For years, “swivel chair work” was shorthand for a very real problem: employees spinning between applications, copying data from one system to another, and chasing approvals through multiple inboxes. The term stuck because it was accurate.
Harvard Business Review reported in 2022 that workers toggle between applications 1,200 times daily, losing nearly 4 hours per week to reorientation. That’s roughly 9% of total work time. For a knowledge worker earning $100,000 annually, the cost translates to $9,000 in lost productivity per employee per year. For a company with 10,000 employees, that fragmentation amounts to roughly $90 million annually, not from bad strategy or underinvestment, but from the basic friction of fragmented digital infrastructure. Almost no organization has a line item for it. It just compounds.
WE GAVE AI THE WRONG JOB
When generative AI arrived, most organizations gave employees better tools for the work they were already doing. Better search, faster drafting, and sharper summaries. But they didn’t change the underlying architecture. The siloed systems, approval chains, and status emails were all still there. We added AI to the workflows without redesigning the workflows. Employees ended up with one more window to toggle to.
In many organizations, AI just made the chairs swivel faster.
The first generation of enterprise AI was primarily generative, retrieval-augmented systems that surface the right document, summarize it well, and hand it back to a human to act on. That’s a meaningful capability, but it’s still a handoff. The AI produces an output for a person to carry it forward. Output is not outcome. In business, no one gets paid for outputs, but rather for completed work.
FINISHING THE WORK IS HARDER THAN IT LOOKS
The next phase of AI is about completing the work.
Agentic AI is defined as systems that can plan a sequence of steps, select the right tools, execute actions across systems, and govern the outcome with auditability. They can now take a request from initiation to resolution without a human manually stitching it together.
The intelligence part—the reasoning, language understanding, and ability to interpret an ambiguous request—is increasingly commoditized. Two years ago, GPT-4-level reasoning was a competitive advantage. Today, that capability is available open source, inference costs have dropped by roughly 10-fold annually, and benchmark performance across the leading models has converged to where the differences are largely invisible to enterprise users. What remains genuinely difficult and differentiated is the execution layer connecting that intelligence to the enterprise’s operational fabric.
Consider what completing a single workflow actually requires. Onboarding a new hire might span a human resources information system, an IT provisioning system, a facilities platform, a payroll processor, and a compliance training tool. Each has its own data model, permissions architecture, and API behavior. And they were never designed to talk to each other.
This is why “just use an LLM” isn’t an enterprise AI strategy. A large language model without enterprise context produces large language model results—fluent, plausible, and often disconnected from your actual business machinery. The grounding layer, the connectors, the context graph, the permissions model, and the workflow orchestration are what turn a capable model into a system that completes work.
WHAT EXECUTION AT SCALE LOOKS LIKE
What makes this possible is the harness around the model. When an employee submits a request, they’re not talking directly to an LLM, but to a layer incorporating business context. The LLM selects the right path, applies the right permissions, and decides which systems need involvement and the order. The model supplies the reasoning and the harness supplies the judgment about how to apply that reasoning inside your organization.
That’s why we’re seeing enterprises resolve the majority of service desk requests without human involvement. They are resolved end to end, with the right action taken, logged, and governed. Live agent chat volumes drop in days because the requests never had to escalate. At one global manufacturer, approval processes that once averaged 10 hours of VP time now close in under 10 minutes. Same policy. Same people. The harness surfaces the right context in sequence, and the inbox archaeology disappears.
The swivel chair became unnecessary.
These productivity improvements are structural changes to how operations run. The people who were spending their days on coordination overhead are now doing the work that requires their expertise. That is the shift worth chasing.
THE MOAT HAS MOVED
McKinsey’s 2025 Global Survey found that 88% of organizations report regular AI use in at least one business function. Individual use cases show promise, including a 14% productivity boost in customer service operations while using AI. That’s meaningful, but incremental.
The competitive question now is: How deeply is your AI connected to how work happens in your organization?
The companies building that depth with context graphs, orchestration layers, and the governed workflow execution are getting more efficient. Every workflow you automate teaches the system more about how your company operates. That compounds. It comes from building AI into how the company runs.
WHAT THIS MEANS RIGHT NOW
If you’re still measuring AI ROI in tokens generated and summaries produced, you’re measuring the wrong thing.
Find the work nobody chose. The status updates nobody asked to send. The approvals stalled in inboxes. The handoffs that require a human because the systems should communicate, but don’t. Map where your most capable people are spending time on coordination instead of judgment.
That’s your AI roadmap. Start with one workflow. Prove value in production. Then let the compounding begin.
The goal was always to make the swivel chair unrecognizable, to reach a point where the next generation of employees doesn’t know what the phrase means because they’ve never experienced the problem. We’re closer to that than most leaders realize. But only by building AI into the workflow, not on top of it.
Bhavin Shah is SVP and GM of Moveworks and AI at ServiceNow.
Leave a comment