview/audit-log-retention-policies; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774] The defensible retention model is category-based, not universal: short-lived high-sensitivity content logs, medium-lived operational traces, and longer-lived audit or incident records retained according to legal and control objectives.

  • F. Resulting design constraints

  • [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://www.w3.org/TR/trace-context/; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors] A single enterprise observability model must combine three coordinated layers: administrative audit events, runtime execution traces, and cross-system connector or tool telemetry, all joined by a portable trace and event-correlation scheme.

  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup] The privacy tension is not resolved by turning content logging off entirely; it is resolved by always capturing reconstructive metadata, then applying selective content capture, redaction, tighter access, and shorter retention where full payloads are needed.

§3 Reasoning

  • [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] The reviewed sources converge on a split between metadata that should always exist and payload content that should be selectively captured, because both AI and low-code platforms expose reliable operational metadata more broadly than they expose safe full-content logging.
  • [inference; source: https://www.w3.org/TR/trace-context/; https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html] Cross-system reconstruction fails unless correlation identifiers and attribution lineage travel together, so trace IDs alone are insufficient without initiator and actor identity fields.
  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://learn.microsoft.com/en-us/purview/audit-log-retention-policies; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html] Retention has to be driven by log purpose and legal basis, because the sources show materially different storage windows for different telemetry categories and no reviewed authority mandates one universal retention term.

§4 Consistency Check

  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] There is no contradiction between Power Automate Purview audit logs and Application Insights telemetry: Microsoft documents that Purview covers lifecycle and permission events, while Application Insights covers runtime runs, triggers, and actions.
  • [fact; source: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] There is no contradiction between OpenTelemetry's opt-in content events and vendor payload logging: all reviewed sources treat full-content capture as configurable and sensitive, not as an unconditional baseline.
  • [fact; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774; https://handbook.apra.gov.au/standard/cps-230] The privacy and audit obligations can coexist because GDPR permits retention where a legal obligation or legal-claims basis applies, while APRA and DORA reviewed texts require evidence, monitoring, and incident recording rather than indiscriminate logging of all content forever.

§5 Depth and Breadth Expansion

  • [fact; source: https://learn.microsoft.com/en-us/azure/foundry/concepts/observability; https://learn.microsoft.com/en-us/azure/azure-monitor/overview] Technical lens: Microsoft now treats evaluation, monitoring, and tracing as a combined observability stack, which means governance telemetry should feed both incident reconstruction and ongoing quality and safety assessment rather than only compliance reporting.
  • [fact; source: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32022R2554; https://handbook.apra.gov.au/ppg/cpg-230] Regulatory lens: the reviewed financial-sector standards emphasize evidence, monitoring, and incident logging, but they leave implementation detail to the entity, which increases the importance of an explicit internal telemetry taxonomy and retention matrix.
  • [inference; source: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html; https://learn.microsoft.com/en-us/purview/audit-log-retention-policies; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata] Economic lens: high-fidelity content logging is materially more expensive than metadata logging because platforms expose very different default retention and storage patterns, so indiscriminate full-payload capture scales poorly and creates avoidable storage and privacy cost.
  • [inference; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://www.w3.org/TR/trace-context/; https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html] Behavioral lens: engineers will only preserve attribution and correlation if the required identifiers are automatically injected and propagated, so the governance model should enforce headers, actor fields, and trace linkage in platform adapters rather than relying on manual developer discipline.

§6 Synthesis

(This section seeds the Findings below.)

Executive summary:

  • [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://www.w3.org/TR/trace-context/; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] Enterprises need a three-layer observability model for governed AI and low-code systems: always-on reconstructive metadata, cross-system distributed tracing, and selective full-content capture for high-risk, sampled, or incident-driven cases.
  • [fact; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] For AI workloads, the mandatory baseline includes prompt or template identity, model requested and model served, retrieval context identifiers, tool calls and results, output status, token usage, latency, and exceptions; full prompts and responses should be captured only under explicit policy because the reviewed standards and vendor guidance treat them as sensitive.
  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata] For low-code systems, governance requires separate administrative audit logs, runtime execution traces, and connector-call telemetry because no single Microsoft log stream captures maker actions, run behavior, and external API activity together.
  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774; https://handbook.apra.gov.au/standard/cps-230] Retention and access should be category-based and purpose-bound, because regulators reviewed here require secure, accessible audit evidence and incident recording, while privacy law limits how long personally revealing content can be kept without a clear legal basis.

Key findings:

    1. [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://www.w3.org/TR/trace-context/; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://csrc.nist.gov/pubs/sp/800/92/final] High confidence: A governable AI and low-code estate needs a dual telemetry baseline of reconstructive metadata for every material event and portable trace correlation across systems, because reconstruction is impossible when either the event details or the linkage chain is missing.
    1. [fact; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/azure/foundry/observability/concepts/trace-agent-concept; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] High confidence: Every governed AI event should record conversation or request identity, agent identity, prompt or template identity, model requested and model served, retrieval set identifiers, tool definitions, tool-call arguments and results, finish status, token usage, latency, and exceptions before full prompt or response bodies are considered.
    1. [inference; source: https://www.w3.org/TR/trace-context/; https://opentelemetry.io/docs/concepts/signals/traces/] Medium confidence: W3C Trace Context and OpenTelemetry provide the strongest vendor-neutral baseline among the reviewed options for cross-system correlation, because they standardize trace identifiers, parent-child span relationships, timestamps, attributes, and events while explicitly prohibiting sensitive payload data in trace headers.
    1. [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors] High confidence: Low-code governance requires three distinct log streams, administrative audit, runtime execution, and connector activity, because Microsoft documents that Purview alone does not capture individual runs, actions, or connector calls at runtime.
    1. [inference; source: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html; https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html] Medium confidence: Enterprise attribution should bind initiator identity, acting machine identity, and session lineage on every material event, because assumed-role chains and agent execution obscure accountability unless the original actor and the workload identity are both preserved.
    1. [fact; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup] High confidence: Full prompt, response, and tool-payload logging should be opt-in, redacted where feasible, and access-restricted, because the reviewed privacy and vendor sources treat those payloads as potentially personal or sensitive data rather than harmless diagnostics.
    1. [inference; source: https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata; https://learn.microsoft.com/en-us/purview/audit-log-retention-policies; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774] High confidence: Retention should be tiered by log purpose instead of standardized into one period, because the reviewed platforms and regulations expose materially different windows for runtime telemetry, audit evidence, and incident records while requiring that each period be justified and secured.
    1. [fact; source: https://csrc.nist.gov/pubs/sp/800/92/final; https://handbook.apra.gov.au/ppg/cpg-230; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32022R2554] Medium confidence: Governance telemetry must be centrally queryable, time-synchronized, tamper-resistant, and reviewable by control functions, because the reviewed standards emphasize synchronized logs, effective monitoring, incident recording, and evidence for control review rather than raw data accumulation alone.

Evidence map:

Assumptions:

  • None.

Analysis:

  • [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] The strongest pattern across sources is structural rather than vendor-specific: AI and low-code platforms both separate administrative audit from execution telemetry, so the enterprise model should formalize that separation instead of expecting one platform log to answer every governance question.
  • [inference; source: https://www.w3.org/TR/trace-context/; https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html] Cross-system correlation and identity attribution solve different problems and must be linked, not conflated, because a trace without actor lineage cannot prove accountability, while actor lineage without trace continuity cannot reconstruct a multi-step execution path.
  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup] The privacy trade-off does not support an all-or-nothing answer on prompt and response logging; the reviewed evidence supports a layered policy in which metadata is mandatory, payload capture is explicit and justified, and erasure or legal-hold decisions are handled per retention class.
  • [inference; source: https://learn.microsoft.com/en-us/purview/audit-log-retention-policies; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata; https://handbook.apra.gov.au/standard/cps-230] Because regulators reviewed here emphasize accessible evidence and effective monitoring rather than one fixed duration, the retention matrix should be anchored to business purpose, regulatory basis, and sensitivity of data, then enforced by platform-specific storage policies.

Risks, gaps, uncertainties:

  • [fact; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://learn.microsoft.com/en-us/azure/foundry/observability/concepts/trace-agent-concept] Microsoft Foundry tracing is generally available only for prompt agents, while workflow, hosted, and custom agents remain in preview, so vendor-native observability coverage is still uneven for some AI execution patterns.
  • [fact; source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] Bedrock invocation logging excludes calls made through some endpoints, such as the Responses API on bedrock-mantle, so an enterprise reference model still needs compensating telemetry outside the vendor-native logging feature.
  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] Power Automate telemetry in Application Insights is not fully lossless according to Microsoft, so authoritative forensic reconstruction should not rely on that stream alone when stronger transactional records exist elsewhere.
  • [inference; source: https://handbook.apra.gov.au/standard/cps-230; https://handbook.apra.gov.au/ppg/cpg-230; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774] The reviewed APRA and DORA materials are principle-based rather than schema-level, so the exact field taxonomy still needs internal policy codification even though the direction of travel is clear.

Open questions:

  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/] Which log classes in the target enterprise can rely on legal-obligation or legal-claims bases strongly enough to justify retaining identifiable prompt or response content beyond short operational windows?
  • [inference; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] How much AI telemetry can be standardized entirely through OpenTelemetry adapters versus requiring vendor-specific side channels for payload capture and retention controls?
  • [inference; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors; https://learn.microsoft.com/en-us/azure/azure-monitor/reference/tables/powerplatformconnectoractivity] Which non-Microsoft low-code platforms expose equivalent connector-level correlation identifiers and runtime schemas, and where will compensating instrumentation be required?

§7 Recursive Review

  • [fact; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://www.w3.org/TR/trace-context/; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/] Every factual or inferential claim in sections 0 through 6 is either explicitly sourced or marked as an inference derived from sourced platform, standards, or legal material.
  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] The key internal contradiction candidate, whether one low-code audit stream could suffice, was resolved by Microsoft's own split between Purview lifecycle logs and Application Insights or Dataverse runtime telemetry.
  • [fact; source: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup] The key privacy contradiction candidate, whether full prompt or response capture must be default for auditability, was resolved in favor of selective payload capture because both reviewed telemetry standards and vendor guidance treat content logging as opt-in and sensitive.

Findings

Executive Summary

  • [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://www.w3.org/TR/trace-context/; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] Enterprises need a three-layer observability model for governed AI and low-code systems: always-on reconstructive metadata, cross-system distributed tracing, and selective full-content capture for high-risk, sampled, or incident-driven cases.
  • [fact; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] The AI baseline should always record prompt or template identity, model requested and model served, retrieval context identifiers, tool calls and results, token usage, timing, and status, while leaving full prompt and response bodies behind an explicit higher-sensitivity policy gate.
  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata] Low-code governance needs separate administrative audit logs, runtime execution traces, and connector-call telemetry because no single reviewed log stream captures maker actions, flow behavior, and downstream API activity together.
  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774; https://handbook.apra.gov.au/standard/cps-230] Retention and access should be category-based and purpose-bound because the reviewed regulators require secure, accessible evidence and incident records, while privacy law prohibits keeping identifying content longer than justified.

Key Findings

  1. [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://www.w3.org/TR/trace-context/; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://csrc.nist.gov/pubs/sp/800/92/final] High confidence: A governable AI and low-code estate needs reconstructive metadata for every material event and portable trace correlation across systems, because post-incident reconstruction fails when event detail exists without linkage or linkage exists without event detail.
  2. [fact; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/azure/foundry/observability/concepts/trace-agent-concept; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] High confidence: Every governed AI event should record conversation or request identity, agent identity, prompt or template identity, model requested and model served, retrieval set identifiers, tool definitions, tool-call arguments and results, finish status, token usage, latency, and exceptions before full prompt or response bodies are considered.
  3. [inference; source: https://www.w3.org/TR/trace-context/; https://opentelemetry.io/docs/concepts/signals/traces/] Medium confidence: W3C Trace Context and OpenTelemetry provide the strongest vendor-neutral baseline among the reviewed options for cross-system correlation because they standardize trace identifiers, parent-child span relationships, timestamps, attributes, and events while explicitly prohibiting sensitive payload data in trace headers.
  4. [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors] High confidence: Low-code governance requires distinct administrative, runtime, and connector telemetry streams because Microsoft documents that Purview alone does not capture individual runs, action executions, or connector calls at runtime.
  5. [inference; source: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html; https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html] Medium confidence: Enterprise attribution should bind initiator identity, acting machine identity, and session lineage on every material event because assumed-role chains and agent execution obscure accountability unless the original actor and workload identity are both preserved.
  6. [fact; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup] High confidence: Full prompt, response, and tool-payload logging should be opt-in, redacted where feasible, and access-restricted because the reviewed privacy and vendor sources treat those payloads as potentially personal or otherwise sensitive data rather than harmless diagnostics.
  7. [inference; source: https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata; https://learn.microsoft.com/en-us/purview/audit-log-retention-policies; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774] High confidence: Retention should be tiered by log purpose instead of standardized into one period because the reviewed platforms and regulations expose materially different windows for runtime telemetry, audit evidence, and incident records while requiring each period to be justified and secured.
  8. [fact; source: https://csrc.nist.gov/pubs/sp/800/92/final; https://handbook.apra.gov.au/ppg/cpg-230; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32022R2554] Medium confidence: Governance telemetry must be centrally queryable, time-synchronized, tamper-resistant, and reviewable by control functions because the reviewed standards emphasize synchronized logs, effective monitoring, incident recording, and evidence for control review rather than raw data accumulation alone.

Evidence Map

Claim Source Confidence Notes
[inference] A governable estate needs reconstructive metadata plus cross-system trace correlation. https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/ ; https://www.w3.org/TR/trace-context/ ; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow ; https://csrc.nist.gov/pubs/sp/800/92/final high Metadata and trace linkage are complementary, not interchangeable, and NIST reinforces the need for log integrity and usable reconstruction.
[fact] AI telemetry must include prompt or template identity, model details, retrieval identifiers, tool activity, output status, token usage, latency, and exceptions. https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/ ; https://learn.microsoft.com/en-us/azure/foundry/observability/concepts/trace-agent-concept ; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html high These sources expose the concrete field families required for reconstruction.
[inference] W3C Trace Context plus OpenTelemetry provide the strongest vendor-neutral baseline among the reviewed options for correlation. https://www.w3.org/TR/trace-context/ ; https://opentelemetry.io/docs/concepts/signals/traces/ medium Trace headers carry linkage only; trace spans carry contextual metadata, but alternative vendor-specific schemes remain possible.
[fact] Low-code governance needs separate administrative, runtime, and connector telemetry. https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate ; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow ; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors ; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata high Purview covers lifecycle and permissions; other stores cover execution and external calls.
[inference] Enterprise attribution should preserve initiator identity, acting machine identity, and session lineage. https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html ; https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html ; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html medium Role-session and machine-identity lineage support accountable automation, but the universal enterprise requirement is a synthesis.
[fact] Full prompt and response capture should be opt-in, redacted, and tightly access-controlled. https://gdpr-info.eu/art-5-gdpr/ ; https://gdpr-info.eu/art-17-gdpr/ ; https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/ ; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup high Payload content is the highest-sensitivity telemetry class in the reviewed evidence.
[inference] Retention should be tiered by purpose rather than forced into one enterprise-wide period. https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata ; https://learn.microsoft.com/en-us/purview/audit-log-retention-policies ; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html ; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774 high The reviewed sources expose 28-day, 90-day, 180-day, one-year, and multi-year patterns, so the policy conclusion is a synthesis.
[fact] Governance telemetry must be synchronized, tamper-resistant, centrally queryable, and reviewable. https://csrc.nist.gov/pubs/sp/800/92/final ; https://handbook.apra.gov.au/ppg/cpg-230 ; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32022R2554 medium The regulatory texts are principle-based, but they consistently demand evidence, monitoring, and incident handling.

Assumptions

  • None.

Analysis

  • [inference; source: https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/; https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-power-automate; https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] The strongest pattern across sources is structural rather than vendor-specific: AI and low-code platforms both separate administrative audit from execution telemetry, so the enterprise model should formalize that separation instead of expecting one platform log to answer every governance question.
  • [inference; source: https://www.w3.org/TR/trace-context/; https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_monitor.html; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html] Cross-system correlation and identity attribution solve different problems and must be linked, not conflated, because a trace without actor lineage cannot prove accountability, while actor lineage without trace continuity cannot reconstruct a multi-step execution path.
  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/; https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/; https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup] The privacy trade-off does not support an all-or-nothing answer on prompt and response logging; the reviewed evidence supports a layered policy in which metadata is mandatory, payload capture is explicit and justified, and erasure or legal-hold decisions are handled per retention class.
  • [inference; source: https://learn.microsoft.com/en-us/purview/audit-log-retention-policies; https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-user-guide.html; https://learn.microsoft.com/en-us/power-automate/dataverse/cloud-flow-run-metadata; https://handbook.apra.gov.au/standard/cps-230] Because regulators reviewed here emphasize accessible evidence and effective monitoring rather than one fixed duration, the retention matrix should be anchored to business purpose, regulatory basis, and sensitivity of data, then enforced by platform-specific storage policies.

Risks, Gaps, and Uncertainties

  • [fact; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://learn.microsoft.com/en-us/azure/foundry/observability/concepts/trace-agent-concept] Microsoft Foundry tracing is generally available only for prompt agents, while workflow, hosted, and custom agents remain in preview, so vendor-native observability coverage is still uneven for some AI execution patterns.
  • [fact; source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] Bedrock invocation logging excludes calls made through some endpoints, such as the Responses API on bedrock-mantle, so an enterprise reference model still needs compensating telemetry outside the vendor-native logging feature.
  • [fact; source: https://learn.microsoft.com/en-us/power-platform/admin/app-insights-cloud-flow] Power Automate telemetry in Application Insights is not fully lossless according to Microsoft, so authoritative forensic reconstruction should not rely on that stream alone when stronger transactional records exist elsewhere.
  • [inference; source: https://handbook.apra.gov.au/standard/cps-230; https://handbook.apra.gov.au/ppg/cpg-230; https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1774] The reviewed APRA and DORA materials are principle-based rather than schema-level, so the exact field taxonomy still needs internal policy codification even though the direction of travel is clear.

Open Questions

  • [inference; source: https://gdpr-info.eu/art-5-gdpr/; https://gdpr-info.eu/art-17-gdpr/] Which log classes in the target enterprise can rely on legal-obligation or legal-claims bases strongly enough to justify retaining identifiable prompt or response content beyond short operational windows?
  • [inference; source: https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/trace-agent-setup; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] How much AI telemetry can be standardized entirely through OpenTelemetry adapters versus requiring vendor-specific side channels for payload capture and retention controls?
  • [inference; source: https://learn.microsoft.com/en-us/power-platform/admin/activity-logging-auditing/activity-logs-connectors; https://learn.microsoft.com/en-us/azure/azure-monitor/reference/tables/powerplatformconnectoractivity] Which non-Microsoft low-code platforms expose equivalent connector-level correlation identifiers and runtime schemas, and where will compensating instrumentation be required?

Output

Navigation

Home

By Tag

bureaucracy

change-management

coase

constraint-analysis

control-model

decision-rights

delegation

delivery-risk

demand-segmentation

enterprise

exception-handling

execution

flow

flow-design

flow-metrics

governance

governance-patterns

incentives

instability

institutional-economics

leading-indicators

operating-model

organisation

organisational-design

queue-design

queueing

regulated-enterprise

routing

throughput

throughput-risk

transaction-costs

triage

williamson

Clone this wiki locally