Enterprise AI vendor comparison: the agentic platforms are converging

The enterprise AI vendor question changed shape sometime in the first half of 2026. For two years the buying conversation was about model capability: whose model scored highest, whose context window was longest, who shipped the newest reasoning mode. That conversation is closing, because the major platforms have converged on the same answer.

By mid-2026 Microsoft, Salesforce, Google, and AWS all ship the same agentic stack: a low-code agent builder, Model Context Protocol (MCP) tool integration, a policy and governance gateway, and observability tracing. The frontier labs that used to sell only model access (OpenAI and Anthropic) have moved up into the same platform layer. When the capability sheets converge, the durable selection criterion moves with them. The question a CIO should now ask is not whose model is best. It is whose accountability surface a buyer can actually audit.

The platforms now ship the same primitives

The convergence is visible in the release notes. Google rebranded Vertex AI as the Gemini Enterprise Agent Platform at Cloud Next on 22 Apr 2026, with an agent builder, a code-first development kit, an Agent Gateway that enforces policy over every tool call, and agent observability with reasoning traces. AWS reached general availability on Amazon Bedrock AgentCore on 13 Oct 2025, a set of modular services covering runtime, gateway, memory, a Cedar-based policy engine (GA Mar 2026), and OpenTelemetry-compatible observability. Microsoft confirmed agent-to-agent communication and remote MCP support generally available in Copilot Studio in May 2026, and shipped Agent 365 on 1 May 2026 as a control plane that inventories and governs agents across vendors. Salesforce documents hosted MCP servers and multi-agent orchestration in Agentforce.

Read those four announcements side by side and the feature lists rhyme. The same four capabilities (build, connect, govern, observe) appear in each, under different brand names.

The strongest single signal is the connective tissue. The Model Context Protocol began as an Anthropic standard; in Dec 2025 Anthropic donated it to a Linux Foundation body, the Agentic AI Foundation, co-signed by OpenAI, Google, Microsoft, and AWS, with more than 10,000 public servers already running and MCP support live in ChatGPT, Gemini, Copilot, and VS Code. When six competitors agree to a shared integration standard and hand its governance to a neutral foundation, they are signalling that the integration layer is no longer where they intend to compete.

The frontier labs moved up-stack

The same convergence pulled OpenAI and Anthropic out of the pure model-API business and into the platform layer the cloud incumbents occupy.

OpenAI shipped AgentKit at its developer event on 6 Oct 2025 (a visual agent builder, a connector registry, and an embeddable chat toolkit), and then announced Frontier, an enterprise agent platform with a managed execution runtime and a forward-deployed professional-services arm, in Feb 2026. Anthropic published the Claude Agent SDK on 29 Sep 2025, generalising the infrastructure behind Claude Code into a toolkit for building production agents, and made MCP the integration spine described above. The labs are no longer only selling tokens. They are selling the systems-integration layer, which we examined separately in frontier labs as systems integrators.

For a buyer this collapses a distinction that used to matter. The choice is no longer “platform vendor or model vendor.” Every serious vendor is now offering a platform.

When capability converges, audit the accountability surface

If the build-connect-govern-observe stack is now table stakes, the differences that remain are in what a vendor will commit to in writing and expose to an auditor. Three are worth scoring directly.

The first is the model-version and deprecation policy. An agent in production is pinned to a model version; when the vendor retires that version, the agent’s behaviour can change. Microsoft’s Azure AI Foundry publishes the most concrete commitment on the public record: a fixed 18-month support window for a generally available model, retirement with at least 60 days’ notice, and a programmatic lifecycle API exposing the retirement date per model. OpenAI commits to at least six months’ notice for GA models and three for specialised variants. Anthropic maintains a dated system-cards index documenting each model’s evaluations and deployment decisions. Salesforce, by contrast, publishes no comparable model-deprecation schedule that we could locate as of 3 Jun 2026, which is itself a finding a procurement team should resolve before signing, not after.

The second is SLA specificity. Most vendors keep agentic-platform SLA terms inside the contract rather than on a public page, so the relevant test is not the headline uptime number but whether the SLA covers the agent runtime, defines a measurable failure, and scales with the autonomous action authority granted. We worked the public terms in the enterprise AI vendor SLA comparison; the short version is that the specificity varies more than the percentages.

The third is compliance documentation. The EU AI Act makes this a regulatory matter, not only a procurement preference. Article 13 obliges providers of high-risk systems to supply deployers with documentation on capabilities, limitations, performance, and log-collection; Article 26 places matching duties on the deployer, including a six-month minimum on log retention. The enforcement clock is contested. The original 2 Aug 2026 high-risk deadline is being pushed back under the 2026 Digital Omnibus, which we tracked in what still applies. The documentation obligation itself is not in doubt; only its start date is. The artefact that satisfies it is an AI Bill of Materials, and the vendors that already publish model cards and lifecycle policies are the ones that can produce it on request.

Scored through our GAUGE lens, these three are where vendors actually separate in 2026. Capability is the entry ticket; the accountability surface is the differentiator.

Why pricing is the wrong axis to decide on

Pricing is where most vendor comparisons start, and it is the axis that ages fastest. The current figures are real and knowable: Microsoft sells Copilot Studio at $200 per 25,000-credit pack and bundled its agent control plane into a $99-per-user Frontier suite that reached GA on 1 May 2026; Salesforce moved to Flex Credits at roughly $0.10 per standard agent action on a rate card dated 21 Apr 2026; AWS meters AgentCore by the vCPU-hour. We compared two of these directly in Agentforce versus Microsoft Copilot pricing.

But our own claim ledger has retracted or revised pricing claims within weeks of publishing them, because a vendor added a tier or shipped a new model version under the same name. A 2026 enterprise AI decision anchored to this month’s price sheet is anchored to the most perishable input available. Price the deployment, but do not select the vendor on it.

What this means for the 2026 vendor decision

For a CIO running a 2026 platform selection, the convergence is good news: it means the build-connect-govern-observe capability is no longer a differentiator to chase, and the evaluation can move to the questions that hold their value.

Ask each vendor for the model-version and deprecation policy in writing, with notice periods. Ask whether the SLA covers the agent runtime and what counts as a failure. Ask for the compliance documentation an EU AI Act deployer would need, whatever the enforcement date settles at. Ask whether the platform speaks MCP, because a shared integration standard is what makes a future migration possible rather than a rebuild. The vendors that answer these cleanly are signalling procurement maturity; the ones that resist are signalling that the deployer will carry the accountability risk on their behalf, the same risk pattern we catalogued in the vendor contract gotchas.

The model race is converging. The accountability race is where the 2026 enterprise decision is actually made.