AEGIS Enterprise Ecosystem Handbook · Implementing Agentic AI Governance Across Adobe, Salesforce, ServiceNow, and Microsoft

How to read this handbook

Governance is overhead in the lab and competitive architecture at scale

Before any platform specifics, fix the mental model. The cost of governance is not constant across the lifecycle. Applying enterprise-scale controls to a lab experiment is waste; carrying lab-stage informality into production is debt that compounds. The architect's job is to know which phase each system is in, and to switch the model at the inflection point.

PHASE 01

Lab & Experimentation

Light-touch is correct

Most experiments fail and understanding is nascent. An ethics checklist and data hygiene are enough. Formal governance architecture here is disproportionate cost. Do not over-engineer.

PHASE 02 ⚡

Pilot → Production

The critical inflection

The moment agents touch real users and real data, the model must update. AEGIS Pillars 1 to 4 active before GA. This is the design moment. Every shortcut taken here creates debt that grows 4 to 7×.

PHASE 03

Enterprise Scale

Governance is the moat

All seven pillars plus the Cost Layer become a procurement differentiator, a regulatory-resilience asset, and a market-access requirement. One ungoverned incident at scale can exceed the entire annual program cost by 30 to 40×.

Founder POV

The platforms in this handbook have all converged on the same realization in the last twelve months: the agent is not the product, the governed control plane around the agent is the product. Salesforce calls it Agent Fabric. Microsoft calls it Agent 365. ServiceNow calls it the AI Control Tower. Adobe calls it the AEP Agent Orchestrator. Four names for one architectural truth that AEGIS has held from the start.

This is good news for the architect. You are no longer bolting governance on from outside; you are configuring controls the platform already exposes as first-class primitives. The work shifts from building governance to mapping AEGIS pillars onto native capabilities and proving the mapping holds under audit. That is what the rest of this handbook does, platform by platform.

Always-on dimensions

Six horizontals that cut across every platform and pillar

No platform chapter stands alone. These dimensions apply to every deployment regardless of which vendor's control plane you build on. Treat them as a checklist that runs orthogonally to the seven pillars: each platform section that follows assumes these are being governed in parallel.

INDUSTRY

Regulated verticals

Four primary verticals set the regulatory ceiling. Financial Services brings CFPB adverse-action, SR 11-7 model risk, and Reg BI. Healthcare brings HIPAA, ONC algorithm transparency, and FDA SaMD. Manufacturing brings GxP, ISO 9001, and on-premise data-sovereignty needs. Energy brings NERC CIP and NIS2 critical-infrastructure obligations. The same agent carries a different governance weight in each. The platform is constant; the pillar intensity is set by the vertical.

Financial ServicesHealthcareManufacturingEnergy

FRONTIER MODELS

Model choice and routing

All four platforms are now multi-model. The governance requirement is model selection governance (Pillar 4): an allowed-list of providers per use case, least-cost model meeting the accuracy bar, and routing logged for audit. Anthropic Claude (Opus 4.8, Sonnet 4.6), OpenAI GPT-5.5, and Google Gemini 3.1 Pro are the production frontier set; coding agents add Cursor, Claude Code, Codex, and Windsurf via the platforms' MCP surfaces. Open models (NVIDIA Nemotron, Llama-derived) cover on-premise and data-residency cases.

Claude Opus 4.8 / Sonnet 4.6OpenAI GPT-5.5Gemini 3.1 ProCursor · Claude Code · CodexNemotron / openFable 5 / Mythos 5 export-suspended

EDGE & INFRA

NVIDIA & AMD

Where the agent runs is a governance fact, not just an ops detail. NVIDIA AI Enterprise (NIM microservices, NeMo Guardrails with content-safety / topic-control / jailbreak NIMs, NeMo observability) is the dominant on-premise and edge guardrail layer. AMD Instinct MI400 series (MI430X for sovereign/HPC, MI440X for on-premise inference, UALink open interconnect) is the credible second source for enterprises wary of single-vendor lock-in. For GxP, NERC CIP, and ITAR cases, on-premise inference with zero cloud egress is the control, not a preference.

NVIDIA NIM / NeMo GuardrailsNVIDIA AI EnterpriseAMD Instinct MI400UALink open interconnect

ROBOTICS

Physical AI

When agents drive actuators, the AEGIS action allow/blocklist (Pillar 4) becomes a physical-safety control and the audit trail (Pillar 6) becomes incident-reconstruction evidence. NVIDIA Isaac GR00T (humanoid VLA models, GR00T-H vision-language-action) and Cosmos world-foundation models anchor the manufacturing and logistics robotics stack. Governance additions over digital agents: emergency-stop authority, geofenced action boundaries, and simulation-validated behavior (Omniverse) before any physical deployment.

NVIDIA Isaac GR00TCosmos world modelsOmniverse simulationGR00T N2 roadmap end-2026

HOSTING

Deployment topology

The hosting choice determines which data-residency and sovereignty obligations apply. Public cloud (Azure, AWS Bedrock, Google Cloud) is the default; sovereign and on-premise deployment is the control for regulated data that cannot leave a boundary. All four platforms now offer a sovereign or local path: Microsoft Foundry local deployment, Google Distributed Cloud, NVIDIA on-premise via AI Enterprise. The architect's job is to tag each system's data-residency requirement in the inventory (Pillar 2) and match the topology to it before GA.

Azure · AWS Bedrock · Google CloudSovereign / Distributed CloudOn-premise (NVIDIA / AMD)

COST & MAINTENANCE

The Cost Governance layer

This is the board-level horizontal most programs miss. Every platform now meters agents differently: Salesforce per-conversation, Microsoft per Copilot credit, ServiceNow per ACV, Adobe per workflow. AEGIS Cost Governance threads through Pillars 1, 2, 4, and 6: per-agent budget caps as a hard stop, cost-as-circuit-breaker on runaway loops, token and API spend in the same dashboard as bias and performance, and a monthly spend report to the CAIO and CFO. Gartner projects 40% of agents will be decommissioned by 2027 for governance and cost reasons; the cost layer is how you avoid being in that 40%.

Per-agent budget capsCost-as-circuit-breakerUnified cost + bias dashboardMonthly CAIO + CFO report

The intellectual core

One framework, four control planes: the AEGIS-to-native mapping

This is the table to keep open while you architect. Each AEGIS pillar already has a home in each platform's native governance tooling. The architect's value is knowing the mapping, configuring it deliberately, and producing the evidence that proves it. Read down a column to govern one platform; read across a row to see how the same obligation is met four different ways.

AEGIS Pillar	Salesforce Agentforce 360	Microsoft Agent 365	ServiceNow AI Control Tower	Adobe Experience Platform
P1Governance Architecture	Agent Fabric governed control plane; per-agent ownership in Agent Builder	Agent 365 control plane; agent approval & publication flow; policy templates	AI Control Tower lifecycle orchestration: intake, review, retire; AI CoE workspaces	AEP Agent Orchestrator; agency system of record preserves accountability
P2AI System Inventory	Agentforce agent registry; Data 360 lineage; AgentExchange catalog	Agent 365 registry; cross-platform sync with Bedrock & Google Cloud	Discover across 30+ integrations (Azure, AWS, GCP, SAP, Workday); CMDB	AEP data and content inventory; Content Credentials provenance
P3Risk & Impact Assessment	Agentforce Testing Center; simulation in Agent Builder before deploy	Defender Agent SPM: posture, excessive-permission & misconfig detection	Govern: 5 risk frameworks aligned to NIST AI RMF & EU AI Act; pre-deploy review	Brand Intelligence validation; output checks vs brand & compliance rules
P4Controls & Human Oversight	Agent Script deterministic control; Agentforce Guardrails; HITL escalation	Entra Conditional Access for agents; network controls; least-privilege	Now Assist Guardian: prompt-injection & output guardrails; Veza least-privilege	AEP approval workflows; deterministic Firefly Creative Production; HITL review/approve
P5Transparency & Rights	Einstein Trust Layer; zero data retention; full action auditability via Data 360	Purview sensitivity-label propagation to agent output; DSPM for AI	AI Risk & Compliance Workspace; explainable decision records	Content Credentials (C2PA) on every asset; AI-content labeling built in
P6Monitoring & Response	Agentforce transcript logs; Agentic Work Unit metering; Data 360 observability	Defender runtime detection; Purview AI Observability; audit logs in Entra	Observe (Traceloop): runtime agent-reasoning observability; financial dashboards	Content Analytics; AEP real-time monitoring of agent-driven experiences
P7Regulatory Intelligence	Spring/Summer release cadence; Trust Layer policy updates	Agent 365 policy-template updates; Compliance Manager	Built-in EU AI Act & NIST RMF frameworks, updated each release (Zurich → Australia)	Continuous Brand Intelligence learning; evolving compliance rule sets
$Cost Governance	Per-conversation metering; Agentic Work Unit tracking; Data 360 cost view	Agent usage estimator (Copilot credits); cost-management eBook; E7 bundling	Measure: financial dashboards for runaway-spend control; ROI analysis	Per-workflow cost in GenStudio; Firefly Services consumption metering

Reading note: a populated cell means the platform exposes a native primitive that substantially satisfies that pillar; it does not mean the pillar is satisfied by default. Configuration, evidence capture, and periodic validation are still the customer's responsibility under every vendor's shared-responsibility model. The cells are where you start, not where you stop.

★

Platform Chapter 01 · Customer-facing autonomy

Salesforce Agentforce 360

Agentforce 360 reached general availability on February 23, 2026 (Spring '26) and is the most production-proven customer-facing agent platform: 22,000+ deals closed in Q4 FY2026 and 85% autonomous resolution on Salesforce's own help portal. For the architect, the governance story is unusually strong because Salesforce shipped Agent Script, a deterministic control language, and Agent Fabric, a multi-vendor governed control plane, in the same cycle. The platform's hard prerequisite is Data 360 (formerly Data Cloud); without it agents have no governed context to reason over.

Native governance primitives

Agent Script

Human-readable expression language for deterministic agent control: conditional logic, precise tool use, guided steps. This is how a consequential decision becomes repeatable and auditable rather than probabilistic.

GA · Spring '26

Agent Fabric

Governed control plane for multi-vendor agents: deterministic orchestration plus centralized agent, tool, and LLM governance across the whole AI landscape, not just Salesforce-built agents.

GA · TDX 2026

Einstein Trust Layer

Zero data retention with third-party LLMs, dynamic grounding, toxicity and bias filtering, configurable guardrails for regulated use cases. The data-protection spine under every agent.

Atlas Reasoning Engine

ReAct-loop reasoning with configurable model choice (Claude, GPT-5.5, Gemini via Bedrock). Reliability is a product of architecture and clean process definitions, not the model alone.

GA · multi-model

AEGIS pillar implementation

P1Governance Architecture

Assign a named product owner per agent in Agent Builder; register multi-vendor agents in Agent Fabric so orchestration and ownership are centralized. Use the portable JSON compile target for version control and change review.

P2AI System Inventory

Maintain the agent registry with Data 360 lineage; tag each agent's action scope and data access. AgentExchange third-party agents must be inventoried too, not just homegrown ones.

P3Risk & Impact Assessment

Run pre-deployment simulations in Agentforce Testing Center; one-click simulations with real-time debugging in Agent Builder. For consequential decisions, test the full Agent Script path, not just the model.

P4Controls & Human Oversight

Encode HITL gates and prohibited actions as deterministic Agent Script steps so a consequential action cannot fire without the required precondition. Enforce least-privilege via inherited Salesforce permissions plus explicit per-agent action authorization. Apply Agentforce Guardrails for prompt-injection defense.

P5Transparency & Rights

Einstein Trust Layer enforces zero data retention and surfaces full action auditability via Data 360. For regulated decisions, configure the agent to emit an explainable decision record, not just an outcome.

P6Monitoring & Response

Audit agent transcript logs on a defined cadence (Salesforce's own guidance: appoint an owner to review and refine prompts). Track Agentic Work Units for behavior and cost-anomaly detection.

$Cost Governance

Per-conversation metering ($2 list, volume-negotiated) plus Data 360 licensing is the cost base. Budget Data 360 as part of the engagement, not a separate project; set per-agent conversation caps and watch Agentic Work Unit growth as the leading cost indicator.

Horizontal lens

Industry & frontier models

Financial Services: Agentforce Voice for Financial Services handles banking and collections; Agent Script enforces Reg BI best-interest checks as deterministic preconditions.
Healthcare: Health Cloud + Agentforce 360 for patient financial counseling; medical-director sign-off encoded as a hard Agent Script gate.
Model choice: Atlas Reasoning Engine supports Claude Sonnet, GPT-5.5, and Gemini on Bedrock. Agentforce Vibes 2.0 brings multi-model coding (Claude Sonnet + GPT-5).
Coding agents: Headless 360 exposes 60+ MCP tools to Claude Code, Cursor, Codex, and Windsurf.

Hosting & maintenance

Hosting: Salesforce-managed cloud; third-party LLMs run on Amazon Bedrock under zero-data-retention terms.
MCP: Spring '26 integrated the Model Context Protocol; agents reach external systems via governed MCP connectors with no custom API build.
Maintenance: Spring/Summer release cadence; the build-test-deploy loop collapses into Agent Builder. Plan for prompt-and-process refinement as ongoing run cost.
Data readiness is the top failure mode: poor data quality undermines agent performance more than platform limits do.

$2 / conv

List price, standard tier (volume-negotiated)

2–6 wks

Plan to production for a real use case

Data 360

Hard prerequisite; budget into the engagement

Architect's watch-out: Agentforce disables PII data masking for agents to improve performance, while keeping zero data retention. For regulated data this shifts the protection model from masking to retention controls plus permission scoping; validate that this satisfies your specific regulator before processing sensitive fields.

▤

Platform Chapter 02 · Identity-governed agents at scale

Microsoft Agent 365

Microsoft's governance story crystallized with Agent 365 (GA May 1, 2026), the centralized control plane that treats every agent as a first-class governed identity. The defining primitive is Entra Agent ID: every agent built in Copilot Studio or Foundry gets a real directory identity, so it is subject to the same Conditional Access, Identity Protection, and audit machinery as a human employee. This is the cleanest expression of AEGIS Pillar 4 least-privilege in the market: an agent calling a tool is governed exactly like an employee accessing a system. Build agents in Copilot Studio (low-code) or Microsoft Foundry (pro-code, multi-agent, sovereign deployment).

Native governance primitives

Entra Agent ID

Every agent is a governed Entra identity under a Copilot Studio identity blueprint. Conditional Access, Identity Governance, sign-in audit logs, and lifecycle management apply to agents exactly as to users.

Defender Agent SPM

Security Posture Management: continuous discovery and risk scoring of every agent, detecting excessive permissions, misconfigurations, shadow agents, and attack paths with prioritized remediation.

Public preview → GA

Purview AI Observability + DSPM

Unified visibility into how agents access and expose sensitive data; sensitivity-label propagation to agent-generated content; data-lifecycle retention for human-to-agent interactions.

GA / DSPM preview

Agent 365 control plane

Registry, approval-and-publication flow, policy templates grouping Entra / Purview / Defender controls, and rules-based lifecycle automation. Syncs with AWS Bedrock and Google Cloud for cross-platform governance.

GA · cross-cloud preview

AEGIS pillar implementation

P1Governance Architecture

Route every agent through the Agent 365 approval and publication flow before it reaches users; apply policy templates at onboarding so Entra, Purview, and Defender controls attach consistently. Build governance gates into the dev lifecycle from day one, since Copilot Studio agents auto-register as Entra identities.

P2AI System Inventory

The Agent 365 registry is your inventory; extend it with registry sync to Bedrock and Google Cloud for multicloud agents, and Defender/Intune discovery of local agents (Claude Code, Copilot CLI) on managed endpoints.

P3Risk & Impact Assessment

Use Defender Agent SPM for pre-deployment posture assessment: excessive-permission detection, misconfiguration surfacing, and attack-path analysis. Treat the SPM risk score as a deployment gate.

P4Controls & Human Oversight

Apply Entra Conditional Access for agents (MFA-equivalent, device compliance, geo, risk-based) and network controls. Autopilots run under their own Entra ID so every autonomous action is attributable. Use Microsoft Execution Containers (MXC) to contain local agents on Windows.

P5Transparency & Rights

Purview propagates sensitivity labels to agent output automatically and governs retention of human-to-agent and agent-to-human interactions; DSPM for AI gives data-risk visibility across Microsoft and non-Microsoft agents.

P6Monitoring & Response

Defender provides runtime detection (prompt injection, risky actions); Entra sign-in logs and Purview audit give the immutable trail. Context mapping, policy-based runtime blocking, and alerts flow through Intune and Defender.

$Cost Governance

The agent usage estimator forecasts Copilot-credit consumption across Copilot Studio and Dynamics 365 in one place. Cost overruns come from underscoped data grounding, not platform fees; budget grounding work explicitly. The E7 bundle consolidates the economics at high adoption.

Horizontal lens

Industry & frontier models

Cross-vertical pattern: Tier-1 IT triage and HR policy Q&A deliver 60–70% of measurable ROI in the first 90 days; both ground in SharePoint/ServiceNow/Workday.
Model choice: Foundry's model catalog includes Claude (Fable 5 available in M365 Copilot since June 10), GPT-5.5, and Gemini; Copilot Studio agents are multi-model.
Coding agents: Microsoft Agent Framework 1.0 (GA, Python + .NET) merges Semantic Kernel and AutoGen; the M365 Agents SDK is the supported path (Bot Framework SDK deprecated, community support to Sept 2026).
Open framework reach: Agent 365 SDK brings LangChain, OpenAI Agents SDK, and Semantic Kernel agents under the same control plane.

Hosting & maintenance

Hosting: Azure-backed; Foundry offers sovereign local deployment for strict topology requirements.
Licensing: Agent 365 needs no hard prerequisite but is most useful with Entra P1/P2 + Purview DLP; the clean path is the M365 E7 bundle (~$99/user/mo) consolidating E5, Copilot, Entra Suite, and Agent 365.
Maintenance: agents built before GA must be rebuilt or republished to get Entra Agent IDs; there is no legacy migration path. Plan a re-platforming pass.
Watch local agents: Defender + Intune now discover OpenClaw, Claude Code, and Copilot CLI on managed devices, this is where shadow AI surfaces.

~$99 / user

M365 E7 bundle, per month (GA May 2026)

$25–100K

Per custom agent build, 6–10 weeks

Entra ID

Every agent is a governed identity

Architect's watch-out: Agent 365 itself does not include E5-level Purview or Defender capability; security posture management is still maturing out of preview. Without Defender Agent SPM, Entra Conditional Access for agents, and Purview classifier propagation actually licensed and configured, you get agent sprawl and unreviewed shadow AI within 12 months. Licensing without an adoption and governance plan buys the tool but not the outcome.

⊞

Platform Chapter 03 · The cross-enterprise governance layer

ServiceNow AI Control Tower

ServiceNow has positioned the AI Control Tower as the governance layer for the whole enterprise, not just ServiceNow-built agents. Following the Zurich release (March 2026) and the Knowledge 2026 expansion (May 2026), it now discovers, observes, governs, secures, and measures AI across 30+ enterprise systems including Azure, AWS, Google Cloud, SAP, and Workday. For the architect whose estate spans multiple vendors, this is the strongest candidate for the single pane of glass that AEGIS Pillar 1 and 2 demand. Crucially, Anthropic is a design partner: Claude Cowork connects to ServiceNow's governed execution layer via Action Fabric, so agents built elsewhere trigger enterprise workflows under ServiceNow governance.

Native governance primitives

AI Control Tower (5 dimensions)

Discover, Observe, Govern, Secure, Measure across any system. Discovery spans 30+ integrations; full lifecycle orchestration from intake to retirement through connected workflows.

GA expanding · Aug 2026

Now Assist Guardian

Real-time guardrails for first- and second-order prompt injection, harmful outputs, sensitive-data exposure, and adversarial behavior. The runtime control point for Pillar 4.

GA · Zurich

Action Fabric + MCP Server

Opens ServiceNow's full system of action (flows, playbooks, approvals, catalogs) to any agent, Claude, Copilot, or homegrown, via the GA MCP Server. Every action is identity-verified, permission-controlled, and audited.

GA · Knowledge 2026

Veza + Armis

Veza's access-graph brings scoped permissions and least-privilege to every AI identity; Armis feeds IT/OT/IoT asset intelligence into the CMDB, turning static inventory into a live attack-surface picture.

GA / integrating

AEGIS pillar implementation

P1Governance Architecture

Run the full AI lifecycle through AI Control Tower workspaces: AI CoE, product owners, and compliance teams collaborate on intake, risk assessment, pre-deployment review, and retirement with case tracking and enterprise-wide accountability.

P2AI System Inventory

Discover across 30+ enterprise integrations gives a cross-vendor inventory no single-platform tool matches; the CMDB plus Armis asset intelligence keeps it live. This is the pillar where ServiceNow is strongest.

P3Risk & Impact Assessment

Govern ships five risk frameworks aligned to NIST AI RMF and EU AI Act out of the box; the AI Risk & Compliance Workspace aggregates bias, drift, and security risks into enterprise risk profiles with system-wide scoring.

P4Controls & Human Oversight

Now Assist Guardian enforces runtime guardrails; Veza enforces least-privilege across human, non-human, and AI identities. Action Fabric ensures every triggered action is identity-verified and permission-controlled before execution.

P5Transparency & Rights

The AI Risk & Compliance Workspace and lineage tracking give explainable, audit-grade decision records; model choice and routing are logged and constrained to an allowed provider list.

P6Monitoring & Response

Observe (via the Traceloop acquisition) delivers runtime observability into how agents reason and decide, replacing periodic audits with live metrics and alerts and the ability to course-correct mid-flight.

P7Regulatory Intelligence

EU AI Act and NIST RMF frameworks are built in and refreshed each release (Zurich → Australia cadence). Monitoring detects dormant or over-privileged agents for continual cleanup.

$Cost Governance

Measure provides financial dashboards built specifically to address runaway AI spend, giving finance and IT ROI analysis and cost control as deployments scale. This is the most explicitly cost-aware of the four control planes.

Horizontal lens

Industry & frontier models

Internal-workflow focus: ITSM, HR onboarding, and L1 support are the heartland; Agentic Playbooks let agents own cross-functional jobs end to end.
Financial Services: HDFC Bank uses AI Control Tower as the common governance layer across every AI use case, the reference pattern for a regulated single-pane-of-glass.
Model choice & routing: 3P model choice across AWS Anthropic, Azure OpenAI, and Google Gemini; the AI CoE can restrict model providers to an allowed list per skill and agent.
Robotics / edge: Project Arc (NVIDIA OpenShell sandbox, AI Control Tower governed) and the NVIDIA Enterprise AI Factory integration extend governance to data-center model workloads.

Hosting & maintenance

Hosting: Now Platform (cloud); FedRAMP-compliant for government; governs agents wherever they run via cross-platform discovery.
Microsoft interop: AI Control Tower integrates with Microsoft Agent 365 (preview); ServiceNow AI specialists will appear in the Agent 365 Marketplace as governed digital employees.
Maintenance: half-yearly named releases (Zurich, Australia); built-in frameworks update with each, lowering the Pillar 7 maintenance burden.
Build discipline: App Engine and Build Agent are deliberately governed-by-default, ServiceNow's explicit counter to ungoverned vibe coding.

30+

Enterprise systems discovered & governed

5 frameworks

NIST & EU AI Act aligned, out of the box

Aug 2026

Full Control Tower enhancements GA

Architect's watch-out: the cross-enterprise discovery and governance breadth is the reason to choose ServiceNow as the meta-layer, but it also makes it the highest-stakes integration to get right. Some Knowledge 2026 capabilities (Project Arc, certain Agent 365 integrations) are in preview, not GA. Validate which specific capabilities you are depending on are GA in your region before architecting a control plane around them.

▲

Platform Chapter 04 · Content provenance & brand-safe agents

Adobe Experience Platform

Adobe's agentic story runs through the AEP Agent Orchestrator and the GenStudio content supply chain, reframed at Summit 2026 (April) as an agentic operating layer. Its distinctive governance contribution is provenance: Content Credentials (C2PA) embed verifiable, tamper-evident attribution into every generated asset, which directly satisfies the AI-content-labeling obligations of AEGIS Pillar 5 in a way no other platform does natively. Adobe's models are commercially safe by design (Firefly is trained on licensed data), and Brand Intelligence turns static brand guidelines into a continuously learning compliance engine. For regulated marketing and customer-experience workloads, this is the brand-and-rights control plane.

Native governance primitives

AEP Agent Orchestrator

Builds, manages, and orchestrates Adobe and third-party agents grounded in customer data and content, with data governance and regulatory compliance built into the platform foundation.

Content Credentials (C2PA)

Verifiable digital provenance embedded in every asset; the Content Authority API (beta) extends this programmatically. Direct, native satisfaction of AI-content disclosure and labeling law.

GA / API beta

Brand Intelligence

A continuously learning engine that moves beyond static brand-guideline PDFs, learning from approvals, rejections, and annotations, then making that understanding available to every content agent.

GA · Summit 2026

Firefly commercially-safe models

Trained on licensed data for IP-safe generation; Custom Models and Firefly Foundry let enterprises train on proprietary brand assets with governance and brand-validation rules enforced.

GA · Image Model 5

AEGIS pillar implementation

P1Governance Architecture

Use the AEP Agent Orchestrator as the accountability hub and the agency system of record to preserve enterprise context, governance, and accountability as work moves across internal teams and external agencies.

P2AI System Inventory

Inventory content agents and the data/content they touch within AEP; Content Credentials provide an asset-level provenance trail that doubles as a generated-content registry.

P3Risk & Impact Assessment

Validate outputs against brand and compliance rules with Brand Intelligence before activation; Firefly Design Intelligence (co-built with Coca-Cola) scales brand-compliance checking across high-volume production.

P4Controls & Human Oversight

Use deterministic Firefly Creative Production workflows and AEP approval gates so content agents propose and humans approve; the campaign-brief agent drives production, review, and approval rather than publishing unsupervised.

P5Transparency & Rights

Content Credentials are the standout: every asset carries verifiable C2PA provenance and AI-content labeling automatically, satisfying CA AI Transparency Act and EU AI Act Art. 50 disclosure natively.

P6Monitoring & Response

Content Analytics measures performance at the attribute level for real-time adjustment; AEP monitors agent-driven experiences continuously across channels.

$Cost Governance

Firefly Services meters by consumption and GenStudio tracks per-workflow cost; reusable Creative Production workflows are the lever to keep per-asset cost falling as volume rises (content demand projected to grow 5× in two years).

Horizontal lens

Industry & frontier models

Cross-vertical: regulated marketing in FinServ, Healthcare, and Pharma where every customer-facing claim must be brand-checked, disclosed, and provenance-tracked.
Interop: Adobe Marketing Agent is GA in M365 Copilot and in beta across Claude Enterprise, ChatGPT Enterprise, Gemini Enterprise, Amazon Q, and IBM watsonx Orchestrate.
Open ecosystem: MCP and Skills with native integration to Anthropic, OpenAI, Google, AWS, and Microsoft.
Robotics / 3D: the Adobe–NVIDIA partnership brings 3D digital twins into enterprise production for accurate product depiction.

Hosting & maintenance

Hosting: Adobe-managed cloud (AEP, Experience Cloud, Creative Cloud); commercially-safe Firefly models reduce IP-indemnity risk versus open generation.
Governed by design: AEP bakes data governance and regulatory compliance into the agent foundation rather than as an add-on.
Maintenance: Brand Intelligence is self-improving from feedback loops, lowering the manual cost of keeping brand rules current; Frame.io review feedback can loop back into changes automatically.
Scope boundary: Adobe governs the content and customer-experience surface; pair it with one of the other three planes for non-content agent workloads.

C2PA

Native content provenance on every asset

5×

Projected content demand growth, 2 years

20,000+

Global brands built on Adobe

Architect's watch-out: Adobe is the narrowest of the four planes by design, it governs the content supply chain and customer experience, not your IT, HR, or core transactional agents. Do not over-extend it. Its unique value is provenance and brand compliance; use it for what only it does natively, and govern the rest of the estate elsewhere.

Putting it to work

The AI-readiness sequence: where to start Monday

A handbook is only useful if it tells you what to do next. This is the sequenced readiness path that applies regardless of which platform or platforms you run. Work it in order; each stage assumes the previous one is done.

Stage 1 · Inventory

See what you already have

Discover every agent in the estate, including shadow and local agents (Defender/Intune, AI Control Tower Discover)
Tag each with purpose, data access, action scope, vertical, and data-residency requirement
Classify risk tier and identify which are consequential-decision systems

Stage 2 · Map

Assign the control plane

Match each agent to its primary platform plane using the control-mapping table
Decide your meta-layer if multi-vendor (ServiceNow AI Control Tower or Microsoft Agent 365 are the cross-estate candidates)
Confirm GA status and regional availability of every primitive you depend on

Stage 3 · Gate

Activate Pillars 1 to 4 before GA

Stand up the approval-and-publication flow and assign agent owners
Encode HITL gates and prohibited actions deterministically (Agent Script, Conditional Access, Guardian)
Run pre-deployment posture and bias assessment as a hard deployment gate

Stage 4 · Run

Monitor, cost-control, evolve

Put token/API cost in the same dashboard as bias and performance; set per-agent budget caps
Stand up runtime observability and immutable audit trails (Observe, Defender, transcript logs)
Schedule the quarterly regulatory horizon scan and annual framework review (Pillar 7)

Founder POV

Notice what Stages 1 to 4 are not: a rip-and-replace. The readiness path is configuration and evidence, not construction, because the platforms have already built the primitives. The enterprises that win the next eighteen months are the ones that treat governance as the design input, not the post-incident cleanup. Gartner's projection that 40% of agents will be decommissioned by 2027 is a forecast about which enterprises did Stage 3 before GA and which ones skipped it.

AEGIS is the constant; the four control planes are how it gets expressed in the tools you already own. Map deliberately, gate before GA, and the same governance that regulators demand becomes the moat that wins enterprise procurement.