AI連携・自動化ツール大全

WHY_THIS_MATTERS_TO_UPDATE

Enterprise AIの実運用では、LLM・agent本体よりも、業務SaaS・基幹系・ファイル・メール・CRM・RPA・Webhook等へ接続する自動化層が実際の権限、credential、side effect、再実行、監査、障害復旧の境界を決める。製品名や「連携数」の一覧だけでは事故条件を説明できないため、Zapier / Make / n8n / Power Automate・Copilot Studio / Google Workspace・Apps Script / UiPath等を共通運用軸で比較できる辞典層が必要。

SCOPE_ROUTING

interoperability / implementation / organization overlay / governance / security / audit / operability / cost / non-LLM automation component / agent-tool boundary

CLAIM_OR_RESEARCH_QUESTION

1. AI連携・自動化製品を、ブランドではなく共通の運用面で比較できるか。 2. connector / plugin / app / workflow / webhook / RPA / agent actionはどこまで同じ問題として扱え、どこから分離すべきか。 3. 「接続できる」ことと「安全に本番運用できる」ことの間に必要なcontrol planeは何か。 4. retry / replay / resubmitがside effectを二重実行する危険をどのように扱うべきか。

EVIDENCE TABLE

E007-S01 — Microsoft Power Platform data policy

- SOURCE_CLASS: official product documentation - TITLE: Implement a data policy strategy / Configure data policies for agents - AUTHOR/ORG: Microsoft - URL: https://learn.microsoft.com/en-us/power-platform/guidance/adoption/dlp-strategy ; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention - ACCESSED: 2026-09-11 - VERSION/STATE: current web docs; Copilot Studio tenant-wide enforcement current as of 2026 - LOCATOR: Business / Non-business / Blocked connector groups; real-time enforcement; connectors in different data groups cannot share data - LABEL: VERIFIED_SPEC + PRODUCT_VOLATILE - RELATION: support / limit - ESTABLISHES: connector利用可否とデータ越境は組織policyで制御可能であり、agent/flowの存在だけでは実行可能性が決まらない。 - DOES_NOT_ESTABLISH: 特定tenantのpolicy実設定、個別connectorの許可状態。

E007-S02 — Power Automate limits and run operations

- SOURCE_CLASS: official product documentation - TITLE: Limits for automated, scheduled, and instant flows / Cancel or resubmit flow runs in bulk - AUTHOR/ORG: Microsoft - URL: https://learn.microsoft.com/en-us/power-automate/limits-and-config ; https://learn.microsoft.com/en-us/power-automate/how-tos-bulk-resubmit - ACCESSED: 2026-09-11 - VERSION/STATE: current 2026 docs - LOCATOR: throughput limits; persistent 14-day over-limit may turn flow off; resubmit subject to connector/API-call limits; run history is 28 days in the bulk-run UI description - LABEL: VERIFIED_SPEC + PRODUCT_VOLATILE - RELATION: support / failure-envelope - ESTABLISHES: automation capacity is bounded by platform and connector throttling; operational recovery includes cancel/resubmit but resubmit is not equivalent to transaction rollback. - DOES_NOT_ESTABLISH: every connector’s retry/idempotency semantics.

E007-S03 — n8n security audit and executions

- SOURCE_CLASS: author-maintained product documentation / OSS practice - TITLE: Security audit / All executions / Source control and environments - AUTHOR/ORG: n8n - URL: https://docs.n8n.io/hosting/securing/security-audit/ ; https://docs.n8n.io/workflows/executions/all-executions/ ; https://docs.n8n.io/source-control-environments/create-environments/ - ACCESSED: 2026-09-11 - VERSION/STATE: n8n current docs; source control available Business/Enterprise; external storage doc references v2.6.4+ - LOCATOR: audit detects risky nodes, community/custom nodes, unprotected webhooks and credential/database risks; failed executions can be retried with current or original workflow; Git-backed multi-instance environments - LABEL: VERIFIED_SPEC + OSS_PRACTICE + PRODUCT_VOLATILE - RELATION: support / counter - ESTABLISHES: self-hostable workflow engines shift part of security/upgrade/storage responsibility to operator; execution replay and environment promotion are first-class concerns. - DOES_NOT_ESTABLISH: self-hosting is inherently safer; security audit is a full compliance audit.

E007-S04 — Make audit / credentials / SSO

- SOURCE_CLASS: official product documentation - TITLE: Audit logs / Credential requests / Okta SAML - AUTHOR/ORG: Make - URL: https://help.make.com/audit-logs ; https://help.make.com/credential-requests ; https://help.make.com/okta-saml - ACCESSED: 2026-09-11 - VERSION/STATE: current docs; audit logs Enterprise; credential requests restricted to eligible Partner/Enterprise users; SAML Enterprise - LOCATOR: audit events cover scenarios, connections, webhooks, keys and credential usage; 12-month audit retention; credential provider can revoke or reauthorize without exposing authentication material to requester - LABEL: VERIFIED_SPEC + PRODUCT_VOLATILE - RELATION: support / availability-gate - ESTABLISHES: credential custody can be separated from workflow authorship; audit/SSO/governance are plan-gated operational features. - DOES_NOT_ESTABLISH: every connector implements least privilege or fine-grained scopes.

E007-S05 — Zapier enterprise governance

- SOURCE_CLASS: official product documentation - TITLE: Set up single sign-on with SAML / Allowed domains for app connections / Security and Compliance / Enterprise security - AUTHOR/ORG: Zapier - URL: https://help.zapier.com/hc/en-us/articles/8496279747085-Set-up-single-sign-on-with-SAML ; https://help.zapier.com/hc/en-us/articles/43996004782349-Allowed-domains-for-app-connections ; https://zapier.com/security-compliance - ACCESSED: 2026-09-11 - VERSION/STATE: SAML article updated 2026-08-05; allowed-domains article current 2026; product controls are plan dependent - LOCATOR: SAML/JIT; allowed OAuth account domains; audit/connection tracking; enterprise observability and permissions - LABEL: VERIFIED_SPEC + PRODUCT_VOLATILE - RELATION: support / availability-gate - ESTABLISHES: no-code automationでもidentity、connection origin、audit、permissionのcontrol planeが必要。 - DOES_NOT_ESTABLISH: SaaS platformがdownstream SaaSのauthorization designを代替すること。

E007-S06 — Google Workspace Add-ons / Apps Script

- SOURCE_CLASS: official platform documentation - TITLE: Build Google Workspace add-ons / Apps Script troubleshooting - AUTHOR/ORG: Google - URL: https://developers.google.com/workspace/add-ons/how-tos/building-workspace-addons ; https://developers.google.com/apps-script/guides/support/troubleshooting - ACCESSED: 2026-09-11 - VERSION/STATE: Add-ons guide updated 2026-07-22; troubleshooting current 2026 - LOCATOR: Add-ons can use Apps Script or HTTP; OAuth scopes should be least-permissive; Apps Script recommended for lightweight/smaller-use cases while other runtimes may suit many users/low latency/full infra control; every Apps Script execution is recorded - LABEL: VERIFIED_SPEC + PRODUCT_VOLATILE - RELATION: support / limit - ESTABLISHES: Workspace integration is not synonymous with Apps Script; runtime choice and ownership/project lifecycle matter. - DOES_NOT_ESTABLISH: Apps Script is unsuitable for all enterprise use.

E007-S07 — UiPath audit and robot logs

- SOURCE_CLASS: official product documentation - TITLE: Automation Cloud — About logs / Audit Logs API - AUTHOR/ORG: UiPath - URL: https://docs.uipath.com/automation-cloud/automation-cloud/latest/admin-guide/about-logs ; https://docs.uipath.com/automation-cloud/automation-cloud/latest/api-guide/audit-logs - ACCESSED: 2026-09-11 - VERSION/STATE: Automation Cloud latest - LOCATOR: organization/tenant audit APIs; robot logs; audit retention/export differs by plan and log class - LABEL: VERIFIED_SPEC + PRODUCT_VOLATILE - RELATION: support / limit - ESTABLISHES: RPAにはworkflow controlだけでなくrobot/runtime execution telemetryと管理auditの二層が存在する。 - DOES_NOT_ESTABLISH: UI automationがAPI automationより常に低信頼であること。

E007-S08 — OpenAPI 3.2.0

- SOURCE_CLASS: open standard - TITLE: OpenAPI Specification v3.2.0 - AUTHOR/ORG: OpenAPI Initiative / Linux Foundation - URL: https://spec.openapis.org/oas/v3.2.0.html - PUBLISHED: 2025-09-19 - ACCESSED: 2026-09-11 - VERSION: 3.2.0 - LOCATOR: standard language-agnostic HTTP API description; paths, callbacks, security schemes, OAuth flows - LABEL: VERIFIED_SPEC - RELATION: support / boundary - ESTABLISHES: typed HTTP API contract is a reusable integration substrate and can describe callbacks/security requirements. - DOES_NOT_ESTABLISH: API description alone provides authorization, approval, idempotency, runtime governance or business safety.

E007-S09 — ChatGPT Plugins/Apps current terminology

- SOURCE_CLASS: official product documentation - TITLE: Apps in ChatGPT - AUTHOR/ORG: OpenAI - URL: https://help.openai.com/en/articles/11487775-connectors-in-chatgpt - ACCESSED: 2026-09-11 - VERSION/STATE: current; Plugin directory migration dated 2026-07-09 - LOCATOR: plugins may include skills/apps/app templates; apps connect external data/actions; workspace admins govern plugin installation and underlying app access/permissions; custom apps can use MCP - LABEL: VERIFIED_SPEC + PRODUCT_VOLATILE - RELATION: support / terminology-boundary - ESTABLISHES: 「Plugin」と「app/connector/action」を同義語として固定してはならず、distribution/workflow packagingとactual external connection/permissionを分離する必要がある。 - DOES_NOT_ESTABLISH: every plugin has external actions or equivalent permission model.

FINDINGS

1. 共通比較単位は「製品」よりautomation execution boundary。 各基盤を少なくとも (a) trigger/event, (b) workflow/orchestration, (c) connector/action, (d) credential/identity, (e) execution runtime, (f) audit/control plane に分解すると、Power Automate・Make・Zapier・n8n・GAS・RPAを同じ表で比較できる。 2. connector countは弱い指標。 本番適格性を左右するのは、connectorごとのauth method/scope、credential owner、shared credential可否、connection domain、DLP classification、rate limit、side-effect semantics、auditability、version/change policy。 3. 「credentialを持てる」と「workflow作者がcredential secretを読める」は別。 Make credential requestsのように、credential providerが認可しworkflow側が利用するが秘密値を閲覧しない構成が成立する。一方n8n workflow sharingでは、workflow editorがworkflow内で使われるcredentialsを利用できる旨が文書化されており、共有境界は製品ごとに異なる。 4. retry/replayは復旧機能であると同時にside-effect risk。 Power Automateのresubmit、n8nのretry-original/current等は再実行を可能にするが、外部API側のidempotency key、dedupe、transaction semanticsが無ければ二重送信・二重作成・二重決済等を防ぐ保証にはならない。 5. DLPはconnector governanceであり、downstream permissionの代替ではない。 MicrosoftのBusiness/Non-business/Blocked分類はデータ経路を制御できるが、接続先でcredentialが過剰権限ならその権限自体を縮小しない。 6. Webhook/API/RPAを一括りにしない。 API/OpenAPIはstructured contract、Webhookはevent delivery、RPAはUI/runtime操作であり、observability・retry・schema drift・auth・failure detectionが異なる。共通章では接続方式を別軸として保持する。 7. LLM/agent integrationは既存automation controlを消さない。 Copilot Studioのdata policyやChatGPT app/plugin管理の例から、agentがworkflow/actionを呼ぶ場合でもworkspace/admin/connector permissionが別control planeとして残る。 8. SaaS vs self-hostedも単純な安全順位にしない。 SaaSはSSO/audit/governanceをplan-gated featureとして提供し得る。self-hosted n8n等はnetwork/storage/upgrade/node riskを組織側がより直接負う。どちらも責任境界の比較が必要。

LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE

- vendor docsは機能存在を示すが、特定tenantの設定・実効権限・運用成熟度は示さない。 - 「監査ログあり」は、必要イベントが全て記録される・十分長く保持される・改ざん耐性がある、を自動的には意味しない。plan、event class、retention、export pathを別確認する。 - workflow retryはtransaction rollbackではない。downstream actionのidempotencyを独立検証する。 - DLP/connector blockはcredential scopeや対象SaaS内部ACLを置換しない。 - OAuth接続はleast privilegeを保証しない。実際のrequested scopeとresource ownerを確認する。 - Webhookはdelivery成功とbusiness processing成功を分ける必要があり、duplicate/out-of-order/replay/schema changeを想定する。 - UI/RPAは画面変更・selector drift・session state・desktop/runtime availabilityのfailureを持ち、API integrationとはfailure envelopeが異なる。 - self-hostingはdata/control localityを増やし得る一方、patching、secret management、network exposure、backup、queue/database運用を組織責任へ移す。

AVAILABILITY_GATES

- plan/tier: SSO、SCIM、audit retention、source control、credential governance、log streaming等がEnterprise/Business等に限定される例が多い。 - tenant/admin policy: Power Platform DLP、ChatGPT app/plugin enablement、OAuth/domain restrictions。 - connector/app support: connectorごとのregion、authentication method、scope、API limit。 - hosting mode: SaaS / self-hosted / desktop robot / cloud robot。 - identity: human OAuth、service account/service principal、shared connection、credential delegation。 - runtime: network egress、private endpoint/VPC、desktop session、queue worker、execution time/throughput quota。

IMPLEMENTATION_STATE

- Power Automate / Copilot Studio / Zapier / Make / n8n / Apps Script / UiPathはいずれもproduction-grade maintained implementations。ただし同一カテゴリではなく、iPaaS/no-code workflow、low-code platform、script runtime、RPA/orchestration等が重なる市場である。 - OpenAPI 3.2.0はinterface description standardでありautomation runtimeではない。 - ChatGPT Plugins/AppsはAI client側のdistribution/action surfaceであり、Power Automate等のgeneral-purpose automation runtimeと同一層ではない。

BENCHMARK_OR_EVAL

この領域に単一の「最強automation tool」benchmarkを置かない。導入評価は最低でも次の実測profileに分ける。 1. SUCCESS RATE: 同一fixtureでN回実行し成功/partial/failureを区別。 2. SIDE-EFFECT SAFETY: timeout後retryでduplicate actionが出ないか。 3. RECOVERY: failed runのresume/retry/replayとmanual compensation所要。 4. THROUGHPUT/LATENCY: connector throttling込みのp50/p95、queue delay。 5. CHANGE RESILIENCE: schema/API/UI change時の破損範囲。 6. GOVERNANCE COVERAGE: identity, RBAC, DLP, secret custody, audit export, environment promotion。 7. COST: task/action/execution/robot/runtime/egress等、課金単位を製品間で正規化して比較。

SECURITY / PRIVACY / AUDIT / COST / OPERABILITY

- SECURITY: connection credentialとworkflow edit permissionを分離し、service identityを人間credentialで代用しない。custom/community connector/nodeはsupply-chain対象とする。 - PRIVACY: dataがどのconnector/runtime/logへ複製されるかをdata-flow diagramで管理。execution payloadのlog retentionも確認。 - AUDIT: authoring change log、credential/connection change、execution log、admin auditを別event classとして確認する。 - COST: 「1 automation」の課金単位は比較不能。task/action/API call/execution/minute/robot/AI token等に分解する。 - OPERABILITY: owner、environment、version、rollback、retry policy、dead-letter/failed-run queue、alert、runbookを本番受入条件とする。

SAFE_WORDING_FOR_FUTURE_BOOK

「AI連携・自動化基盤は、コネクタ数ではなく、誰の資格情報で・どの権限を使い・どの実行環境から・何を変更し・失敗時にどう再実行され・その履歴を誰が監査できるかで評価する。Plugin、connector、Webhook、workflow、RPA、agent actionは接続面が重なる一方、同じ安全境界ではない。特に再実行は復旧手段であると同時に二重side effectの入口であり、外部操作のidempotencyと補償手順を独立して設計する必要がある。」

UPDATE_WORK_HANDOFF

1. 公開後本編では「ブランド一覧」より先に共通比較matrixを作る。推奨列: integration type / trigger / action / auth / credential owner / least-privilege support / DLP / human approval / retry / idempotency support / run history / audit retention/export / env promotion / self-host / private network / rate limits / cost meter / AI-agent surface。 2. Zapier, Make, n8n, Power Automate/Copilot Studio, GAS/Workspace add-ons, UiPathを代表実装として同matrixへ入れる。必要ならWorkato/Pipedream/IFTTT等を追加するが、製品数を成果指標にしない。 3. EA-160（connector/OpenAPI comparison）ではE007-S08を再利用し、OpenAPI contract、generic HTTP action、native connector、MCP tool、RPAの境界を詳細化する。 4. EA-063/159（MCP gateway）とは、connector inventory・approval・credential mediation・auditを接続する。 5. TEST_PROFILE-AUTO-001（未実施）: sandbox対象APIでcreate actionを故意にtimeout/retryし、idempotency key有/無でduplicate side effectを比較。 6. TEST_PROFILE-AUTO-002（未実施）: OAuth scope、workflow editor、credential owner、runtime identityの4者を分離した最小構成を各主要platformで照合。 7. PRODUCT_VOLATILE再確認: plan/tier、audit retention、SSO/SCIM、DLP、AI/plugin terminology、connector availability、rate limitsは出版直前に再確認。

OPEN_UNKNOWN / STOP_CONDITION