← Back to Insights
Briefing · Platform Strategy · September 2026

No Single Vendor Owns the Enterprise Action Layer

A placement lens for deciding where state, identity, policy, execution, evidence, and recovery should live when agents cross enterprise systems.

Conditional sign-off verdict

It is safe to place an agent only when the architecture names which system authorizes the consequential act and which evidence audit will accept. Systems of record, cloud AI platforms, model vendors, and independent gateways each control part of the path. Product breadth does not settle transaction authority.

Map the control path before the vendor

State

Tickets, cases, orders, employee records, and other systems that own the transaction state.

Identity

The user, agent, service, token audience, delegated chain, and current entitlement.

Policy

Purpose, object, action, limits, approval, rate, cost, and default-deny rules.

Execution and evidence

The actual write channel, receipt, result verification, recovery, and accepted audit record.

Owner, briefing, proof

Owner

The enterprise architecture and security owners decide the control point; the business process owner keeps outcome accountability.

Briefing

One cross-system action path, the candidate placements, release status, lock-in, recovery, and the authoritative system for each decision.

Proof

A drill showing authorization, denial, stop, replay, evidence continuity, and recovery across every system and protocol hop.

What leaders should take from it

1
Systems of record retain transaction gravity.

Service, CRM, ERP, HCM, and work-management platforms own process state and increasingly embed agent builders, hubs, and controls.

2
Cloud and identity platforms contest the runtime.

They connect models, agents, data, identities, and tools across applications, but a connected agent is not automatically production-ready.

3
Model vendors supply important execution tools.

Tool use and computer use widen reach. The enterprise still has to define policy, receiver authorization, evidence, and recovery.

4
Independent gateways can narrow access.

They may centralize tool identity, allowlists, and audit across platforms. The evidence does not support calling them the sole control plane.

5
General availability and outcomes are separate claims.

Preview connectors, product announcements, and vendor percentage claims should not be treated as production evidence or comparative superiority.

Where the evidence stops

First-party product pages establish current capability and positioning. They do not provide neutral benchmarks for operating cost, control effectiveness, portability, or customer outcomes. Workday announcement language does not prove historical general availability; Microsoft labels the inspected connections preview; SAP sample percentages are marketing estimates.

First moves before hiring anyone

01
Diagram one irreversible write path.

Name every system, identity, protocol hop, policy decision, approval, action, evidence store, and recovery owner.

02
Choose the authoritative deny.

Decide which layer can refuse the transaction even when an upstream agent requests it with valid credentials.

03
Contract and test portability.

Export policy and evidence, switch one representative path, and measure behavior delta, exceptions, recovery time, and evidence loss.

Where to start

Choose one consequential action that crosses two systems of record. Compare placements on authoritative state, identity, policy enforcement, execution, evidence, recovery, release status, and lock-in.

Claim ledger

8
Checked
source clusters reopened or traced in the September 2026 research pass
0
Fabricated
unsupported source clusters removed from the public claim set
3
Corrected
claims narrowed for source status, scope, or authority
1
Demoted
useful signals kept out of the headline conclusion
ConfirmedServiceNow AI Control Tower: First-party documentation describes discovery, governance, monitoring, and measurement across enterprise AI. It is capability evidence.servicenow.com
CorrectedMicrosoft connected agents: Microsoft documents connections from Copilot Studio to other agent systems and labels the inspected feature public preview and not for production use.microsoft.com
CorrectedWorkday Agent System of Record: Workday positions a registry and governance layer for Workday and third-party agents. Positioning does not establish historical availability or customer outcomes.workday.com
DemotedSAP Joule Agents and AI Agent Hub: SAP describes agents, assistants, and hub capabilities. Its sample percentage is a marketing estimate and stays out of the conclusion.sap.com
ConfirmedAtlassian Rovo agents: Atlassian documents agents inside chat, automation, Jira, Confluence, and Studio. This is first-party capability evidence.atlassian.com
ConfirmedSalesforce Agentforce: Salesforce describes agent building, integration, security, and observability. It is product and marketing evidence, not an independent outcome comparison.salesforce.com
ConfirmedOpenAI computer use: OpenAI documents a harness that executes model-returned UI actions through a screenshot loop. Enterprise policy and authorization remain outside the model call.openai.com
CorrectedAnthropic autonomy research: Anthropic publishes a framework for measuring agent autonomy. It supports separating capability from granted authority, not a claim that one lab owns enterprise governance.anthropic.com
What would change this conclusion
Verified research · 8 source clusters checked · Systems of Action treated as a proposed operating lens