MCP & A2A in
HR Technology:
Two open protocols are quietly redrawing the architecture of enterprise HR technology. Model Context Protocol (MCP) and Agent-to-Agent Protocol (A2A) have moved from emerging standards to production infrastructure in under eighteen months. Every major enterprise software vendor has now taken a public position on them — and those positions are not equal.
This paper maps the current state of MCP and A2A adoption across the HR technology and adjacent enterprise platform landscape. It assesses each major vendor's architectural position, assigns an openness classification, and draws out the strategic implications for HR leaders making platform decisions in 2026 and 2027.
The finding is unambiguous: the HR platform market has split into distinct camps. The architectural choices vendors are making now will determine what is possible for enterprise customers for the next five to seven years. This paper gives HR leaders the framework to understand those choices before they commit.
MCP — Model Context Protocol
MCP was developed by Anthropic and released as an open standard in November 2024. In December 2025 it was donated to the Linux Foundation's Agentic AI Foundation (AAIF), co-governed by Anthropic, OpenAI, and Block, with Google, Microsoft, AWS, and Cloudflare as platinum members.
The adoption trajectory is extraordinary. From 100,000 downloads at launch, MCP reached 97 million monthly SDK downloads by March 2026 — a 970-fold increase in eighteen months. As of June 2026, the public MCP server registry contains over 10,000 production servers, and 78% of enterprise AI teams report at least one MCP-backed agent in production.
The analogy: USB-C for AI. Before USB-C, every device required a proprietary connector. MCP creates a universal interface between an AI agent and any external data source or system capability. An enterprise with a Cornerstone MCP server, a ServiceNow MCP server, and a Workday MCP server can connect any MCP-compatible LLM to all three — without custom integration code for each pairing.
A2A — Agent-to-Agent Protocol
A2A was developed by Google and released in April 2025 with over 50 founding partners. By April 2026 — its one-year mark — over 150 organisations supported A2A in production, including Microsoft, AWS, Salesforce, SAP, ServiceNow, Workday, and IBM.
Where MCP solves "how does my AI agent access your data and tools?", A2A solves a different problem: "how do AI agents built by different vendors coordinate with each other?" Together, MCP and A2A form a complete open interoperability stack: MCP connects agents to tools and data (vertical); A2A connects agents to other agents (horizontal). A 2026 CTO survey found that 42% of CTOs plan to run both protocols in production stacks.
Protocol Adoption at a Glance — Q2 2026
| Protocol | Monthly Downloads | Production Servers / Orgs | Enterprise Adoption | Governance |
|---|---|---|---|---|
| MCP | 97M+ SDK downloads | 10,000+ servers | 78% of enterprise AI teams | Linux Foundation AAIF |
| A2A | N/A (protocol standard) | 150+ organisations | 23% of enterprise AI teams | Linux Foundation |
Sources: AAIF Report, VantagePoint Protocol Comparison, April 2026.
The table below summarises each vendor's current position across MCP status, A2A status, access model, and openness rating. Full assessments follow.
| Vendor | MCP | A2A | Access Model | Openness |
|---|---|---|---|---|
| Cornerstone (Workforce AI) | GA | In dev (Salesforce live) | Open — any LLM | Open |
| Oracle (Fusion HCM) | GA | GA | Open — MCP + A2A | Open |
| ServiceNow (Now Platform) | GA | GA | Open — native MCP | Open |
| Salesforce (Agentforce 3) | GA | GA | Open — hosted MCP + A2A | Open |
| Microsoft (M365 Copilot) | GA | GA | Open with governance layer | Open |
| Workday (Agent Gateway) | GA (early access) | GA | Open — MCP + certification | Conditional |
| SAP (Joule / SuccessFactors) | GA (metered gateway) | GA (Joule-routed) | Gated — Joule required | Gated |
| Phenom | Partnership (ServiceNow) | GA via ServiceNow | Partnership-gated | Conditional |
| Eightfold | Partnership (Salesforce) | Not confirmed | Partnership-gated | Not Committed |
| Beamery | Not published | Not confirmed | N/A | Not Committed |
| iCIMS | Proprietary layer | Not confirmed | Proprietary | Not Committed |
Each vendor is assessed against a consistent framework: MCP and A2A status, access model, openness rating, strategic context, and analyst verdict.
Cornerstone launched Cornerstone Workforce AI on May 20, 2026 with MCP and A2A as foundational architecture. The platform exposes its data layer and APIs as MCP servers, meaning any enterprise LLM can connect directly to the Cornerstone People Graph, Skills Engine, and workforce intelligence layer without routing through Cornerstone's own inference engine.
Cornerstone's Chief AI Officer has confirmed on record that they do not require customers to use Cornerstone as the orchestrating agent. Customers bring their own orchestration layer. Cornerstone provides the data and the tools.
Core assets underpinning the MCP layer: the Cornerstone People Graph (45M user profiles, 55,000-skill ontology, 28TB daily labour market data ingestion); the Skills Engine powered by SkyHive (self-cleaning, continuously refreshed); ISO/IEC 42001 certification governing AI management systems and human-in-the-loop requirements.
The Salesforce partnership (announced May 21) demonstrates the A2A direction: a headless integration connecting the People Graph into Agentforce, HR Service, ITSM, and Data Cloud. Pricing model: Fixed fee, headcount-based — no per-call or per-agent charges.
What is not yet resolved: MCP onboarding requires forward-deployed engineers for data pipeline configuration. Self-serve A2A for customer-built agents outside named partner platforms is not yet publicly documented. The H2 2026 timeline for broader A2A availability needs to materialise.
Cornerstone has made the most explicit LLM-agnostic architectural commitment of any HR-native platform in this review. The MCP and A2A foundation is well-documented and supported by public technical materials. The Salesforce partnership is the market's first major A2A production integration in HR. The onboarding friction is real but not disqualifying. The commercial model is aligned with openness.
SAP's position on MCP and A2A is the most consequential in the market. In April 2026, SAP published API Policy v4/2026. Section 2.2.2 prohibits external AI systems from independently sequencing API calls to SAP. From June 9, 2026, a security patch technically enforces this.
The practical effect: a Microsoft Copilot agent, a Salesforce Agentforce agent, or a customer-built agent that wants to act on SAP data must now route through SAP's infrastructure. SAP frames this as governance. Forrester framed it as SAP "attempting to become the gatekeeper of enterprise AI" and published a direct advisory to CIOs to push back.
The adoption gap: Production adoption of Joule among SAP enterprise customers was approximately 3% as of Q1 2026. Meanwhile, 77% of AI-active SAP enterprises use Microsoft Copilot. The policy routes the 97% through a system they are not using to reach data that powers the AI 77% of them are running.
Despite the gatekeeping criticism, the Sapphire announcements were substantively significant: Joule Studio 2.0 GA, over 200 specialised AI agents, 50+ Joule Assistants across HR/finance/procurement, Autonomous HCM reaching GA in June 2026 — all running on Anthropic Claude. SAP committed €100 million to partners building on Joule Studio.
SAP has built a serious AI platform. The semantic enrichment argument for routing through Joule has genuine technical merit. But governance-by-gatekeeping and governance-by-design are different things. The policy restricts the 97% using Copilot to serve the 3% using Joule. SAP's Sapphire implementation partners used the words "walled garden" and "agent tax" to journalists. When your implementation partners are using that language, the governance argument has already lost the room.
Workday's trajectory on open standards is positive and the direction is unambiguous. The DevCon 2026 announcements represent a significant acceleration, with 28,000+ MCP tool calls in a single hackathon demonstrating genuine developer ecosystem momentum.
Agent-Ready Tools: Workday's MCP-based enterprise connectors give AI agents governed access to HR and finance data. These tools carry business context, reduce hallucination, and automatically inherit Workday's security model, business process controls, and audit trail.
Agent Passport: A first-of-its-kind verification framework that tests, validates, and continuously monitors AI agents before they access critical Workday operations — payroll, benefits, ledger. Rather than routing through a proprietary AI layer, Workday certifies the agent.
Google Cloud partnership (May 28): Workday and Google Cloud announced an expanded partnership combining Workday's Agent System of Record with Google's enterprise agent platform. Nick Moores at DevCon: "We are choosing to use open standards, and not building our own bespoke things."
Workday is building the right things and building them correctly. The AgentSkills open standard, the certification model in Agent Passport, and the explicit open-standards commitment at DevCon are serious architectural signals. The execution gap between direction and availability is the primary concern — some capabilities are still in early access. The H2 2026 GA timelines need to hold.
Oracle is the most underreported positive story in the MCP and A2A landscape within the HR analyst community. In March 2026, Oracle launched Oracle Fusion Agentic Applications — embedding over 600 pre-built AI agents directly into Fusion Cloud HCM, ERP, SCM, and CX. The HCM deployment covers workforce scheduling, absence management, payroll, recruiting, and HR service delivery.
MCP implementation: Oracle AI Agent Studio provides a full professional toolchain including MCP integration, debugger, CLI, policy engine, long-term memory, and cross-platform agent interoperability.
A2A implementation: From Release 26A, published Oracle Fusion agents are discoverable and callable by AI agents on other platforms via A2A. Oracle Integration Cloud enforces human-in-the-loop controls and audit requirements across agent interactions.
Oracle's commitment to open standards in Fusion is genuine, well-documented, and production-ready in a way that exceeds what Oracle receives credit for in HR analyst coverage. The 600+ embedded agents, MCP server exposure, and A2A interoperability represent a comprehensive architectural bet. The binding constraint: agentic AI is exclusively available in Fusion Cloud. Organisations on legacy Oracle environments cannot access the layer without migration.
ServiceNow's MCP position is production-ready. The MCP Server Console was described at Knowledge 2026 as "native" — already available. Employee Works, ServiceNow's AI-powered employee experience front door, routes requests across IT, HR, finance, and facilities and explicitly includes MCP and A2A in its Tools layer for multi-vendor agent ecosystems.
Phenom partnership (June 11, 2026): ServiceNow and Phenom announced an A2A and MCP integration embedding Phenom's AI hiring agents into the ServiceNow AI Platform. Hiring managers can autonomously conduct intake meetings, generate job descriptions, source and screen candidates, and move candidates to interview — all within ServiceNow, with Phenom's agents coordinating via A2A.
ServiceNow is building a genuine multi-vendor agent ecosystem using open standards. The native MCP implementation and the Phenom A2A partnership demonstrate both technical capability and partner strategy. Relevant to HR leaders whose workflows intersect significantly with IT service management — a large proportion of enterprise HR.
Three adjacent platforms are directly relevant to HR technology architecture decisions, as they represent the agent runtimes and data layers that HR platforms must integrate with.
Salesforce — Agentforce 3 / AgentExchange
OpenSalesforce has built one of the most comprehensive MCP implementations in enterprise software. Agentforce hosted MCP servers are GA; A2A allows Agentforce agents to coordinate with third-party agents across platforms. The AgentExchange marketplace hosts 1,000+ pre-built agents from 200+ partners. Cornerstone's May 21 partnership — connecting the People Graph into Agentforce, HR Service, ITSM, and Data Cloud — is the most significant current HR expression of this architecture.
Microsoft — Microsoft 365 Copilot / Work IQ
OpenMicrosoft's Work IQ API went GA on June 16, 2026, exposing 10 MCP tools covering email, calendar, Teams, files, SharePoint, and OneDrive. A2A graduated to GA in the May 2026 Copilot Studio update. 77% of AI-active SAP enterprises use Microsoft Copilot — making the SAP Joule A2A integration announced at Sapphire the most immediately relevant MCP/A2A development for the large SAP installed base.
Google — Gemini Enterprise Agent Platform / Agentspace
OpenGoogle rebranded Vertex AI to the Gemini Enterprise Agent Platform at Google Cloud Next 2026, unifying Agentspace into a single product. MCP data stores and A2A agent registration are both in Public Preview. The Agent Garden includes pre-built partner agents from Box, Workday, Salesforce, and ServiceNow. Google is the co-creator and primary developer of A2A.
As of Q2 2026, none of these vendors has published a platform-level MCP server or confirmed A2A support independently of a named partner integration.
Phenom — Openness: Conditional
ConditionalServiceNow A2A + MCP partnership announced June 11, 2026 — the most significant open-standards development from a talent acquisition specialist in this review. No platform-level MCP server published independently. The ServiceNow integration demonstrates A2A is viable for specialist talent platforms; the platform-level MCP story is still to come.
Eightfold AI — Openness: Not Committed
Not CommittedMCP position remains at partnership layer only — Salesforce AgentExchange. No platform-level MCP server. No confirmed A2A commitment. For enterprises building multi-LLM, multi-orchestration stacks, the absence of a platform-level position creates dependency on a single integration pathway.
Beamery — Openness: Not Committed
Not CommittedOperates as an intelligence layer above existing ATS infrastructure. No MCP server published as of Q2 2026. No confirmed A2A position. The platform's strength is integration above Workday, SAP SuccessFactors, and Oracle Taleo — but the agentic interoperability layer has not been publicly committed to.
iCIMS — Openness: Not Committed
Not CommittedIntroduced an "Agent Access Layer" at its 2026 Executive Summit. The architecture is agent-accessible but uses iCIMS's own integration framework rather than the MCP standard. No platform-level MCP or A2A commitment confirmed as of this report.
AI lock-in has a new mechanism
The old model of vendor lock-in was data lock-in: your workforce data in a vendor's system, expensive to migrate. The new mechanism is AI lock-in: a vendor that requires you to use their AI inference layer has locked you in at the intelligence level. Data lock-in is visible and contractually manageable. AI lock-in is harder to see, harder to quantify, and significantly harder to unpick once a multi-year agentic deployment has been built on top of a proprietary layer.
Composability now requires open standards as a prerequisite
The composable HR technology stack has been a credible architectural ambition for several years. But composability is only real if the underlying platforms support open interoperability standards. A stack that claims to be composable but routes AI agent traffic through each vendor's proprietary layer is multi-vendor fragmented, not composable. MCP and A2A are what make a genuinely composable agentic architecture possible.
The ecosystem advantage compounds
Platforms that commit genuinely to MCP and A2A attract more third-party agent builders, more integration partners, and more developer activity. That ecosystem advantage compounds. The HR platform with the richest ecosystem of compatible agents in 2028 will have a structural advantage that cannot be replicated by proprietary re-architecture.
The governance question is real — and solvable without gatekeeping
SAP's argument that routing through a managed layer provides governance is technically valid. But the market has established that governance and openness are not mutually exclusive. Oracle's human-in-the-loop controls, Workday's Agent Passport certification, Microsoft's Rego-based policy engine, and Cornerstone's ISO 42001-certified governance layer all demonstrate that an enterprise can have governed, auditable, agentic AI without routing everything through a single vendor's inference layer.
The liability gap in open architectures is emerging
When an enterprise uses its own LLM via MCP and removes the human-in-the-loop, accountability for AI-driven decisions sits with the enterprise. Vendors including Cornerstone have confirmed that the orchestrating customer owns liability when operating outside the vendor's default governance stack. This is a governance design question that must be resolved before deployment.
Is your MCP server self-serve for our enterprise to connect our own LLM, or does it require routing through your AI layer?
What does A2A agent coordination cost per call, and what is the pricing model for agentic usage at scale?
Can a customer-built agent, running outside any named partner platform, connect directly via A2A?
Where does liability sit when our LLM acts on your data outside your default human-in-the-loop system?
What is your twelve-month roadmap for MCP and A2A, and what is the gap between what is on stage and what is in production today?
The MCP and A2A protocols have completed the transition from emerging standards to production infrastructure in eighteen months. Every major enterprise software vendor has taken a position. Those positions are now stable enough to evaluate.
The HR platform market is consolidating around two broad architectural philosophies. The first: open data layers that allow enterprises to bring their own AI and connect their own agents, with governance built into the data and permission model rather than the routing layer. The second: managed AI layers that aggregate governance, context, and access control in a single vendor-owned inference layer.
Both philosophies have legitimate arguments. The decision that matters is not which is philosophically correct. The decision that matters is which your vendors are building — and whether that matches the AI strategy your organisation is actually executing.
As of June 2026, the data is clear: 78% of enterprise AI teams have at least one MCP-backed agent in production. 77% of AI-active SAP enterprises use Microsoft Copilot, not Joule. The enterprise market has already voted — in deployment decisions — on which architectural direction it prefers. The vendors who read that signal clearly will compound the advantage. The vendors who read it late will face an acceleration problem they cannot catch up with through announcement alone.
© 2026 Elev8 Group. This research may be shared with attribution. Not for commercial redistribution without written permission.
About the Author
Chris Long is an independent HR technology analyst, researcher, and keynote speaker based in Sydney, Australia. He is the founder of Elev8 Group and the author of The Long View, an independent research newsletter with 35,000+ subscribers. Articul8 is the research publication of Elev8 Group, publishing primary analysis on workforce systems and HR technology for operators who take the field seriously.