The idea that generative AI would weaken the traditional software suite was persuasive because the interface appeared capable of replacing so much of it. Instead of opening a separate application for finance, research, coding or customer support, an employee could describe an objective to a model, give it access to the necessary systems and allow it to complete the work.
Anthropic’s public IPO filing suggests that a different structure is emerging. Individual applications may become less visible, while the AI provider sitting above them becomes considerably more powerful.
As of September 30, Anthropic had more than 11,900 customers generating at least $100,000 in annual revenue, including more than 1,900 generating at least $1 million and approximately 250 generating $10 million or more. More than half of the customers spending at least $100,000 now use both the Claude developer platform and a first-party application such as Claude Code or Cowork, up from approximately 30% at the end of 2025.
The highest-spending 5% of Claude Enterprise customers using Claude Code spend approximately twenty times more per month than the average enterprise customer that does not use the coding product. Anthropic also reported net revenue retention of approximately 900% at the end of the third quarter, driven by customers expanding their use of Claude across more parts of the organization. Taken together, these numbers reveal how Anthropic intends to convert a model into an enterprise ecosystem.
From model access to organizational infrastructure
Anthropic describes its business as three connected layers. Claude Platform gives developers the models and tools required to build their own products, Claude Code brings the company directly into software development, and Claude Cowork extends its reach into research, analysis and other forms of knowledge work.
An improvement at the model level can make every application more capable. Wider adoption of the applications increases model consumption and gives Anthropic more opportunities to introduce adjacent products. Customers entering through the API can expand into Claude Code, while organizations adopting Claude Code can introduce Cowork to employees outside engineering.
This structure creates legitimate value. A company using the same provider across several workloads can reduce integration work, consolidate administration and apply common security controls. Employees encounter a consistent interface, while technical teams can reuse integrations, evaluations and deployment knowledge. Anthropic’s filing argues that this integrated approach allows improvements at one layer to benefit the others.
It also creates a familiar commercial dynamic. The provider enters through one workload, becomes integrated with the customer’s systems and expands into adjacent departments. Each new product increases the value of the existing relationship and raises the cost of replacing it.
The expansion no longer depends solely on software distribution. Anthropic has internal Applied AI teams that work with customers to place Claude inside existing technology stacks. It has also partnered with private equity firm Hg to create a European consultancy that will embed engineers and consultants inside businesses, beginning with companies in Hg’s portfolio.
Consulting gives Anthropic a way to close the distance between model capability and operational adoption. It also allows the company to participate directly in designing the workflows, integrations and internal applications through which customers use Claude.
The engineers implementing those systems are creating business value. They are also creating dependency – once a provider’s models, applications, agent skills and implementation teams become involved in how work is organized, replacing that provider becomes a transformation project rather than an API change.
The enterprise vendors are moving from the opposite direction
AI laboratories are expanding upward from models into applications and implementation. Enterprise-software companies are moving toward the same position from the systems beneath them.
Oracle and OpenAI announced that Oracle applications and enterprise data will become accessible through ChatGPT. Employees will be able to describe a business objective in a conversational interface while Oracle AI Workspace retrieves information and executes approved actions across Fusion Applications, NetSuite, Oracle AI Database and Oracle Cloud Infrastructure.
A finance employee, for example, could ask ChatGPT to investigate costs, reconstruct purchasing decisions and identify outstanding work before the monthly close. The interface would bring together information from finance, procurement and planning systems without requiring the employee to navigate each application separately.
Oracle contributes the data, transactions, permissions and established enterprise relationships. OpenAI contributes the model and the interface through which the employee expresses intent. Oracle also reported 130,000 active ChatGPT users and more than 95,000 active Codex users inside its own organization, giving the partnership an internal deployment story alongside the commercial announcement.
SAP is pursuing its own version of this architecture. At TechEd, the company presented an autonomous-enterprise stack combining agents with SAP applications, business data, knowledge graphs, domain models and governance. Its advantage comes from the context already encoded inside systems that manage finance, supply chains, procurement and human resources.
SAP does not need to persuade customers that these systems belong in critical operations. They are already there. Its task is to make them accessible to agents without surrendering control of the data and transactional authority they contain.
The AI providers and enterprise-software incumbents are therefore approaching the same strategic position from opposite directions. The laboratories have intelligence and increasingly influential interfaces, but they need access to enterprise data and permissions. The incumbent vendors already possess data, workflows and authority, but they need models capable of reasoning across them.
The most important competition may concern which layer coordinates the others.
Lock-in is changing shape
Traditional enterprise lock-in was relatively visible. It came from proprietary file formats, difficult data migrations, employee training, licensing agreements and workflows built around one vendor’s product.
AI-era dependency can accumulate across a wider set of technical and organizational decisions. A company may develop model-specific prompts, evaluation sets, agent instructions, safety policies and tool integrations. Employees begin to rely on the behavior of a particular model, agents accumulate access to internal systems, and conversations and completed tasks create organizational memory inside the provider’s environment.
The company may still own its underlying data, but the operational meaning of that data increasingly depends on the context and policies assembled around the AI system.
This concern extends beyond Anthropic. A Gartner analysis of vendor-agnostic AI strategy argues that enterprises should keep machine-readable context accessible across platforms and develop interoperable metadata layers. Separate Gartner research warns that deploying multi-agent systems from a single large provider can constrain future flexibility as agents become embedded across operations.
An IBM Institute for Business Value survey of 1,000 senior executives found that 71% believed switching their main AI vendor or model would be difficult. Ninety-one percent said they did not fully understand their dependencies across vendors, models and infrastructure, while 81% expected a seven-day outage at a major provider to cause severe or critical disruption.
The figures come from vendor-sponsored research and should be read in that context. The operational problem they describe is nevertheless credible. AI dependencies can spread through an organization without appearing in a conventional software inventory. One model may be accessed directly, embedded inside several SaaS products and called by internal agents through third-party infrastructure. A change in price, availability, policy or model behavior can therefore affect several workflows at once.
Anthropic acknowledges a related dependency from the provider side. Amazon, Google and Microsoft are simultaneously investors, infrastructure suppliers, distribution partners and competitors. The filing warns that these companies could prioritize their own capacity or competing products, making it harder for Anthropic to secure compute on favorable terms.
The stack may be modular at the technical level while remaining deeply interconnected at the commercial one.
Open standards provide a counterweight
The return of the suite is not inevitable. Enterprises have more options for preserving portability than they did during previous software cycles.
The Model Context Protocol provides an open method for connecting models with external data and tools. Its adoption by model providers, developer platforms and enterprise-software companies can reduce the need to build a separate connector for every model and application. Organizations can also route workloads across several providers, use open-weight models in controlled environments and retain their systems of record outside the model layer.
SAP offers an instructive example. Its CFO told Reuters that the company integrates more than 70 models from different providers and increasingly receives questions about model origin, data residency and geopolitical exposure. SAP’s provider-neutral approach allows a customer to select different models according to workload, jurisdiction and security requirements.
The flexibility is valuable, although interoperability protocols solve only part of the problem. MCP can standardize how a model connects to a tool. It does not automatically transfer the model-specific evaluations, accumulated memory, operating policies or employee habits surrounding that connection. Two providers may call the same application through the same protocol while interpreting the task differently.
Model portability and organizational portability are therefore separate problems. The first concerns whether another model can access the same tools and data. The second concerns whether the company can preserve the behavior, controls and institutional knowledge built around the previous provider.
The architecture decision has moved above the application
For CTOs, the response should extend beyond avoiding long-term commitments. Deep integration can be justified when it produces better outcomes, reduces implementation time or gives employees access to capabilities they would otherwise struggle to use. Attempting to eliminate every dependency can leave an organization with an expensive abstraction layer and weaker products.
The more useful objective is to decide deliberately which assets the company must control.
Business context, evaluation data, access policies and audit history should remain available outside one model provider wherever practical. Agent permissions should derive from the organization’s identity and authorization systems. Workflows with significant operational consequences should expose the models, tools and approvals involved in each action. Applications should distinguish the provider responsible for reasoning from the system responsible for recording the transaction.
A multi-model strategy also requires more than contracting with several vendors. Companies need a credible method for evaluating which model should handle each workload and for moving that workload when cost, capability or regulatory requirements change. Without shared evaluations and portable context, multiple contracts create supplier diversity without operational flexibility.
Anthropic’s filing shows why these decisions cannot wait until switching becomes necessary. The company’s strategy is designed to expand organically from one successful use case into several others. From Anthropic’s perspective, that is efficient distribution and deeper customer value. From the customer’s perspective, each expansion changes the architecture of dependency.
The suite is returning around intelligence
Generative AI was expected to weaken enterprise suites because a conversational layer could replace many application interfaces. That much may still happen. Employees may interact with fewer dashboards and move through fewer menus as agents coordinate work across systems.
The commercial structure behind those interactions is beginning to look familiar. Anthropic is combining models, a developer platform, coding software, knowledge-work applications, cloud distribution and implementation services; OpenAI is bringing enterprise applications into ChatGPT; Oracle and SAP are reorganizing their products around agents that can act across the systems they already control.
That layer can become more consequential than the individual applications beneath it because it mediates how employees reach them. AI may make individual applications less visible while making the provider behind them considerably more powerful. The suite is returning, rebuilt around intelligence rather than software modules.
