The Enterprise Agent Platform Landscape

When the agent market is presented as a logo soup of fifty vendors, leaders freeze. The fix is to hold one durable mental model: the market has two layers, and almost every product sits in one of them.

1. The model layer (the intelligence). A small number of frontier labs — principally Anthropic (Claude) and OpenAI, with Google and a few others — build the raw reasoning engines. This is the "brain" an agent thinks with.
2. The platform layer (the agent + your systems). The large enterprise-software vendors wrap that intelligence in tooling that connects to your data, your applications, and your workflows — and adds the governance, identity, and audit controls an enterprise needs.

The strategic point: the model layer is commoditizing, and value is moving up to the platform layer — to workflow integration and proprietary data. You will rarely "buy a model" directly. You will buy (or build on) a platform that uses a model, and increasingly lets you swap which one.

Inference cost to reach a given capability level fell roughly 280× in about 18 months (~$20 → ~$0.07 per million tokens, late 2022 to late 2024). Source: Stanford HAI, AI Index 2025. The intelligence you buy keeps getting cheaper — a reason not to over-anchor on today's model choice.

At the bottom of the stack sit the frontier labs. For an agent, the model is the reasoning engine — and Anthropic's own framing is useful here: an agent is Model + Tools + Memory + Planning. The model is one of four parts, not the whole product.

Two names dominate the enterprise model + agent layer:

• Anthropic (Claude). Leads enterprise API spend in recent surveys, with particular strength in coding, writing, and a safety/governance posture. Anthropic has also pushed into the agent layer directly with Claude Cowork and an open Agent Skills standard for packaging reusable agent capabilities. (Product names and positioning — verify live.)
• OpenAI. The consumer category leader (ChatGPT) with broad general-purpose reasoning and the widest brand recognition, also building agent tooling on top of its models.

Google, Meta (open-weight Llama), and others round out the layer with multimodal, sovereignty, or cost plays. The durable insight, repeated across the research, is that foundation models are becoming a "strategic commodity" (Gartner). Open-weight models have nearly closed the quality gap, and price keeps collapsing.

Model names, rankings, and prices are the single most volatile thing in this lesson. Recognize that two labs lead the enterprise model + agent layer; verify which and on what terms before you decide.

This is where leaders make real decisions. The major software vendors you already pay are racing to add agent layers on top of their existing platforms — because the platform that already holds your data and sits inside your employees' day has a structural advantage. The market is consolidating around major software vendors plus the model labs.

The single most useful way to read the field is by category — what kind of work the platform is closest to — because that is durable even as product names churn.

Notice the pattern: each vendor's claim is proximity to a kind of work and the data that surrounds it. The decision usually starts not with "best agent" but with "where does the work — and the data — already live?" A Microsoft-and-Teams shop reaches for Copilot Studio; a Salesforce-run revenue org reaches for Agentforce. That proximity often beats abstract model debates.

Every specific name, ranking, and price in this table is volatile. The categories — customer-facing, productivity, workflow/IT, cloud-build, orchestration — are the durable layer. Learn the categories; verify the products live before a procurement decision.

Three roads lead to an enterprise agent, and they carry very different cost, speed, and risk. Most leaders should know which road they are on before signing anything.

The evidence points firmly toward a default-to-buy posture. MIT's NANDA "GenAI Divide" study found that buying from or partnering with specialized vendors succeeded ~67% of the time, versus roughly one-third of that rate for internal builds, and that about 76% of enterprise use cases are bought, not built (Menlo Ventures, 2025). Source: MIT NANDA, State of AI in Business 2025; Menlo Ventures, 2025.

The deciding question is BCG's two-axis build-vs-buy logic: value potential versus competitors × differentiated data access versus vendors. High value and a real data edge → build or own. Low value or no data edge → buy. The middle → partner. Build only where the agent is a genuine source of advantage — everywhere else, buy or build on a platform and spend your energy on integration and adoption.

Here is the part the vendor demos hide: the model is the easy part; reaching your data and your workflows is the hard part — and it is where the value actually lives. An agent that can reason brilliantly but cannot see your customer records, your policies, or your systems of record is a clever intern locked out of the building.

The research is blunt about this. Roughly 80% of the real AI work is data, governance, and workflow integration (MIT Sloan). The reason most enterprise pilots stall is not weak models — MIT NANDA found ~95% of enterprise GenAI pilots delivered no measurable P&L return, and named the root cause the "learning gap": failure to integrate AI into workflows, structures, and culture, not model quality. Source: MIT NANDA, State of AI in Business 2025.

This is exactly why the platform vendors have an edge: Agentforce starts inside Salesforce's customer data; Copilot Studio starts inside your Microsoft 365 estate; ServiceNow starts inside your workflows. They are selling you a shortcut through the hardest part. And it is why proprietary, well-governed data is the one moat that compounds — "AI agents cannot query what they cannot find."

The questions a leader should ask any agent vendor:

• What data must this agent access to be useful, and where does that data go — including to the vendor's own model providers?
• How does it connect to our systems of record, and who owns that integration?
• If we change models or platforms later, what is portable and what is locked in?
• Does our data foundation actually support this — is it clean, governed, and reachable?

The single most useful skill in this fast-moving market is separating the durable from the volatile. The durable layer is worth committing to memory; the volatile layer must be re-checked before any decision — and will be wrong if you memorize it.

The market is consolidating around the major software vendors plus the model labs — that structure is stable. The names inside the boxes are not. A leader who learns the left column and disciplines their team to verify the right column before every procurement decision will navigate this market far better than one who memorizes today's leaderboard.

A note on pricing discipline: recognize the shapes — per-seat subscriptions, usage/token API pricing, and committed cloud + model deals. A common, expensive mistake is buying per-seat licenses for the entire workforce on day one; reports note 30–40% of blanket seats sit unused within 90 days. Start with high-value cohorts; expand on evidence.