2026 05 02 ms copilot vs aws bedrock enterprise ai capability model

Vendor-agnostic enterprise Artificial Intelligence (AI) capability model: Microsoft Copilot and GitHub families vs AWS Bedrock ecosystem

Research Question

What is the complete set of architectural capabilities required to run Artificial Intelligence (AI) safely at scale in a regulated enterprise, how do Microsoft's Copilot family (Microsoft 365 Copilot Chat, Copilot Retrieval-Augmented Generation, Copilot Studio, Copilot Cowork) and Microsoft's GitHub family (GitHub Copilot, GitHub Actions, GitHub Advanced Security, GitHub Models) each address those capabilities, and how does Amazon Web Services' (AWS) AI ecosystem (Bedrock, Bedrock Agent Core, Strands Agents, Bedrock Tool Gateway, Bedrock Guardrails) compare across the same vendor-agnostic capability map?

Scope

In scope:

Deriving a complete, vendor-agnostic capability model for enterprise AI in regulated environments, extending and validating the issue's initial list through published literature and existing corpus research
Mapping each capability domain to specific Microsoft Copilot family products: Microsoft 365 (M365) Copilot Chat, Copilot Retrieval-Augmented Generation (RAG), Copilot Studio, and Copilot Cowork
Mapping each capability domain to specific Microsoft GitHub family products: GitHub Copilot, GitHub Actions, GitHub Advanced Security (GHAS), GitHub Models, and GitHub Copilot Workspace
Mapping each capability domain to AWS AI products: Bedrock, Bedrock Agent Core, Strands Agents Software Development Kit (SDK), Bedrock Tool Gateway, Bedrock Guardrails, and Bedrock Knowledge Bases
Identifying coverage gaps for each vendor ecosystem and documenting compensating controls or third-party tooling required to close them
Capability domains to assess (minimum): Knowledge Management, Compliance Training and Testing of Agents, Continuous Integration and Continuous Delivery (CI/CD) Pipeline for Agents, Change Control, Agent and Tool Registry and Discovery, Capacity Management and Financial Operations (FinOps), Application Programming Interface (API) Integration, Egress Gateway, Identity, Access Management, Auditing of Agent Build, Auditing of Agent Actions, Data Stewardship, Benefit Tracking, Observability and Application Performance Monitoring (APM), Alerting and Incident Raising, Dynamic Model Routing, Agent Gateways, Tool Gateways and Model Context Protocol (MCP) Server Hosting, Agent Throttling and Kill Switches, Control Plane, and Governance Plane

Out of scope:

Consumer-tier or non-enterprise features of either vendor's products
On-premises or air-gapped deployments not covered by a named cloud product
Third-party AI platforms (Google Vertex AI, Salesforce Einstein, ServiceNow Now Assist) unless referenced as compensating controls
Financial cost modelling or commercial licensing comparison
Full legal or compliance assessment for any specific organisation

Constraints:

Use primary vendor documentation (official product docs, architectural reference guides, conference sessions) as first-choice sources; secondary analysis acceptable where primary is unavailable
Sources must be dated within the past 24 months unless foundational architecture documentation still in active use
Treat all product capabilities as current at time of research, noting preview or roadmap status explicitly
Expand all acronyms on first use; do not present inferences as facts

Context

[fact; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html] This item asks for an architectural comparison of the Microsoft Copilot suite versus the Amazon Web Services (AWS) Bedrock ecosystem, evaluated through the lens of enterprise-grade Artificial Intelligence (AI) operating safely in a regulated environment.

[inference; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html] The comparison is more useful if Microsoft is split into the Copilot family and the GitHub family, because the two product lines serve different users, expose different governance surfaces, and solve different slices of the enterprise operating model.

[inference; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html] Prior completed work already covered control-plane gaps and a reuse-oriented capability model, but it did not map the Microsoft families and the AWS family against the same full enterprise capability taxonomy, which is the gap this item addresses.

Approach

Derive the complete capability model: start from the 22 domains named in the issue, augment through literature review of enterprise AI operating model frameworks (National Institute of Standards and Technology (NIST) AI Risk Management Framework (RMF), International Organization for Standardization (ISO)/International Electrotechnical Commission (IEC) 42001, Cloud Security Alliance (CSA) AI Safety Initiative, and vendor-published reference architectures), and produce a canonical, vendor-agnostic capability taxonomy grouped into: Foundation, Delivery Pipeline, Runtime Operations, Security and Trust, Governance and Compliance, and Economics and Value.
Map Microsoft Copilot family: for each capability domain, identify the specific M365 Copilot product that addresses it (or that gaps exist), referencing Microsoft Learn documentation, Copilot Studio admin documentation, and Purview Data Loss Prevention (DLP) controls.
Map Microsoft GitHub family: for each capability domain, identify the specific GitHub product (GitHub Copilot, GitHub Actions, GitHub Advanced Security, GitHub Models, GitHub Copilot Workspace), noting that the GitHub family primarily addresses the delivery-pipeline and developer-tooling domains rather than runtime business-user AI.
Map AWS AI ecosystem: for each capability domain, identify the specific Bedrock service (Bedrock base model access, Bedrock Knowledge Bases, Bedrock Agent Core, Strands Agents Software Development Kit (SDK), Bedrock Tool Gateway, Bedrock Guardrails, Amazon CloudWatch), referencing AWS official documentation and architecture blog posts.
Produce the coverage comparison matrix: a structured table of capability domains vs vendor products, with coverage ratings (Native, Partial, Compensating Control Required, Not Addressed) and a notes column citing the specific product feature or gap.
Identify gaps and compensating controls: for each gap, identify the most common third-party or cross-vendor compensating control (for example, API Management (APIM) gateways, Datadog observability, Open Policy Agent (OPA) policy enforcement, Microsoft Entra ID for AWS workloads).
Synthesise minimum viable architecture: for a regulated enterprise operating under each vendor family, define the minimum set of capabilities that must be present before the first production AI agent can safely run, and which products supply them.

Sources

Research Skill Output

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

§0 Initialise

[fact; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html] Research question restated: derive a vendor-agnostic capability model for regulated-enterprise Artificial Intelligence (AI), then score how the Microsoft Copilot family, the GitHub family, and the Amazon Web Services (AWS) Bedrock ecosystem cover those capabilities.
[fact; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html] Scope confirmed: business-user Copilot surfaces, developer-tooling GitHub surfaces, and AWS runtime surfaces are all in scope, with emphasis on knowledge, pipeline, runtime, security, governance, and economics controls.
[fact; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-vendor-platform-governance-constraints-compensating-controls.html] Prior completed repository work already established that regulated enterprises need a shared control plane above local agent surfaces, that governance ownership should be centralized with federated delivery, and that vendor-native controls usually require compensating layers.
[inference; source: https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models] The practical comparison unit is therefore not a single chat product but a family-level operating model, meaning the combination of first-party runtime, authoring, policy, audit, and delivery services that an enterprise can actually buy and govern together.
Metadata: output format = knowledge.

§1 Question Decomposition

Branch A: baseline capability model
- A1. Which capabilities are explicitly required by NIST Artificial Intelligence Risk Management Framework (AI RMF) and ISO/IEC 42001?
- A2. How should the named domains be grouped into Foundation, Delivery Pipeline, Runtime Operations, Security and Trust, Governance and Compliance, and Economics and Value?
Branch B: Microsoft Copilot family
- B1. Which Microsoft 365 Copilot, Agent Builder, Copilot Studio, Cowork, Purview, and agent-admin controls address knowledge, governance, and runtime oversight?
- B2. Which required capabilities remain outside the Copilot family and need adjacent Microsoft services or third-party controls?
Branch C: GitHub family
- C1. Which GitHub Copilot, GitHub Actions, GitHub Advanced Security, GitHub Models, and Copilot Workspace surfaces address delivery, model governance, and build assurance?
- C2. Which runtime and enterprise-data controls are missing from the GitHub family?
Branch D: AWS ecosystem
- D1. Which Bedrock, Bedrock Guardrails, Bedrock Knowledge Bases, AgentCore, AgentCore Gateway, and Strands Agents services cover runtime, tool, and model controls?
- D2. Which capabilities still depend on customer-assembled Identity and Access Management (IAM), CloudWatch, CloudTrail, and budgeting patterns?
Branch E: comparison
- E1. Which family is strongest on business-user knowledge and data stewardship?
- E2. Which family is strongest on developer delivery and build governance?
- E3. Which family is strongest on modular runtime orchestration?
- E4. Which capabilities remain incomplete across all three families?

§2 Investigation

Source audit and replacement notes

[fact; source: https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/] Direct GitHub Docs retrieval for GitHub Copilot Workspace returned 404 in this runtime, so the official GitHub Blog launch post and GitHub Next microsite were used to establish current public positioning; this product does not materially change enterprise control-surface scoring because the public evidence remains preview-oriented.
[fact; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/agentic-ai-frameworks/strands-agents.html; https://aws.amazon.com/blogs/opensource/introducing-strands-agents-an-open-source-ai-agents-sdk/] The seeded Strands blog URL was no longer valid, so the investigation used the official AWS Prescriptive Guidance page and the official AWS Open Source Blog launch post instead.
[fact; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html] The governance baseline is recent and primary: NIST AI RMF and ISO/IEC 42001 remain current enough to use as the vendor-agnostic reference model for this comparison.

A. Vendor-agnostic capability model

[fact; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html] NIST AI RMF defines Govern, Map, Measure, and Manage as the core enterprise functions for trustworthy AI, and ISO/IEC 42001 defines an Artificial Intelligence Management System that must establish, implement, maintain, and continually improve organization-wide AI controls.
[inference; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html] The named issue domains consolidate cleanly into six groups: Foundation, Delivery Pipeline, Runtime Operations, Security and Trust, Governance and Compliance, and Economics and Value.
[inference; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html] A complete regulated-enterprise model needs all 22 domains named in the issue, spanning knowledge, delivery, runtime operations, security and trust, governance and compliance, and economics and value.

B. Microsoft Copilot family findings

[fact; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy] Microsoft 365 Copilot grounds responses in Microsoft Graph and app context, uses the Microsoft 365 service boundary, and only surfaces data that the requesting user is already permitted to access.
[fact; source: https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/purview/audit-copilot] Microsoft provides first-party data-stewardship controls for Copilot through Data Loss Prevention (DLP), sensitivity-label exclusions, prompt blocking, accessed-resource logging, and agent identifiers in audit records.
[fact; source: https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio] Copilot Studio adds low-code agent authoring, connectors, agent flows, real-time data-policy enforcement, and detailed authoring and usage audit events.
[fact; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/agent-builder; https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/agents-overview] Agent Builder enables declarative agents grounded in SharePoint and Microsoft 365 connector data, while custom engine agents support deeper orchestration at the cost of extra hosting and enterprise responsibility.
[fact; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent?view=o365-worldwide] Microsoft now exposes tenant-wide agent registry and Frontier tool-governance pages that can inventory agents, identify ownerless agents, and approve or block MCP servers.
[fact; source: https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage?view=o365-worldwide; https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management] Microsoft provides adoption and prompt metrics for Copilot and metered Copilot Credits accounting for Copilot Studio, but the economics plane is split across license reports, usage reports, and credits rather than one unified runtime FinOps console.
[inference; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide] The Microsoft Copilot family is strongest where business-user productivity, enterprise knowledge, and tenant-bound governance are tightly coupled, especially knowledge management, data stewardship, and approved-agent distribution.
[inference; source: https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio; https://learn.microsoft.com/en-us/microsoft-365/copilot/agent-essentials/m365-agents-admin-guide; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization] The same family is weaker on code-centric CI/CD, generalized egress governance, and portable runtime kill switches because those controls live in adjacent Microsoft admin, Azure, or GitHub surfaces rather than inside Copilot alone.

C. GitHub family findings

[fact; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security] GitHub Actions provides the core CI/CD automation substrate, while GitHub Advanced Security adds code scanning, dependency review, secret scanning, push protection, and security campaigns across repositories.
[fact; source: https://docs.github.com/en/copilot/concepts/policies; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/manage-enterprise-policies] GitHub Copilot exposes enterprise policy controls for features, models, privacy behaviors, agents, and MCP availability, with enterprise-level override and organization-level delegation.
[fact; source: https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs] GitHub provides adoption, code-generation, and pull-request lifecycle metrics, plus audit logs for policy changes, license changes, and GitHub website agent activity, but it explicitly excludes local prompt session data unless customers add their own logging hooks.
[fact; source: https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://docs.github.com/en/enterprise-cloud@latest/github-models/github-models-at-scale/manage-models-at-scale] GitHub Models provides a model catalog, prompt configurations stored in repositories, side-by-side model comparison, quantitative evaluators, publisher and model allowlists, and bring-your-own-key integration, although organization features remain in public preview and some extensions bypass organization restrictions.
[fact; source: https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/] GitHub Copilot Workspace was publicly positioned as an issue-to-plan-to-code preview surface that hands work back into pull requests, GitHub Actions, and human review, rather than as a persistent enterprise runtime control plane.
[inference; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models] The GitHub family is strongest on delivery-pipeline rigor, build auditability, model experimentation, and developer benefit tracking.
[inference; source: https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.github.com/en/enterprise-cloud@latest/github-models/github-models-at-scale/manage-models-at-scale; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/] The GitHub family does not natively provide full local prompt-session logging, tenant-bound business-data stewardship, or a clearly durable runtime control surface beyond preview-oriented workspace material, so it is a delivery-plane family rather than a full runtime governance family.

D. AWS Bedrock ecosystem findings

[fact; source: https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html] Amazon Bedrock is a managed multi-model platform with default model availability, provider-specific enablement steps, and account-level IAM and marketplace prerequisites.
[fact; source: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html; https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html] Bedrock Guardrails adds configurable content, sensitive-information, grounding, and automated-reasoning checks, while Bedrock Knowledge Bases adds managed RAG, citations, multimodal retrieval, and structured-query support.
[fact; source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html; https://docs.aws.amazon.com/bedrock/latest/userguide/geographic-cross-region-inference.html] Bedrock supports invocation logging to Amazon Simple Storage Service (Amazon S3) and CloudWatch, encryption and CloudTrail under the shared-responsibility model, and geography-bound cross-Region inference for residency-sensitive workloads.
[fact; source: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html] AgentCore extends the Bedrock stack with secure runtime, memory, identity, gateway, observability, evaluations, policy, and registry services, and Gateway can convert APIs, Lambda functions, and services into MCP-compatible tools with both ingress and egress authentication.
[fact; source: https://docs.aws.amazon.com/prescriptive-guidance/latest/agentic-ai-frameworks/strands-agents.html; https://aws.amazon.com/blogs/opensource/introducing-strands-agents-an-open-source-ai-agents-sdk/] Strands Agents is an AWS-originated open-source SDK that supports model-first agent loops, multi-agent collaboration patterns, MCP integration, and broad AWS service connectivity.
[inference; source: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html] The AWS ecosystem is the strongest of the three on modular runtime architecture, because it combines model access, RAG, guardrails, runtime, registry, evaluations, and tool gateway services in one family.
[inference; source: https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://docs.aws.amazon.com/bedrock/latest/userguide/geographic-cross-region-inference.html] The same ecosystem still expects customers to assemble enterprise completeness from IAM, Service Control Policies, CloudWatch, CloudTrail, budget controls, and external CI/CD, so strong runtime primitives do not eliminate governance integration work.

E. Cross-family comparison

[inference; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html] The three families specialize in different slices of the enterprise capability map: Microsoft Copilot on tenant-grounded business productivity, GitHub on software delivery governance, and AWS on modular runtime control.
[inference; source: https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] None of the three families independently provides a complete regulated-enterprise operating model that spans build pipeline, runtime enforcement, cost attribution, business-data stewardship, registry, benefit tracking, and emergency control with one native surface.
[inference; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-26-vendor-platform-governance-constraints-compensating-controls.html; https://www.nist.gov/itl/ai-risk-management-framework] The minimum safe design remains layered: vendor-native strengths should be used where they are strongest, and cross-cutting control-plane, policy, and FinOps gaps should be filled with adjacent first-party or third-party controls.

§3 Reasoning

[inference; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html] A capability model for regulated enterprise AI is only decision-useful if it spans both pre-production and runtime control surfaces, because NIST and ISO require governance, measurement, and continual improvement rather than isolated product administration.
[inference; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention] Microsoft Copilot earns the highest score on knowledge management and data stewardship because it is the only family in this comparison whose core user experience is natively grounded in the same enterprise permissions and sensitivity-label systems that govern the underlying content.
[inference; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models] GitHub earns the highest score on delivery-pipeline capability because version control, change control, automation, model evaluation, and code security all operate inside the same repository-centric workflow.
[inference; source: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html; https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html] AWS earns the highest score on modular runtime architecture because Bedrock plus AgentCore exposes the broadest set of first-party runtime services for tool mediation, policy, observability, and evaluations.
[inference; source: https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage?view=o365-worldwide; https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html] Benefit tracking and FinOps are incomplete across all three families because each exposes some metrics or pricing signals, but none combines business impact, runtime spend, and per-agent accountability into one closed-loop enterprise value system.

§4 Consistency Check

[fact; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/agent-builder; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent?view=o365-worldwide] The Microsoft evidence is internally consistent on the main thesis: first-party registry and tool-governance surfaces exist, but some of them are limited by licensing, Frontier rollout, or product boundary, so Microsoft coverage is strongest where tenant-bound data and admin surfaces already exist.
[fact; source: https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/manage-enterprise-policies; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.github.com/en/enterprise-cloud@latest/github-models/github-models-at-scale/manage-models-at-scale] The GitHub evidence is also consistent: enterprise policy, metrics, and model-governance primitives are real, but runtime prompt logging, enterprise data stewardship, and organization-model restrictions all have explicit limitations.
[fact; source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html] The AWS evidence is consistent on a different boundary: runtime services are broad, but permissions, subscriptions, Region controls, and logging destinations remain customer-configured rather than fully opinionated by the product family.
[inference; source: https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/] GitHub Copilot Workspace remains the least stable element in the comparison, so it is treated as a preview adjunct that does not materially improve the GitHub family's regulated-enterprise control-plane completeness.

§5 Depth and Breadth Expansion

[inference; source: https://www.iso.org/standard/81230.html; https://learn.microsoft.com/en-us/purview/audit-copilot; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html] From a regulatory lens, the decisive difference is not model quality but how well each family can evidence traceability, retention, policy enforcement, and accountable ownership.
[inference; source: https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage?view=o365-worldwide; https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html] From an economic lens, Microsoft and GitHub provide stronger native adoption analytics than AWS, while AWS provides stronger runtime modularity, so Microsoft-first and GitHub-first estates learn faster about usage but AWS-first estates can shape runtime mechanics more precisely.
[inference; source: https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html] From a security lens, Microsoft is strongest on tenant-bound content controls, GitHub is strongest on software-supply-chain controls, and AWS is strongest on tool mediation and runtime policy hooks.
[inference; source: https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html] From an operating-model lens, regulated enterprises that use more than one family should not split governance by vendor alone; they should centralize control-plane ownership and let delivery teams specialize by user group or product surface.

§6 Synthesis

Executive summary:

Neither the Microsoft Copilot family, the GitHub family, nor the AWS Bedrock ecosystem natively delivers the full regulated-enterprise AI capability stack on its own. [inference; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html]
Microsoft Copilot has the strongest fit when the enterprise problem is governed reuse of tenant-bound work data, because Purview, Microsoft Graph permissions, and first-party agent administration sit close to the underlying business content. [inference; source: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html]
GitHub has the strongest fit when the enterprise problem is delivery-pipeline rigor, developer telemetry, and build assurance, because Actions, GitHub Advanced Security, GitHub Copilot policies, and GitHub Models all operate inside the same repository workflow. [inference; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html]
AWS has the strongest fit when the enterprise problem is modular runtime architecture, because Bedrock, Bedrock Guardrails, Knowledge Bases, and AgentCore together cover model access, tool mediation, registry, policy, observability, and evaluation more completely than the other two families. [inference; source: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html; https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html]
Those rankings would change somewhat if Microsoft adjacent admin services were excluded, if GitHub were judged only as a delivery plane, or if AWS were weighted more heavily on business-user content governance, which is why the comparison is more reliable as a boundary-aware synthesis than as an absolute winner table. [inference; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html]
A layered target architecture is usually the safer fit for regulated enterprises that need breadth across business knowledge, software delivery, and modular runtime operations, while a single-vendor estate plus adjacent controls can still be the better trade-off where operational simplicity matters more than capability breadth. [inference; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-control-plane-architecture-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization]

Key findings:

A regulated-enterprise AI capability model needs at least 22 control domains across foundation, delivery, runtime, security, governance, and economics, and no single family in this comparison covers them all natively. ([inference]; high confidence; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html)
Under the family boundaries used in this item, the Microsoft Copilot family is strongest for knowledge management, data stewardship, and business-user governance because Microsoft Graph permissions, Purview controls, Copilot Studio policy enforcement, and the Microsoft agent registry sit close to the underlying work data. ([inference]; medium confidence; source: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html)
The Microsoft Copilot family remains incomplete for generalized CI/CD, portable egress mediation, and fleet-level kill-switch control, so a regulated deployment still needs adjacent Microsoft admin services, GitHub workflows, or third-party control points. ([inference]; medium confidence; source: https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio; https://learn.microsoft.com/en-us/microsoft-365/copilot/agent-essentials/m365-agents-admin-guide; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization)
Under the family boundaries used in this item, the GitHub family is strongest for agent delivery pipelines because GitHub Actions, GitHub Advanced Security, GitHub Copilot policies, audit logs, and GitHub Models combine change control, security assurance, and model evaluation inside one repository-centered operating loop. ([inference]; medium confidence; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/copilot/concepts/policies; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html)
The GitHub family does not natively provide the enterprise runtime controls needed for regulated business-user agents, especially full local prompt-session logging, tenant-bound business-data stewardship, and a clearly durable runtime control surface beyond preview-oriented workspace material. ([inference]; medium confidence; source: https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.github.com/en/enterprise-cloud@latest/github-models/github-models-at-scale/manage-models-at-scale; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/)
Under the family boundaries used in this item, the AWS Bedrock ecosystem is strongest for modular runtime capability because Bedrock plus AgentCore provides model access, guardrails, retrieval, runtime, registry, policy, observability, evaluations, and tool gateway services within one family. ([inference]; medium confidence; source: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html; https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html)
The AWS Bedrock ecosystem still requires customer-built governance integration for identity boundaries, Region policy, logging destinations, and economic accountability, so it is a strong runtime platform but not a complete enterprise governance plane by itself. ([inference]; medium confidence; source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html; https://docs.aws.amazon.com/bedrock/latest/userguide/geographic-cross-region-inference.html)
A layered multi-family architecture is usually the most defensible target state for regulated enterprises that need broad capability coverage, although a single-vendor estate plus adjacent controls can still be the better trade-off where integration simplicity outweighs capability breadth. ([inference]; medium confidence; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-26-vendor-platform-governance-constraints-compensating-controls.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-control-plane-architecture-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization)

Evidence map:

Claim	Source	Confidence	Notes
[inference] The capability model must span 22 domains across governance, delivery, runtime, security, and economics, and no one family covers all of them.	https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html	high	baseline synthesis
[inference] Under the family boundaries used here, Microsoft Copilot is strongest on knowledge management, data stewardship, and business-user governance.	https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html	medium	boundary-sensitive ranking
[inference] Microsoft Copilot remains incomplete for generalized CI/CD, egress mediation, and fleet kill switches.	https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio; https://learn.microsoft.com/en-us/microsoft-365/copilot/agent-essentials/m365-agents-admin-guide; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization	medium	adjacent controls required
[inference] Under the family boundaries used here, GitHub is strongest on delivery-pipeline capability, change control, and model experimentation.	https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html	medium	boundary-sensitive ranking
[inference] GitHub lacks native business-data stewardship and a durable runtime control surface beyond preview-oriented workspace material.	https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.github.com/en/enterprise-cloud@latest/github-models/github-models-at-scale/manage-models-at-scale; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/	medium	runtime gap
[inference] Under the family boundaries used here, AWS Bedrock is strongest on modular runtime architecture.	https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html; https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html	medium	boundary-sensitive ranking
[inference] AWS still needs customer-built governance integration for identity, logging, Region policy, and economics.	https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html; https://docs.aws.amazon.com/bedrock/latest/userguide/geographic-cross-region-inference.html	medium	shared responsibility
[inference] A layered multi-family architecture is usually the most defensible target state, although some estates may still prefer single-vendor simplicity plus adjacent controls.	https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-26-vendor-platform-governance-constraints-compensating-controls.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-control-plane-architecture-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization	medium	trade-off-sensitive synthesis

Assumptions:

[assumption; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent?view=o365-worldwide] Microsoft Agent Registry and Tools are counted as Microsoft-side capability coverage even though they are adjacent admin services, because the research question asks about the complete Microsoft family operating model rather than one isolated chat surface. Justification: they are first-party governance surfaces used to manage the same agents and tools.
[assumption; source: https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/] GitHub Copilot Workspace is excluded from durable regulated-enterprise scoring because current public evidence remains preview-oriented and does not show a stable long-term control surface comparable to GitHub Actions, GitHub Copilot, or GitHub Models. Justification: counting it as fully shipping enterprise coverage would overstate GitHub's runtime completeness.

Analysis:

The evidence favors a family-level comparison because each vendor distributes capabilities across multiple products and admin planes rather than inside one monolith. [inference; source: https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models]
Microsoft Copilot leads where the enterprise problem is secure reuse of existing work data and managed business-user access, not where the problem is generalized software delivery or cloud-neutral runtime control. [inference; source: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html]
GitHub leads where the enterprise problem is turning agent changes into reviewable, testable, and auditable software artifacts, not where the problem is protecting non-repository business content or mediating live business actions. [inference; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html]
Microsoft can still be a reasonable single-vendor choice for some regulated estates if they prioritize operational simplicity and are willing to accept narrower delivery-pipeline flexibility than a layered multi-family design would provide. [inference; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/agent-essentials/m365-agents-admin-guide; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html]
AWS leads where the enterprise problem is building and operating modular autonomous runtimes, but its shared-responsibility model means governance completeness still depends on how the customer wires identity, logs, budgets, and Region policy around the runtime. [inference; source: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html]
The ranking results are therefore conditional rather than absolute: excluding Microsoft adjacent admin surfaces would weaken Microsoft, scoping GitHub only as a delivery plane would strengthen GitHub's relative fit, and weighting business-data governance above runtime modularity would weaken AWS's headline advantage. [inference; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html]

Risks, gaps, uncertainties:

GitHub Copilot Workspace remains a weak evidence surface for current enterprise planning because the publicly accessible material is still preview-centered rather than a current stable operations manual. [inference; source: https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/]
The Microsoft comparison partly relies on adjacent governance surfaces such as Agent Registry, Tools, Foundry Control Plane, and Purview, which means some "Microsoft-native" coverage is family-native rather than Copilot-surface-native. [inference; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent?view=o365-worldwide; https://learn.microsoft.com/en-us/azure/foundry/control-plane/overview]
The AWS economics plane is still fragmented across pricing pages, quotas, logging, and account-level billing rather than one clearly documented per-agent value-management surface. [inference; source: https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html]

Open questions:

How much of Microsoft 365 agents governance will become generally available outside Frontier, and how quickly will its MCP governance become standard rather than preview-limited?
Will GitHub Models move from preview governance features to a durable enterprise control surface with tighter audit and runtime-policy integration?
How quickly will AWS add first-party benefit-tracking and budget-governance patterns that connect runtime spend to business outcomes rather than only to infrastructure telemetry?

§7 Recursive Review

[fact; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html] Recursive review result: the final synthesis stays inside the evidence gathered, uses explicit fact and inference labels, and grounds the main comparative claims in current primary vendor documentation plus directly relevant prior completed items.
[fact; source: https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html] No unresolved contradiction changes the final answer, although each family retains product-boundary caveats that lower confidence on some coverage scores from high to medium where evidence shows preview limits or missing runtime telemetry.

Findings

(Populated from §6 Synthesis above.)

Executive Summary

Neither the Microsoft Copilot family, the GitHub family, nor the AWS Bedrock ecosystem natively delivers the full regulated-enterprise Artificial Intelligence (AI) capability stack on its own. [inference; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html]
Microsoft Copilot has the strongest fit when the enterprise problem is governed reuse of tenant-bound work data, because Purview, Microsoft Graph permissions, and first-party agent administration sit close to the underlying business content. [inference; source: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html]
GitHub has the strongest fit when the enterprise problem is delivery-pipeline rigor, developer telemetry, and build assurance, because Actions, GitHub Advanced Security, GitHub Copilot policies, and GitHub Models all operate inside the same repository workflow. [inference; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html]
AWS has the strongest fit when the enterprise problem is modular runtime architecture, because Bedrock, Bedrock Guardrails, Knowledge Bases, and AgentCore together cover model access, tool mediation, registry, policy, observability, and evaluation more completely than the other two families. [inference; source: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html; https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html]
Those rankings would change somewhat if Microsoft adjacent admin services were excluded, if GitHub were judged only as a delivery plane, or if AWS were weighted more heavily on business-user content governance, which is why the comparison is more reliable as a boundary-aware synthesis than as an absolute winner table. [inference; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html]
A layered multi-family architecture is usually the safer fit for regulated enterprises that need breadth across business knowledge, software delivery, and modular runtime operations, while a single-vendor estate plus adjacent controls can still be the better trade-off where operational simplicity matters more than capability breadth. [inference; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-26-vendor-platform-governance-constraints-compensating-controls.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-control-plane-architecture-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization]

Key Findings

A regulated-enterprise AI capability model needs at least 22 control domains across foundation, delivery, runtime, security, governance, and economics, and no single family in this comparison covers them all natively. ([inference]; high confidence; source: https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html)
Under the family boundaries used in this item, the Microsoft Copilot family is strongest for knowledge management, data stewardship, and business-user governance because Microsoft Graph permissions, Purview controls, Copilot Studio policy enforcement, and the Microsoft agent registry sit close to the underlying work data. ([inference]; medium confidence; source: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html)
The Microsoft Copilot family remains incomplete for generalized CI/CD, portable egress mediation, and fleet-level kill-switch control, so a regulated deployment still needs adjacent Microsoft admin services, GitHub workflows, or third-party control points. ([inference]; medium confidence; source: https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio; https://learn.microsoft.com/en-us/microsoft-365/copilot/agent-essentials/m365-agents-admin-guide; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization)
Under the family boundaries used in this item, the GitHub family is strongest for agent delivery pipelines because GitHub Actions, GitHub Advanced Security, GitHub Copilot policies, audit logs, and GitHub Models combine change control, security assurance, and model evaluation inside one repository-centered operating loop. ([inference]; medium confidence; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/copilot/concepts/policies; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html)
The GitHub family does not natively provide the enterprise runtime controls needed for regulated business-user agents, especially full local prompt-session logging, tenant-bound business-data stewardship, and a clearly durable runtime control surface beyond preview-oriented workspace material. ([inference]; medium confidence; source: https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.github.com/en/enterprise-cloud@latest/github-models/github-models-at-scale/manage-models-at-scale; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/)
Under the family boundaries used in this item, the AWS Bedrock ecosystem is strongest for modular runtime capability because Bedrock plus AgentCore provides model access, guardrails, retrieval, runtime, registry, policy, observability, evaluations, and tool gateway services within one family. ([inference]; medium confidence; source: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html; https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html)
The AWS Bedrock ecosystem still requires customer-built governance integration for identity boundaries, Region policy, logging destinations, and economic accountability, so it is a strong runtime platform but not a complete enterprise governance plane by itself. ([inference]; medium confidence; source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html; https://docs.aws.amazon.com/bedrock/latest/userguide/geographic-cross-region-inference.html)
A layered multi-family architecture is usually the most defensible target state for regulated enterprises that need broad capability coverage, although a single-vendor estate plus adjacent controls can still be the better trade-off where integration simplicity outweighs capability breadth. ([inference]; medium confidence; source: https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-26-vendor-platform-governance-constraints-compensating-controls.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-control-plane-architecture-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization)

Evidence Map

Claim	Source	Confidence	Notes
[inference] The capability model must span 22 domains across governance, delivery, runtime, security, and economics, and no one family covers all of them.	https://www.nist.gov/itl/ai-risk-management-framework; https://www.iso.org/standard/81230.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-capability-model.html	high	baseline synthesis
[inference] Under the family boundaries used here, Microsoft Copilot is strongest on knowledge management, data stewardship, and business-user governance.	https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html	medium	boundary-sensitive ranking
[inference] Microsoft Copilot remains incomplete for generalized CI/CD, egress mediation, and fleet kill switches.	https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio; https://learn.microsoft.com/en-us/microsoft-365/copilot/agent-essentials/m365-agents-admin-guide; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization	medium	adjacent controls required
[inference] Under the family boundaries used here, GitHub is strongest on delivery-pipeline capability, change control, and model experimentation.	https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html	medium	boundary-sensitive ranking
[inference] GitHub lacks native business-data stewardship and a durable runtime control surface beyond preview-oriented workspace material.	https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.github.com/en/enterprise-cloud@latest/github-models/github-models-at-scale/manage-models-at-scale; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/	medium	runtime gap
[inference] Under the family boundaries used here, AWS Bedrock is strongest on modular runtime architecture.	https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html; https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html	medium	boundary-sensitive ranking
[inference] AWS still needs customer-built governance integration for identity, logging, Region policy, and economics.	https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html; https://docs.aws.amazon.com/bedrock/latest/userguide/geographic-cross-region-inference.html	medium	shared responsibility
[inference] A layered multi-family architecture is usually the most defensible target state, although some estates may still prefer single-vendor simplicity plus adjacent controls.	https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-26-vendor-platform-governance-constraints-compensating-controls.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-control-plane-architecture-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html; https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization	medium	trade-off-sensitive synthesis

Assumptions

[assumption; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent?view=o365-worldwide] Microsoft Agent Registry and Tools are counted as Microsoft-side capability coverage even though they are adjacent admin services, because the research question asks about the complete Microsoft family operating model rather than one isolated chat surface. Justification: they are first-party governance surfaces used to manage the same agents and tools.
[assumption; source: https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/] GitHub Copilot Workspace is excluded from durable regulated-enterprise scoring because current public evidence remains preview-oriented and does not show a stable long-term control surface comparable to GitHub Actions, GitHub Copilot, or GitHub Models. Justification: counting it as fully shipping enterprise coverage would overstate GitHub's runtime completeness.

Analysis

The evidence favors a family-level comparison because each vendor distributes capabilities across multiple products and admin planes rather than inside one monolith. [inference; source: https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models]
Microsoft Copilot leads where the enterprise problem is secure reuse of existing work data and managed business-user access, not where the problem is generalized software delivery or cloud-neutral runtime control. [inference; source: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html]
GitHub leads where the enterprise problem is turning agent changes into reviewable, testable, and auditable software artifacts, not where the problem is protecting non-repository business content or mediating live business actions. [inference; source: https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.github.com/en/get-started/learning-about-github/about-github-advanced-security; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html]
Microsoft can still be a reasonable single-vendor choice for some regulated estates if they prioritize operational simplicity and are willing to accept narrower delivery-pipeline flexibility than a layered multi-family design would provide. [inference; source: https://learn.microsoft.com/en-us/microsoft-365/copilot/agent-essentials/m365-agents-admin-guide; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-28-alternative-pipeline-platforms-copilot-studio-agents.html]
AWS leads where the enterprise problem is building and operating modular autonomous runtimes, but its shared-responsibility model means governance completeness still depends on how the customer wires identity, logs, budgets, and Region policy around the runtime. [inference; source: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html]
The ranking results are therefore conditional rather than absolute: excluding Microsoft adjacent admin surfaces would weaken Microsoft, scoping GitHub only as a delivery plane would strengthen GitHub's relative fit, and weighting business-data governance above runtime modularity would weaken AWS's headline advantage. [inference; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html]

Capability comparison matrix

Capability domain	Microsoft Copilot family	GitHub family	AWS ecosystem	Source	Notes
[inference] Knowledge management	Native	Partial	Native	https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/agent-builder; https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html	Microsoft Graph and Bedrock Knowledge Bases are first-party grounding layers.
[inference] Compliance training and testing of agents	Partial	Native	Native	https://learn.microsoft.com/en-us/azure/foundry/control-plane/overview; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html	Microsoft needs adjacent Foundry or admin controls.
[inference] CI/CD pipeline for agents	Partial	Native	Partial	https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html	GitHub owns the cleanest build pipeline story.
[inference] Change control	Partial	Native	Partial	https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html	Microsoft and AWS require adjacent workflow discipline.
[inference] Agent and tool registry and discovery	Partial	Partial	Native	https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent?view=o365-worldwide; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html	AWS AgentCore Registry is the clearest runtime registry.
[inference] Capacity management and FinOps	Partial	Partial	Partial	https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage?view=o365-worldwide; https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management; https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html	Metrics exist, unified per-agent FinOps does not.
[inference] API integration	Native	Native	Native	https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio; https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/agents-overview; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html	All three families expose integration surfaces.
[inference] Egress gateway	Compensating control required	Compensating control required	Native	https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://docs.github.com/en/copilot/concepts/policies; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html	Only AWS documents an explicit managed egress layer.
[inference] Identity	Native	Partial	Native	https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://docs.github.com/en/copilot/concepts/policies; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html	GitHub identity is user-centric rather than agent-centric.
[inference] Access management	Native	Partial	Native	https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-data-loss-prevention; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/manage-enterprise-policies; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-identity-access-management-enterprise.html	AWS and Microsoft expose stronger policy detail.
[inference] Auditing of agent build	Partial	Native	Partial	https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html	repository-centered build audit trail
[inference] Auditing of agent actions	Native	Partial	Native	https://learn.microsoft.com/en-us/purview/audit-copilot; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html	GitHub lacks local prompt visibility by default.
[inference] Data stewardship	Native	Partial	Partial	https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; https://docs.github.com/en/copilot/concepts/policies; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html	Microsoft is strongest because the data boundary is product-native.
[inference] Benefit tracking	Partial	Native	Partial	https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage?view=o365-worldwide; https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html	GitHub has the clearest developer impact telemetry.
[inference] Observability and APM	Partial	Partial	Native	https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-logging-copilot-studio; https://learn.microsoft.com/en-us/azure/foundry/control-plane/overview; https://docs.github.com/en/copilot/concepts/copilot-usage-metrics/copilot-metrics; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html	managed runtime telemetry emphasis
[inference] Alerting and incident raising	Partial	Partial	Partial	https://learn.microsoft.com/en-us/azure/foundry/control-plane/overview; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/review-audit-logs; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html	All three rely on adjacent alerting stacks.
[inference] Dynamic model routing	Compensating control required	Partial	Partial	https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://docs.github.com/en/enterprise-cloud@latest/github-models/about-github-models; https://docs.aws.amazon.com/bedrock/latest/userguide/geographic-cross-region-inference.html	None documents rich policy-based model brokering end to end.
[inference] Agent gateways	Compensating control required	Compensating control required	Native	https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization; https://github.blog/news-insights/product-news/github-copilot-workspace/; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html	explicit managed gateway documentation
[inference] Tool gateways and approved external tool connectivity	Partial	Partial	Native	https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent?view=o365-worldwide; https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/manage-enterprise-policies; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html	Microsoft and GitHub govern Model Context Protocol (MCP) usage more than they host the tool endpoints themselves.
[inference] Agent throttling and kill switches	Partial	Partial	Partial	https://learn.microsoft.com/en-us/purview/dlp-microsoft365-copilot-location-learn-about; https://docs.github.com/en/copilot/concepts/policies; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html	Fast disable exists, dedicated kill-switch plane does not.
[inference] Cross-plane coordination for agents, policy, and operations	Partial	Partial	Partial	https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html; https://davidamitchell.github.io/Research/research/2026-04-26-ai-agent-control-plane-architecture-enterprise.html; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html	Family-local operating surfaces exist, but estate-wide coordination still requires layering.
[inference] Enterprise-wide governance coverage	Partial	Partial	Partial	https://www.nist.gov/itl/ai-risk-management-framework; https://learn.microsoft.com/en-us/purview/audit-copilot; https://docs.github.com/en/copilot/concepts/policies; https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html	Governance obligations remain broader than any one family.

Risks, Gaps, and Uncertainties

GitHub Copilot Workspace remains a weak evidence surface for current enterprise planning because the publicly accessible material is still preview-centered rather than a current stable operations manual. [inference; source: https://github.blog/news-insights/product-news/github-copilot-workspace/; https://githubnext.com/projects/copilot-workspace/]
The Microsoft comparison partly relies on adjacent governance surfaces such as Agent Registry, Tools, Foundry Control Plane, and Purview, which means some "Microsoft-native" coverage is family-native rather than Copilot-surface-native. [inference; source: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-registry?view=o365-worldwide; https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent?view=o365-worldwide; https://learn.microsoft.com/en-us/azure/foundry/control-plane/overview]
The AWS economics plane is still fragmented across pricing pages, quotas, logging, and account-level billing rather than one clearly documented per-agent value-management surface. [inference; source: https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html; https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html]

Open Questions

How much of Microsoft 365 agents governance will become generally available outside Frontier, and how quickly will its MCP governance become standard rather than preview-limited?
Will GitHub Models move from preview governance features to a durable enterprise control surface with tighter audit and runtime-policy integration?
How quickly will AWS add first-party benefit-tracking and budget-governance patterns that connect runtime spend to business outcomes rather than only to infrastructure telemetry?

Output

(Fill in when completing, what was produced as a result of this research?)

Type: knowledge
Description: A vendor-agnostic enterprise Artificial Intelligence (AI) capability model and comparison showing that, under the family boundaries used here, Microsoft Copilot is strongest on tenant-grounded knowledge and stewardship, GitHub is strongest on delivery-pipeline governance, and AWS is strongest on modular runtime architecture, with compensating controls still required for complete regulated deployment. [inference; source: https://www.nist.gov/itl/ai-risk-management-framework; https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview; https://docs.github.com/en/actions/get-started/understand-github-actions; https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html; https://davidamitchell.github.io/Research/research/2026-04-22-enterprise-ai-platform-operating-models.html; https://davidamitchell.github.io/Research/research/2026-04-26-multi-ai-provider-control-planes.html]
Links:
https://www.nist.gov/itl/ai-risk-management-framework
https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview
https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html