タスクマイニング／デスクトップアクティビティマイニング → AI化

Process Miningが業務システムのevent logから工程を観測するのに対し、Task Miningはdesktop上の人間操作を記録・分析し、個々のtaskの実行実態、variation、manual friction、automation candidateを発見できる。一方、screen/keyboard/click等の収集はemployee privacy、secret/PII capture、sampling bias、surveillance misuseを直接生むため、AI/RPA候補抽出の前にcapture governanceをCoreとして定義する必要がある。

Enterprise AI Core / organization overlay / governance / implementation / benchmark-evaluation / non-LLM component / interoperability with Process Mining and RPA.

• Task MiningはProcess Miningと同一ではなく、desktop/user-interaction粒度を補完する観測層として扱うべきか。
• 観測された高頻度・高時間taskを、そのままAI/RPA自動化候補とみなしてよいか。
• capture対象、consent/notice、masking/exclusion、retention、access、deletion、analysis provenanceをどこまで監査可能にすべきか。

• Official product documentation | “Overview of process mining and task mining in Power Automate” | Microsoft | https://learn.microsoft.com/en-us/power-automate/process-advisor-overview | accessed 2026-09-16 | current web documentation; licensing/capacity are PRODUCT_VOLATILE | locator: Task mining / When to use task mining / Licensing | VERIFIED_SPEC + PRODUCT_VOLATILE | support/context. Microsoft states task mining records and analyzes desktop user actions, supports discovery of common actions/errors and automation opportunities, and distinguishes it from process mining.
• Official product documentation | “Overview of task mining” | Microsoft | https://learn.microsoft.com/en-us/power-automate/task-mining-overview | updated 2025-12-05; accessed 2026-09-16 | current web documentation | locator: Components; process map/application analytics/automation recommendations | VERIFIED_SPEC | support/context.
• Official tutorial | “Get started with the task mining capability” | Microsoft | https://learn.microsoft.com/en-us/power-automate/task-mining-tutorial | accessed 2026-09-16 | current web documentation | locator: Analyze a process / application report / Identify automation opportunities | VERIFIED_SPEC | support/limit. Product can surface variants, time and candidate activities, but recommendation is not proof of business correctness or AI suitability.
• Official product/privacy documentation | “Task Mining - Protection of personal data and privacy of the users” | UiPath | https://docs.uipath.com/task-mining/automation-cloud/latest/user-guide/protection-of-personal-data-and-privacy-of-the-users | accessed 2026-09-16 | Automation Cloud latest docs | locator: opening privacy/data-subject section | VERIFIED_SPEC + PRODUCT_VOLATILE | support/limit. UiPath documents capture of screenshots, keyboard actions, clicks and movements when enabled and frames recorded employees as data subjects.
• Official security documentation | “Task Mining - Security” | UiPath | https://docs.uipath.com/task-mining/automation-cloud/latest/user-guide/security | accessed 2026-09-16 | Automation Cloud latest docs | locator: Secrets management / Security assurances | VERIFIED_SPEC + PRODUCT_VOLATILE | context/limit. SaaS/security controls do not by themselves establish tenant-specific lawful/appropriate workforce monitoring.

• Task Mining and Process Mining should be separated by observation grain. Process Mining primarily reconstructs business-process flow from event logs; Task Mining observes desktop-level execution of particular work. They can be joined, but neither substitutes for the other.
• observed frequent/slow/repetitive action ≠ automation-ready task ≠ AI-suitable task. Candidate ranking must add business rule stability, exception rate, authorization, data sensitivity, deterministic alternative, failure consequence, human judgment requirement and expected value.
• Desktop recording creates a materially different privacy/security surface from ordinary event-log analysis. Screenshots and keyboard actions can incidentally capture credentials, customer data, health/HR/legal content, unrelated applications and personal communications. Capture minimization and exclusion are therefore preconditions, not post-processing niceties.
• Worker behavior captured during a study can be non-representative because of sampling window, participant selection, Hawthorne effect, application/version differences, role differences and rare exceptions. Frequency/time summaries must retain cohort and capture-window provenance.
• AI should not be used to infer employee intent, productivity, misconduct or performance from interaction traces unless an independently governed use case and validation supports that inference. Task Mining evidence establishes observed interaction patterns, not latent motive or employee quality.

• Recording coverage can omit offline work, phone calls, meetings, tacit decisions, second devices and applications excluded from capture.
• UI events may be technically captured but semantically mis-grouped into activities; variant maps are model outputs dependent on labeling/grouping rules.
• High-frequency actions may be symptoms of upstream policy/data defects; automating them can preserve or accelerate a bad process.
• Rare but high-risk exceptions can disappear under dominant-path optimization.
• Screens/keystrokes can capture secrets/PII even when the target business task does not require them.
• Vendor privacy/security documentation is evidence of product mechanisms and stated controls, not proof of legal compliance or tenant-specific employee notice/consent.
• Automation recommendation/connectors exposed by a product are implementation suggestions, not independent evidence that the recommended automation is safe or economically justified.

Microsoft Power Automate Task Mining availability depends on tenant/environment, licensing/capacity and administrative configuration; these are PRODUCT_VOLATILE and must be re-pinned at publication/update time. UiPath Task Mining availability likewise depends on deployment/product edition, tenant configuration, region and current release. Public docs do not prove a specific enterprise tenant has recording enabled or permitted.

PROPOSAL: TASK_MINING_CAPTURE_AND_AUTOMATION_CONTRACT with: study_id; business_process_ref; task_scope; participant/cohort definition; notice/consent/legal-basis record where applicable; device/app allowlist and denylist; capture modality; screenshot/keyboard policy; masking/redaction rules; capture window/timezone; collector/product/version; raw artifact hash/location; retention/deletion policy; analyst access; activity-label/grouping rules/version; variants/time/error metrics; automation-candidate rationale; deterministic/RPA/AI alternatives; risk/authorization gate; pilot design; human-review requirement; before/after metrics; approval; rollback/stop trigger.

SAFE TEST_PROFILE (not executed): synthetic desktop workflow with normal path, shortcut path, rework, rare exception, sensitive-window interruption, credential-field event, copy/paste, app switching and idle/wait periods. Compare raw action trace to expected semantic task segmentation. Measure segmentation precision/recall, variant recovery, sensitive-capture rate, redaction recall/false-positive rate, automation-candidate precision, rare-exception preservation and analyst reproducibility. Do not use real employee surveillance data for an initial reproducibility test.

Privacy: data minimization, purpose limitation, app/field exclusion, masking, retention and deletion must be explicit. Security: raw screenshots/interaction traces are sensitive telemetry and require least-privilege access, encryption and export controls. Audit: retain collector/version, grouping rules, cohort/window and transformations so findings can be reproduced. Cost: include recorder deployment/support, storage, analyst review, privacy/legal review and automation implementation—not only software license. Operability: recorder health and missing-capture rate must be monitored or the analysis can silently become biased.

“Task Mining can complement process-level event-log analysis by observing how selected desktop tasks are actually executed and by surfacing variants, friction and possible automation candidates. Its recordings may contain highly sensitive workforce and business data, and observed frequency or duration alone does not establish that a task should be automated or delegated to AI. Capture governance, representativeness, exception analysis and independent automation-risk evaluation are required.”

Connect EA-096 Process Mining and EA-097 Task Mining as two observation grains. Future update Work should decide whether PROCESS_TO_AI_CANDIDATE_CONTRACT becomes a parent contract with this task-level capture contract as a child. Re-pin current Microsoft/UiPath plan, region, capacity, retention, recording controls and product version before publication. If commercial cross-vendor comparison is added, collect each vendor’s exact capture modalities, exclusion/masking controls, retention/deletion, admin roles and export boundaries rather than comparing marketing “automation opportunity” claims.

/ STOP_CONDITION

exact current Microsoft Task Mining plan/capacity/region matrix; exact UiPath current release/region/retention matrix; cross-vendor measurement comparability; tenant-specific legal/works-council/employee-monitoring requirements; empirical false-positive/false-negative rates of vendor automation recommendations. STOP_CONDITION for this packet: Core boundary, primary product evidence, privacy failure envelope, availability gates, safe wording and reproducible non-employee TEST_PROFILE are established; remaining product/tenant specifics are explicitly volatile or tenant-dependent.

RUN_SCOPE: EA-407..EA-410