Anthropic Just Did to SaaS What SaaS Did to On-Premise Software
Most of the coverage of Anthropic’s enterprise strategy has focused on model benchmarks and pricing tiers. That is the wrong frame. What Anthropic has quietly executed is one of the more significant platform power grabs in recent enterprise technology history, and the incumbents who enabled it are only beginning to understand what they agreed to.
The thesis is simple. SaaS companies spent a decade dismantling on-premise software by moving where value and dependency lived. They did not replace the capability. They just repositioned themselves between the user and the underlying infrastructure, commoditised the servers and the IT teams, and captured the margin. Anthropic has done the same thing to SaaS, and the SaaS vendors built the connectors themselves.
What MCP Actually Is
The Model Context Protocol gets described as an open standard for connecting AI tools to external services. That description is technically accurate and strategically misleading.
What MCP actually does in practice is let Claude call a third party service, receive the results, and synthesise them for the user. When Anthropic markets Claude Enterprise as having enterprise search capability, Claude is not doing the search. It is calling Atlassian’s search, or Microsoft’s Graph API, or whatever backend the connector points to. The third party does the retrieval. Claude presents the output.
That distinction matters enormously because of where the credit lands. When the experience is good, the user credits Claude. When it is bad, the attribution is murky. Anthropic is reselling retrieval infrastructure it does not own or operate, presenting it as a native product capability, and the actual infrastructure owners are invisible in the transaction.
Why Every Vendor Had No Real Choice
The coercion in MCP was structural rather than explicit. Anthropic never pressured anyone. They published an open standard and let competitive dynamics do the rest.
The message to every SaaS vendor was implicit but unmistakable. If you do not build an MCP connector, you become invisible to AI-native workflows. Your competitor will build one. Your users will notice. The standard being open is precisely what made adoption frictionless, which is what accelerated the consolidation of power at the AI layer.
The result is that vendors with significant engineering resources and sophisticated technical leadership built the connectors anyway. It was the rational individual decision. Collectively it was a terrible outcome for the SaaS layer as a whole because every integration deepened Claude’s position as the central orchestration layer sitting above their products.
They funded their own displacement. And they did it voluntarily.
The SaaS Parallel
The structural parallel to what SaaS did to on-premise software is exact enough to be uncomfortable if you are running a SaaS business today.
SaaS did not replace the capability that on-premise software provided. Finance teams still needed accounting software. Sales teams still needed pipeline management. The capability stayed constant. What changed was where the dependency lived. SaaS moved it away from the software installation and toward the vendor relationship, the subscription, and increasingly the data that lived in the vendor’s cloud.
On-premise vendors who thought SaaS was just a different delivery model missed that it was a fundamental restructuring of where margin lives. By the time they understood the game had changed, the switching costs had already moved.
Anthropic has executed the same move on SaaS. The capability stays constant. Teams still need project management, documentation, CRM, data analytics. What is changing is where the dependency lives. Users who increasingly rely on Claude to interact with, query, and act on the data inside these tools are migrating their cognitive dependency away from the SaaS application and toward Claude.
Once that behavioural shift completes, whether the underlying data store is Jira or an open source alternative or something purpose built by Claude Code stops mattering in the way it used to. The SaaS application becomes infrastructure. Infrastructure gets commoditised.
Anthropic’s position is arguably stronger than early SaaS vendors had because they do not even need to build and maintain the application layer. The partners do that. Anthropic owns the reasoning layer and lets everyone else maintain the plumbing.
Atlassian Is the Biggest Loser
Atlassian had the perfect strategic position to own enterprise AI and chose not to take it.
They already had the knowledge graph of how organisations actually work. Confluence held institutional memory. Jira held execution state. They were embedded across engineering, product, and operations at most serious technology companies. They had a living map of organisational intelligence that any AI company would have spent years trying to acquire.
The opportunity was to become the organisational intelligence layer. Not just storing what teams know and do, but reasoning across it, surfacing it, and acting on it. Notion saw a version of this opportunity and built toward it, treating the document as a database, the database as a workflow, and the workflow as something AI could operate across.
Atlassian instead treated Confluence as a wiki and Jira as a ticket system and kept adding features to a product mental model that was formed in 2005. They are now a structured data store that Claude queries.
The replacement risk they face is more severe than most analysis acknowledges. Historically Jira and Confluence carried enormous switching costs. Not because of technical superiority but because migration was painful, institutional knowledge was buried in the tooling, and nobody wanted to rebuild workflows from scratch.
Claude Code removes all three of those switching costs simultaneously. Migrating an entire Jira instance to an alternative is now a legitimate Claude Code project rather than a six month professional services engagement. The open source alternatives, Plane for project management and Outline for documentation, are genuinely capable. The only barrier was setup and maintenance complexity. That barrier is largely gone.
More fundamentally, a team can now simply ask Claude Code to build them a project management tool built precisely to their requirements, self-hosted, with no licensing cost, maintained through natural language instructions. The replacement is not migration to a competitor. It is replacement with something that did not exist as a practical option eighteen months ago.
By making their data accessible via MCP, Atlassian has given Claude a complete map of everything stored in their systems. That is exactly the information needed to migrate away from them cleanly. They handed over the blueprint.
Microsoft Is Complicated
Microsoft’s position is harder to summarise because they are winning and losing simultaneously depending on which product you look at.
The Excel and PowerPoint moats are real and durable in ways that most AI analysis underestimates. The threat to those products is not that Claude replaces them. It is subtler and more dangerous than that. Anthropic releasing a Claude addin for Excel and PowerPoint looks like a partnership. The actual strategic logic is that Claude inserts itself as the cognitive layer on top of applications that have thirty years of human muscle memory behind them.
Users keep using Excel. But increasingly Claude understands the data, writes the formulas, builds the models, and interprets the outputs. The human’s dependency quietly migrates from Excel to Claude. Excel becomes the rendering engine, the thing that holds the grid, while Claude owns the thinking. Once that behavioural shift completes, whether the grid is Excel or Google Sheets or something else stops mattering. Anthropic has used Microsoft’s own distribution to detach users from Microsoft’s core stickiness. That is a genuinely aggressive move dressed as a productivity feature.
Microsoft’s own execution has made this worse. The Copilot rollout was one of the more damaging own goals in recent enterprise technology history. They attached the Copilot brand to every product in their portfolio before any of it worked properly. Teams Copilot, Word Copilot, Security Copilot, Azure Copilot, Dynamics Copilot. Enterprise sales teams were selling a vision the product could not yet deliver. CIOs bought licences, switched it on, got underwhelming results, and Copilot became a thirty dollar per user per month punchline inside most organisations that tried it.
The SharePoint activation decision compounded the damage. SharePoint is famously where organisational knowledge goes to die. Decades of poorly structured, ungoverned, duplicated content sitting in folder hierarchies nobody maintains. Pointing an AI at that and calling it enterprise intelligence was never going to produce good results. The garbage in garbage out problem was entirely predictable. They ignored it and shipped anyway.
The strategy that was available to them was almost embarrassingly simple in hindsight. Take OpenAI’s best model, make it the engine, build genuinely excellent connectors for Microsoft 365, go deep on Excel and PowerPoint where AI assistance has obvious daily value for every enterprise user, build everything else AI-native from scratch rather than retrofitting Copilot onto legacy surfaces, and price basic functionality at zero for existing enterprise clients. Absorb the cost, embed the behaviour, create the dependency, and monetise once the habit is unbreakable.
They had Teams, Outlook, Excel, and PowerPoint inside essentially every large organisation on earth. The distribution advantage was extraordinary. Instead they charged a premium before the product deserved it, generated widespread rejection, and now face the considerably harder task of re-convincing buyers who already evaluated and dismissed them.
Dynamics 365 is likely the next product they lose through the same pattern. Dynamics was always a second tier ERP and CRM that won on Microsoft relationship and bundling rather than product merit. It has always suffered from being a collection of acquired products, Navision, Axapta, Great Plains, that were never truly unified. The data models do not naturally share context. Workflows between finance, sales, and operations require heavy implementation work. Microsoft’s response has been to add Copilot features to paper over the seams rather than fix the underlying architecture.
The modern replacement stack for something like Dynamics is now a purpose built database layer, clean APIs, and Claude handling workflow orchestration, data synthesis, and user interaction. The implementation complexity that made switching feel risky is the same complexity that Claude Code is systematically eliminating.
Google Read the Room
Google is the one major incumbent that correctly diagnosed what MCP actually meant for them and built around it rather than into it.
The clearest example is Gmail. Google built MCP connectors for Gmail and Google Workspace. They participated in the standard. They did not block it or ignore it. But connect Claude to Gmail via MCP and the experience is noticeably limited compared to working inside Google’s own surface. The threading, the full conversation history, the deep integration with Calendar and Drive, the search quality across your entire mail history, none of that comes through the MCP layer with the same richness. The connector exists. It is just not very good.
That is not an accident. Google controls exactly what the connector exposes and has every reason to ensure the really useful context stays native to their own products. Meanwhile Gemini inside Gmail has full access to everything. The comparison is not even close, and Google engineered the gap deliberately.
This is a more sophisticated move than simply refusing to participate. Refusing to build MCP connectors would have looked obstructionist and driven enterprise customers to ask uncomfortable questions. Building connectors that are just good enough to avoid that conversation, whilst keeping the genuinely valuable integration experience locked inside their own ecosystem, is the smarter play. It lets Google say yes to MCP whilst making sure yes does not actually cost them anything.
Their broader defensive strategy is coherent for the same reason. Gemini is deeply embedded in Workspace, in Docs, Sheets, Gmail, and Drive. Google owns the file layer, the identity layer, and the collaboration layer simultaneously in a way that is genuinely difficult to route around. Simplified enterprise pricing reduces the CFO conversation that might otherwise trigger a migration evaluation. Gemini improving rapidly means the quality gap that might push enterprise users toward Claude-centric workflows narrows over time.
The core insight Google is operating from is that they cannot win by being the backend for Claude’s ecosystem. They can only win by making their own ecosystem so integrated and frictionless that adopting a Claude-centric workflow has a real organisational cost. They are the one incumbent that understood that MCP participation is not the same as MCP commitment, and acted accordingly.
What This Means If You Are Buying Enterprise Software Today
The switching cost calculus for enterprise software has changed in ways that most procurement processes have not caught up with yet.
The complexity that historically justified paying premium SaaS pricing, the implementation risk, the migration cost, the maintenance burden, is the specific complexity that AI is eliminating fastest. That means the risk premium built into staying with an established vendor is eroding at the same time as the cost of evaluating alternatives is falling.
Data infrastructure is where this plays out most visibly right now and Microsoft’s Fabric is the clearest example of a product that should not exist in its current form.
Fabric was always a questionable proposition. A sprawling PaaS that attempted to unify data engineering, warehousing, and analytics into one Microsoft-billed surface. The integration convenience it sold was real but expensive, opinionated, and deeply tied to the Microsoft estate. Then they added Copilot and rebranded parts of it as Fabric IQ, which is AI washing on top of a platform that was already struggling to justify its complexity.
The actual replacement stack is straightforward. Take a ClickHouse SaaS instance, which is cheaper to run, faster for analytical workloads, and operationally simple. Connect it via MCP. Let Claude handle the querying, the interpretation, and the insight layer. The complexity that Fabric was charging you to manage disappears, and you are not locked into Microsoft’s data estate to access it.
Snowflake appears to have understood the direction of travel. Snowflake Intelligence is a genuine attempt to put a chat and reasoning layer directly on top of your data, removing the abstraction complexity rather than adding to it. Snowflake Code takes that further by tackling the data engineering plumbing itself, the pipelines, the transformations, the infrastructure work that has always been the expensive and brittle part of running a serious data platform. That is the right problem to solve. Make the hard stuff disappear rather than building new interfaces on top of it.
Databricks has gone the opposite direction and it is going to cost them. Rather than simplifying, they have kept shipping new features and new complexity, including new ways to build BI tooling. That is exactly the wrong bet. BI as a category is being hollowed out in real time. The entire premise of BI, that you need a specialised tool and trained analysts to surface insights from data, falls apart when anyone in the organisation can ask Claude a question and get an answer directly from the underlying data. Building new BI features in 2025 is the Fabric Copilot mistake made twice. It is adding a layer of complexity to solve a problem that the AI layer is eliminating from the other direction.
For project management and documentation, the question worth asking honestly is how much of the switching cost is real technical complexity versus accumulated organisational inertia. Claude Code has changed that answer significantly in the past year.
The businesses that figure this out before their competitors will carry a structural cost and flexibility advantage that compounds over time. The businesses that stay locked into legacy SaaS pricing out of habit will eventually face the same reckoning the on-premise vendors did when SaaS matured. By that point the window to move cheaply will have closed.
The One Sentence Version
Anthropic got the SaaS layer to build the connectors that make the SaaS layer optional, and most of the companies that built those connectors are still describing it as a partnership.
Top comments (0)