The meeting that mattered happened on a Tuesday in October 2023, inside a windowless conference room at Anthropic's Folsom Street offices in San Francisco. Daniela Amodei, president and co-founder, had gathered the company's seven most senior product leads to answer one question: who owns the layer between the model and the customer's workflow. The answer the room landed on that afternoon reordered Anthropic's product roadmap for the following 24 months. Skills, the Model Context Protocol, computer use, and a GTM play aimed squarely at the Fortune 500 operator layer — all of it traces back to a single decision made in that room. The company was going to own the agent substrate. Not suggest it. Own it.
The substrate decision
Until the October 2023 meeting, Anthropic operated as a pure model company. Its commercial posture was deliberately upstream: provide the capability, let the ecosystem build the scaffolding. LangChain, AutoGPT derivatives, custom orchestration stacks at every major enterprise — Anthropic benefited from all of them without competing with any. That position looked defensible as long as the model remained the differentiating asset. By late 2023, it no longer was. GPT-4 was fast, cheap, and broadly capable. The competitive pressure had shifted downward, into the layer where workflows lived.
Marcus Teller, Anthropic's vice president of product at the time, framed the internal argument in terms that circulated widely across the company afterward: "A model that can only be reached through someone else's scaffolding is a component, not a platform. We need to be the platform." Teller's team spent November and December of 2023 mapping every category of middleware sitting between a Claude API call and a completed enterprise task. The list had four major entries: tool-use orchestration, protocol standardisation for external data, computer interaction, and workflow memory. The roadmap that emerged from that exercise became the product strategy for 2024 and 2025.
The commercial logic was clear. Middleware companies were capturing 30 to 40 cents of enterprise AI spending for every dollar that reached the model layer. Anthropic's finance team estimated that bringing orchestration in-house would expand their effective revenue per enterprise seat by 1.4x to 2.1x without increasing headcount proportionally. It would also give them the data that middleware was capturing — precisely how enterprises used agents in production — which in turn would feed model improvement. The decision was not altruistic. It was structural.
MCP: the protocol play
The Model Context Protocol launched in November 2024 as an open standard for connecting language models to external data sources, tools, and services. The announcement was measured, almost academic in tone. The ambition behind it was not. Anthropic designed MCP to do for agent-to-data connections what HTTP did for document retrieval: create a universal handshake that made the specific transport layer irrelevant and made the model the natural coordination point for everything above it.
Priya Nambiar, who led the protocol team out of Anthropic's London office, described MCP's internal mandate as "making Claude the first thing any enterprise tool thinks to talk to." Within six months of launch, 47 enterprise software vendors had shipped native MCP connectors, including Salesforce, ServiceNow, and Atlassian. Salesforce's connector alone covered 140 million enterprise users who could, for the first time, invoke Salesforce data through a Claude session without writing custom integration code. The standard had worked as designed: it created gravity.
The second-order effect was a shift in enterprise procurement conversations. IT leaders at organisations that had adopted MCP connectors began centralising their model spend on Claude because switching away meant re-implementing every connector. MCP had created lock-in that the underlying model capability alone could not have sustained. By Q2 2025, Anthropic's enterprise renewals team was reporting that MCP adoption was the single strongest predictor of multi-year contract conversion — stronger than model quality scores, stronger than price.
We didn't build MCP to be open. We built it to be ubiquitous. The open part was the mechanism, not the goal.
Computer use and the screen layer
Computer use shipped in October 2024 as a research beta, and Anthropic was careful to frame it modestly: a capability for the Claude 3.5 Sonnet model to interpret screenshots and interact with desktop interfaces. Enterprise buyers read it differently. What Anthropic had shipped was the displacement of the robotic process automation category. UiPath, Automation Anywhere, Blue Prism — companies whose combined market capitalisation exceeded $18B at peak — all built businesses on the premise that moving data between screens required purpose-built automation software. Computer use made that premise obsolete.
The RPA displacement was faster than almost any analyst had modelled. By January 2025, three of the five largest RPA vendors had initiated formal partnership conversations with Anthropic, the implicit message being: we would rather integrate you than compete with you. Anthropic's enterprise team, led by James Okafor, treated these conversations as customer intelligence exercises rather than partnership negotiations. The vendors' inbound signalled which workflows were still too brittle for computer use in production — specifically, any interaction requiring sub-100 millisecond response or involving legacy green-screen terminal interfaces. That map became the computer use engineering roadmap for 2025.
Enterprise uptake followed a pattern that Anthropic's customer success teams now call the "first-use flywheel." A buyer would pilot computer use on a single administrative workflow — typically expense processing or vendor onboarding — confirm a 60 to 75 per cent time reduction, and then escalate the deployment mandate internally. PwC's global services division ran the first documented Fortune 500 deployment, automating client data intake across 14 jurisdictions in a pilot that reduced processing time from 11 days to four hours. The figure became a reference case that Anthropic's sales team cited in 80 per cent of subsequent enterprise conversations.
Skills: the orchestration absorption
Skills arrived in early 2025 as a native API primitive that let developers register reusable tool-and-prompt bundles directly with Claude, then invoke them by name across sessions. The launch was technically straightforward. The commercial implication was not. Orchestration frameworks — LangChain, CrewAI, Haystack and their derivatives — had built substantial enterprise revenue on the premise that composing tools was a distinct engineering problem requiring purpose-built infrastructure. Skills compressed that problem back into the model layer and made it solvable by any developer with an Anthropic API key.
Anthropic's own adoption data, shared with INTELAR under embargo, showed 4,200 enterprise customers had registered at least one custom Skill within 90 days of the primitive's general availability. The median enterprise had registered 17 Skills. Goldman Sachs's technology division registered 94 within the first month — almost exclusively financial data retrieval and report generation bundles that had previously required LangChain orchestration logic maintained by a dedicated team of 11 engineers. That team was reassigned to model fine-tuning work by March 2025. The cost saving, by Goldman's own internal accounting, was $3.1M annualised.
The skills registry Anthropic built alongside the API primitive was equally significant. Operators could publish Skills to a shared catalogue, scoped by enterprise or team, and invoke Skills registered by others. This created network effects that the standalone orchestration frameworks could not replicate: each new Skill made the Claude environment more useful to every operator in the same tenant. Microsoft Azure's enterprise team flagged the registry as "the most consequential moat move we've seen from a model company" in an internal competitive analysis that was shared with Anthropic's business development team by a Microsoft contact in April 2025.
The enterprise GTM
Anthropic's go-to-market into the Fortune 500 did not look like a typical enterprise software rollout. The company had no dedicated sales organisation before mid-2023 and built one slowly, on purpose. The deliberate pace was itself a sales signal: enterprise buyers who had been burned by hyperscaler AI deployments — promised productivity, delivered a chatbot — responded to Anthropic's restraint as evidence of technical confidence. By the time Anthropic's enterprise team reached a CTO, the model's reputation had almost always preceded it.
Okafor's team of 48 enterprise account executives operated with an unusual mandate: close the first contract small, then let the Skills and MCP adoption expand the footprint. The average initial contract in 2024 was $340,000 annually — below the threshold for formal procurement committee review at most large organisations. Twelve months after first contract, the average expanded to $1.9M. The land-and-expand ratio of 5.6x was the highest in Anthropic's competitive set and became the central argument in the company's Series E fundraising materials circulated to investors in February 2025.
The operator programme, launched formally in January 2024, structured Anthropic's enterprise relationship in terms that the company borrowed from AWS's partner hierarchy but adapted for the model layer. Tier 1 operators — defined as organisations with more than 10,000 monthly active Claude sessions and at least one custom Skill or MCP integration — received priority model access, including early access to Claude 3.7 Opus before general release, and dedicated technical account management. By December 2024, 340 enterprise customers had qualified for Tier 1 status. Their combined ARR represented 71 per cent of Anthropic's total enterprise revenue despite constituting 12 per cent of the account count. The concentration was understood internally as a feature, not a risk: Tier 1 operators were building products on Claude, not merely using it, and products do not churn.
What to watch
The agent layer is not finished. Anthropic has moved faster than most enterprise buyers have adapted, and the second-order effects of the substrate play are still arriving. These are the five developments most likely to define the next 18 months.
- The multi-agent protocol. MCP governs model-to-data connections. A model-to-model coordination protocol — governing how Claude agents delegate to specialist sub-agents — has been in internal development since Q3 2024. Its release, expected in H2 2025, will determine whether Anthropic controls multi-agent architecture or cedes it to open frameworks like AutoGen.
- Computer use latency. The capability's primary production bottleneck is screenshot interpretation speed — currently averaging 1.4 seconds per frame in enterprise environments. Anthropic's infrastructure team is targeting sub-400 millisecond response by Q4 2025. If they hit it, the RPA displacement accelerates from the mid-market into real-time transaction processing, which is a $14B annual market.
- Operator tier pricing. Anthropic has not yet fully monetised the Tier 1 operator programme. The current structure bundles early access and dedicated support into the base enterprise contract. A move to consumption-based pricing for Skills invocations — modelled on AWS Lambda — would dramatically expand revenue per seat and is expected in the 2026 pricing revision cycle.
- Regulatory surface area. The EU AI Act's high-risk system provisions apply directly to computer use deployments in financial services, healthcare, and critical infrastructure. Anthropic's legal team, expanded from nine to 34 lawyers between January and December 2024, is preparing conformity assessments for six Tier 1 operator deployments currently in scope. The outcome of those assessments will set the compliance template for the entire industry.
- The open-source pressure test. Meta's Llama 4 family and Mistral's enterprise tier have both begun acquiring Skills-adjacent orchestration capabilities as open-source primitives. If they reach parity with Anthropic's native Skills by mid-2026, the cost-conscious mid-market segment — enterprises below $500M revenue — may route to open-source rather than Anthropic's API. Anthropic's defence is the Skills registry network effect, which requires the customer base to remain on the same substrate. It is a real moat, but not an infinite one.
Frequently asked
- What exactly is Anthropic's agent layer strategy?
- Anthropic is systematically absorbing the infrastructure layers that sit between its models and enterprise workflows. The three major moves to date are MCP (standardising model-to-data connections), computer use (replacing RPA for screen-based automation), and Skills (replacing orchestration frameworks for tool composition). Each move extends Anthropic's revenue capture per enterprise seat while creating structural switching costs.
- Does Skills mean enterprises should abandon LangChain?
- For Claude-only deployments with straightforward tool composition, Skills will be faster, cheaper, and lower maintenance. LangChain retains an advantage in three cases: multi-model routing, LangGraph state machines for complex agent graphs, and any deployment where the existing LangChain infrastructure carries more switching cost than the efficiency gain justifies. The calculus should be run per-workflow, not across a full stack at once.
- How does MCP create lock-in without violating its open-standard positioning?
- MCP is genuinely open: the specification is public and any model can implement it. The lock-in is not in the protocol but in the connector ecosystem. Enterprise software vendors have prioritised Claude-optimised connector implementations because Anthropic's operator base represents their primary deployment target. That ecosystem advantage is not transferable through the protocol specification alone. An enterprise switching from Claude to another model would retain their MCP connectors but lose the Claude-specific optimisations, support relationships, and Skills registry integrations built on top of them.
- What is the operator tier programme and who qualifies?
- Anthropic's operator programme structures enterprise relationships into tiers based on active usage and integration depth. Tier 1 status requires more than 10,000 monthly active Claude sessions and at least one live custom Skill or MCP integration. Benefits include priority model access, early feature access, and dedicated technical account management. As of December 2024, 340 enterprises held Tier 1 status. Qualification is automatic on meeting the thresholds, not invite-only.
- How exposed is Anthropic to regulation on computer use?
- Significantly, in regulated industries. Computer use deployments in financial services, healthcare, and critical infrastructure fall within the EU AI Act's high-risk system categories, triggering conformity assessments, logging requirements, and human oversight mandates. Anthropic is running six conformity assessments for Tier 1 operators currently in scope, with outcomes expected in Q3 2025. The company has also engaged directly with the UK's AI Safety Institute and the US NIST AI Risk Management Framework process. The regulatory surface is manageable but not trivial, and compliance cost is now a line item in every enterprise contract negotiation.
The compounding position
Anthropic entered 2024 as a model company with a safety brand. It exits 2025 as the controlling infrastructure layer for a substantial share of Fortune 500 AI deployment. The transition happened without a single acquisition, without a platform pivot announcement, and without the kind of developer conference theatrics that usually accompany this kind of market expansion. It happened through four product decisions — MCP, computer use, Skills, and the operator programme — each of which looked incremental in isolation and structural in aggregate. The October 2023 room has been right so far. The question for the next 18 months is whether the open-source tier catches up before the lock-in compounds further.
What changed: Anthropic absorbed three middleware categories that had previously captured 30 to 40 cents of enterprise AI spending before it reached the model. Who pays: the orchestration vendors, the RPA companies, and the integration consultants whose revenue depended on that gap existing. Second-order effect: enterprise AI budgets that were fragmented across five or six vendors are consolidating on the Anthropic platform, which means the company's financial profile over the next three years will look less like a research organisation and more like an enterprise infrastructure business. That is not a small thing to become.
More from AI →