The End of App-Switching: AI Becomes the Interface to Work

Most office work is not thinking time. It is context-switching time.

The average knowledge worker does not spend their day in one application, doing one kind of work, in one sustained flow. They spend it bouncing — from email to a project management tool, from the project management tool to a spreadsheet, from the spreadsheet to a messaging app to ask a colleague about a number, from the messaging app back to the spreadsheet, and then to a slide deck to paste the updated chart. Each switch takes seconds. Each recovery takes minutes. The cumulative cost is staggering: researchers at the University of California, Irvine found that it takes an average of 23 minutes and 15 seconds to fully refocus after a task interruption. And a Harvard Business Review study tracking 20 teams across three Fortune 500 companies found that the average knowledge worker toggles between applications roughly 1,200 times per day — each switch costing just over 2 seconds, adding up to nearly 4 hours per week spent reorienting. That is 9% of annual work time — roughly five working weeks per year — lost to the act of switching, not the work itself. In one extreme case from the study, supply-chain employees toggled 3,600 times per day, with a single transaction requiring 350 toggles across 22 different applications.

This is not a technology problem. It is an architecture problem. The applications are fine individually. The problem is that the human is the integration layer between them. Every time you copy a number from one tool and paste it into another, every time you re-explain context that exists in a different system, every time you manually trigger a workflow that spans three platforms — you are doing the work that software should be doing for you. You are the API.

That architecture is changing. AI is becoming the control plane over work software — a layer that sits above individual applications, understands context across them, and executes tasks that span multiple tools without requiring the human to be the bridge. OpenAI is building it through ChatGPT's app directory and connected project sources. Microsoft is building it through Copilot's agent-first architecture and cross-application orchestration. Anthropic is building it through the Model Context Protocol, which lets AI connect to any tool through a standardized interface. The convergence is unmistakable: the chat window is becoming the operating layer for work.

The Hidden Cost of Being the Human API

The productivity cost of app-switching has been studied extensively, and the numbers are worse than most people assume.

Gloria Mark, a professor at UC Irvine who has spent over two decades studying attention in the workplace, documented in her 2023 book "Attention Span" that the average attention span on any single screen has dropped from 2.5 minutes in 2004 to 75 seconds in 2012 to just 47 seconds by 2020 — a decline of over 68% in sixteen years. Her work shows that people self-interrupt 49% of the time — nearly half of all context switches are not caused by external notifications but by the worker's own restlessness. And when people are interrupted during a task, they experience "attention residue" — a cognitive penalty first identified by researcher Sophie Leroy at the University of Washington, where part of the mind remains on the previous task even after switching. The result is that people work faster after switching but produce lower-quality work and experience measurably higher physiological stress. The American Psychological Association found that task-switching can reduce productive output by as much as 40%.

The enterprise tool landscape makes this worse. Okta's 2025 Businesses at Work report found that the average company now uses 101 SaaS applications — the first time the number has crossed 100. Zylo's SaaS Management Index puts the number even higher at 275, with organizations spending an average of $49 million annually on SaaS — $4,830 per employee — and only 49% of provisioned licenses actually being used. The average company runs 15 online training apps, 11 project management tools, and 10 collaboration platforms, with a third of all applications qualifying as shadow IT that the IT department does not even know exists. Each tool has its own interface, its own notification system, its own data model, and its own context. None of them share state by default. The human worker is expected to maintain coherence across all of them — remembering which information lives where, translating between formats, and keeping the full picture of a project in their head when no single tool holds it.

A study by Qatalog and Cornell University quantified this as the "toggle tax" — the cognitive and time cost of switching between tools. They found that 45% of workers say context switching makes them less productive, 43% say it is mentally exhausting, and the average worker takes 9.5 minutes to regain productive workflow after toggling to a different application. Employees spend 59 minutes each day just searching for information scattered across fragmented systems. Slack's 2023 State of Work survey of 18,000 desk workers confirmed the pattern: 68% spend at least 30 minutes per day toggling between tools, and knowledge workers spend 60% of their day on "work about work" — status updates, tool management, and chasing information rather than doing the work itself.

The problem has gotten worse, not better, as more tools have been added. Microsoft's 2025 Work Trend Index found that workers are now interrupted every 2 minutes during core work hours — 275 times per day — by meetings, emails, and notifications. The average worker receives 117 emails and 153 Teams messages per weekday, and 68% say they do not have enough uninterrupted focus time. Workers need roughly 20 hours per week of focus time but get only about 11. And the irony of AI adoption is that most organizations now have their primary productivity suite, a separate AI chat tool, maybe an AI coding assistant, an AI note-taker, and specialized AI tools for specific departments — each one another tab, another context, another place where information lives but does not connect. Tools designed to reduce work are creating more switching.

What a "Control Plane" Means for Work Software

In infrastructure engineering, a control plane is the layer that manages and coordinates the components below it. In a network, the control plane decides where traffic should go while the data plane actually moves the packets. In Kubernetes, the control plane manages containers — scheduling them, monitoring them, restarting them when they fail — while the containers do the actual computing. The control plane does not replace the components. It orchestrates them.

This is precisely what AI is becoming for work software. The applications — your CRM, your project manager, your email, your spreadsheets, your documents — are the components. They still store the data, run the logic, and provide the specialized interfaces for their specific domains. But the AI layer sits above them, reading context from all of them, reasoning about tasks that span multiple systems, and executing actions across tools without requiring the human to manually bridge the gaps.

The control plane metaphor matters because it clarifies what AI is not doing. It is not replacing your applications. It is not trying to be a worse version of your CRM or a worse version of your spreadsheet. It is providing the orchestration layer that those applications have always lacked — the ability to coordinate work across systems in a way that previously required a human (or a fragile chain of automations) to manage.

Three Convergent Patterns

Three major AI platforms are building this control plane, and while their approaches differ, they are converging on the same destination.

OpenAI: The Chat Hub. ChatGPT has evolved from a question-and-answer tool into a work coordination platform. The introduction of Project Sources — which connect ChatGPT to Slack channels, Google Drive folders, and other data sources — means the AI can maintain persistent context from the tools where your work actually lives. You do not need to copy-paste information into the chat window. The AI reads it from the source. The ChatGPT app directory extends this further: rather than switching to a separate tool for a specific task, third-party capabilities come to you inside the conversation. And with Codex — OpenAI's autonomous coding agent launched in May 2025 — the pattern extends to engineering workflows, where the AI moves between codebase, terminal, and development environment without the developer switching contexts. ChatGPT now serves over 800 million weekly active users, with over 1 million business customers and 9 million paying enterprise seats — a 4x increase from September 2025. The ChatGPT app directory, launched in December 2025 and built on MCP, brings third-party capabilities from partners like Figma, Canva, and Salesforce directly into the conversation. Connectors shipped in June 2025 link to SharePoint, Google Drive, Gmail, Slack, Asana, GitHub, HubSpot, and more. The trajectory is clear: the conversation is becoming the workspace.

Microsoft: The Embedded Agent. Microsoft's approach is different — instead of building a new hub, it is embedding AI agents into the tools people already use. Microsoft 365 Copilot operates across Word, Excel, PowerPoint, Outlook, and Teams as a unified AI layer, with BizChat serving as the cross-application interface where you can ask questions and execute tasks that span the entire Microsoft 365 ecosystem. The agent-first architecture means Copilot agents can be built to automate specific workflows — processing invoices that arrive in Outlook, updating data in Excel, and generating a summary in Teams — all without the user leaving their current application. Outlook is increasingly becoming a command center where agents handle tasks across the entire productivity stack. At Ignite 2025, Microsoft introduced Agent 365 — explicitly described as the "control plane for agents" — enabling organizations to manage and secure agents across the enterprise. OneDrive agents can understand an entire set of related documents, plans, and meeting notes. Copilot in Outlook mobile summarizes unread emails and guides actions through voice. With 15 million paid Copilot seats across 450 million Microsoft 365 commercial subscribers and 82% of leaders saying 2025 is a pivotal year to rethink operations with AI, the embedded approach reaches the largest installed base of any AI productivity tool.

Anthropic: The Protocol Layer. Anthropic has taken the most infrastructure-oriented approach with the Model Context Protocol (MCP). Rather than building integrations into a specific product, MCP is an open standard that lets any AI model connect to any tool through a standardized interface. The protocol has seen extraordinary adoption since its launch in November 2024: over 97 million monthly SDK downloads, more than 10,000 active public MCP servers, and adoption by every major AI provider — OpenAI in March 2025, Google in April 2025, Microsoft by year-end. In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, co-founded with OpenAI and Block, ensuring the standard remains vendor-neutral. Through Claude's Connectors Directory — now spanning 50+ curated integrations including Google Workspace, Slack, Figma, Asana, and DocuSign — and the Cowork platform, users can orchestrate work across tools without leaving the conversation. Anthropic now commands 40% of the enterprise LLM API market and serves over 300,000 business customers. The protocol approach means the control plane is not locked to a single vendor — any AI system that speaks MCP can orchestrate any tool that supports it.

And these three are not alone. Google launched the Workspace CLI in March 2026, which starts an MCP server exposing Gmail, Docs, Sheets, and Drive as structured tools — with 100+ pre-built agent skills. Zapier, which connects to 8,000+ apps, reports that 72% of enterprises now have AI agents deployed and operating autonomously. The convergence is unmistakable. Despite different strategies — hub, embedded, protocol, automation platform — all are building toward the same outcome: the human no longer needs to be the router between applications.

Why Retrieval Inside the Flow Matters

The key capability that makes AI a control plane rather than just another app is retrieval — the ability to pull relevant information from connected systems in real time, without the user having to specify where it lives or how to find it.

The cost of this manual retrieval is estimated at $450 billion annually in the United States alone — lost productivity from workers navigating between tools to assemble information that should flow automatically. In the old model, the human is the retrieval engine. You know the quarterly numbers are in an Excel file on SharePoint. You know the project timeline is in Asana. You know the client's latest email is in Outlook. You navigate to each one, find what you need, and assemble the picture in your head (or in a slide deck). This works when you have three sources. It breaks down when you have thirty.

In the AI control plane model, the retrieval happens inside the conversation. You ask the AI to prepare a project update, and it pulls the latest task status from your project manager, the budget figures from your spreadsheet, and the key decisions from recent meeting notes. The AI does the navigation. You do the thinking.

This is why connected context — Project Sources in ChatGPT, Microsoft Graph in Copilot, MCP connectors in Claude — is the foundational capability of the control plane. Without retrieval, the AI is just another tool you have to feed information to. With retrieval, the AI becomes the layer that connects all your other tools.

The shift changes what "using AI at work" means. It stops being about prompting a chatbot and starts being about directing an agent that has access to your full work context. The prompt evolves from "write me an email about the project status" (where you then paste in the status) to "send the project team an update based on this week's progress" (where the AI knows what the progress is because it can read it from connected systems).

What This Means for SaaS Product Design

The rise of the AI control plane creates an existential question for SaaS companies: if users interact with your product through an AI layer rather than your interface, what is your product?

For two decades, SaaS companies competed on user experience. The interface was the product. Salesforce won because its interface made CRM manageable. Slack won because its interface made messaging delightful. Notion won because its interface made documentation flexible. The business model depended on users spending time inside the application — because time in the app meant engagement, retention, and expansion.

The AI control plane threatens this model by making the interface less relevant. If a user can query their CRM, send a message, and update their documentation through a single AI conversation — without ever opening those individual applications — then the SaaS company's interface becomes a backend. The application is still storing the data and running the logic, but the user interaction happens somewhere else.

The disruption is already measurable. Enterprise AI investment tripled from $11.5 billion to $37 billion in a single year. Gartner predicts that generative AI and AI agents will create the first true challenge to mainstream productivity tools in 35 years, prompting a $58 billion market shake-up. Seat-based SaaS pricing has declined from 21% to 15% of market share in one year as usage-based and outcome-based models surge. IDC predicts that by 2028, pure seat-based pricing will be obsolete, with 70% of software vendors refactoring pricing around consumption or outcomes. And 80% of enterprise buyers now cite AI-driven commoditization as the number-one risk to SaaS valuations.

This is why every major SaaS company is racing to embed AI directly into their product. Salesforce has Agentforce — already running 3 billion automated workflows monthly across 18,500 enterprise customers for $540 million in ARR. Slack has AI-powered agents where 95% of users say using an app inside Slack makes those tools more valuable. Notion 3.0 introduced autonomous agents that can do 20 minutes of independent work at a time across hundreds of pages, connected to 70+ applications via MCP. HubSpot has Breeze. The strategy is the same: if you cannot stop the AI layer from abstracting your interface, make sure your product is the best-in-class service behind that layer — and give users enough AI capability inside your product that they do not need the external layer.

The SaaS companies that thrive will be the ones that treat their API as their primary product and their interface as one of many frontends. The ones that struggle will be the ones that assumed the interface was the moat.

The Risks of Too Much Abstraction

The AI control plane is not without risks. Abstraction always trades direct control for convenience, and the trade is not always worth it.

Loss of direct manipulation. When you work inside a specialized application, you see the data, the structure, the relationships. You can click into a specific record, examine a formula, verify a chart. When you work through an AI intermediary, you see the AI's interpretation of that data — which may be accurate, partially accurate, or subtly wrong in ways that are hard to detect without going to the source. The UK government's Copilot trial found that AI-assisted work was faster but produced lower-quality and less accurate results — precisely because the abstraction layer made it harder to verify the underlying work.

Context collapse. A control plane that connects to everything creates a new problem: the AI has too much context and no guaranteed ability to determine what is relevant. When you ask for a project update and the AI has access to your email, your project manager, your spreadsheet, your messaging history, and your calendar, it must make judgment calls about what to include. Those judgment calls can be wrong — surfacing irrelevant information, missing critical details, or conflating similar items from different contexts.

Vendor concentration. If most of your work flows through a single AI control plane, that vendor has extraordinary visibility into your operations and extraordinary leverage over your workflow. The interoperability that MCP promises is important precisely because it mitigates this risk — letting organizations connect to multiple AI systems rather than depending on one.

Skill erosion. If you never navigate your own tools, you lose understanding of how your data is organized, where information lives, and how systems connect. This matters when the AI gets it wrong — because debugging an AI-orchestrated workflow requires understanding the underlying systems that the AI was supposed to abstract away.

These risks do not invalidate the control plane pattern. They define its boundaries. The AI layer is most valuable for routine coordination — the 80% of app-switching that is mechanical, not creative. It is least valuable when judgment, verification, and direct manipulation matter most.

The Next Interface War: Who Owns the Control Plane

The biggest question in enterprise software is no longer which application is best at a specific task. It is which AI layer becomes the default control plane for work.

Microsoft has the distribution advantage — 450 million commercial subscribers already inside Microsoft 365, with Copilot embedded in every application. OpenAI has the consumer momentum — 800 million weekly active ChatGPT users who are already forming the habit of starting work in a conversation. Anthropic has the infrastructure advantage — MCP is becoming the standard protocol for AI-tool connectivity, which means Claude's control plane can connect to anything that speaks the protocol.

But the real competition may not be between these three. It may be between the AI control plane and the application itself. Every SaaS company is trying to make its own product the hub — Salesforce wants to be the control plane for sales, Notion for knowledge work, Slack for communication. The question is whether users will prefer one universal AI layer or multiple embedded AI assistants, each deep in its own domain.

The answer is probably both — and the split will happen along a predictable line. For work that spans multiple tools (project coordination, reporting, communication), the universal control plane wins because no single application holds the full context. For work that is deep in one domain (data analysis, design, coding), the embedded AI wins because domain-specific tools have deeper capabilities than any general-purpose layer can match.

The end state is not one AI to rule them all. It is an AI layer that makes the boundaries between applications irrelevant for the work that crosses them — while specialized tools remain essential for the work that lives within them.

The human was the integration layer. The AI is taking that job. And the applications that adapt to being orchestrated, rather than insisting on being the destination, are the ones that will matter most in what comes next.

---

At AIReady.fit↗, we help professionals and teams build productive AI workflows. Our AI Foundations track covers how to work effectively with AI across your entire tool stack — practical skills for professionals who want to spend less time switching and more time thinking.