Feature
White Paper
The Future of Leadership Development: AI-Driven Manager Enablement
Enterprise AI leaders have a metrics problem. Adoption dashboards are green, seat counts are climbing and quarterly business reviews are full of slide decks showing deployment progress.
The problem is a widely used tool is not the same as a structurally embedded one. The difference only becomes visible when someone asks: if it switched off tomorrow, would work break, or would people just find another way?
Take Claude. Anthropic now commands an estimated 40% of enterprise LLM expenses, ahead of OpenAI's 27%, according to Menlo Ventures' December 2025 report. Seventy percent of Fortune 100 companies use it, and Deloitte rolled it out to more than 470,000 employees.
Those figures come from third-party analysts and press coverage, because Anthropic does not publish its own enterprise adoption data. By any adoption metric available, the numbers are exceptional. But Anthropic does not publish metrics such as retention rates, workflow penetration or substitution behavior, which means enterprise buyers deploying it at scale are working with an incomplete picture.
Metrics such as seats, tokens used and monthly active users "measure adoption, not value," according to Alastair Paterson, CEO and co-founder of Harmonic Security.
Then there’s the other side of the coin. When employees turn to personal accounts for AI tools, Paterson's research found that nearly two-thirds of that activity is business use. Workers are routing around the sanctioned stack to get things done, and the enterprise is measuring deployment while dependency forms somewhere else.
That means the tools employees rely on are not the tools the enterprise thinks it is managing.
The Difference Between Usage and Indispensability
What happens if you take the tool away?
"If you remove the tool and people can still ship work the same way with a small inconvenience, that is usage,” said Dustin Engel, co-founder of Elegant Disruption, who advises enterprise leadership teams on AI adoption and operating models. “If removal forces measurable degradation in throughput or quality, or a meaningful reallocation of labor and time, that is indispensability."
Most enterprises are not measuring that because doing so requires connecting AI activity to workflows and outcomes rather than use.
Metrics that capture structural embedding are workflow-level, not platform-level: cycle time on defined processes, revision loops, rework rates, escalation frequency and whether AI-assisted outputs can be consistently tied back to the source.
Moreover, they are not being tracked systematically at most organizations, which makes it a governance problem as well.
Engel recommends:
- Identify five to 10 workflows where AI affects revenue, risk or customer experience.
- Run them without the tool for a defined period.
- Measure the difference.
It shows where AI is embedded rather than a convenience, and shows whether the organization is model-agnostic at the process level or locked into a tool-specific way of working.
Measuring Minutes Per Task
Task duration data offers a practical approximation of the same answer.
Tools with the deepest engagement, measured by time spent on individual tasks rather than how often employees open the application, tend to be the ones that have become structurally embedded, Harmonic Security’s research found. For example, Perplexity Enterprise averaged nearly 12 minutes per task, while Claude Enterprise averaged just over 10 and Gemini Pro measured 5.5.
Short task times suggest a tool handling quick lookups and convenience queries. Longer engagement suggests something more complex and harder to replicate, a tool woven into how consequential work gets done. It can also mean the tool is experiencing a slowdown.
In addition, departmental breakdowns show which ones are most dependent on the tool. "What's favored by marketing might not be the tool of choice for legal departments," Paterson said. Enterprises that do not make that distinction miss that nuance.
The 10.4-minute average for Claude Enterprise is the closest proxy available for indispensability and suggests embedding. But task duration is a third-party measure from Harmonic Security's research, not something Anthropic publishes.
What happens inside those deployments, whether removing Claude would break workflows or merely inconvenience them, remains opaque. Other major AI vendors don’t publish these metrics either.
Measuring Human Oversight Cost
Another metric that models have trouble capturing is the cost of human oversight with agentic AI.
Oversight requires a reviewer with sufficient AI literacy to recognize when an output is wrong, the training to develop that literacy and the time to perform the review. Those costs are not always considered when putting together the business case.
Engaged, But Not Convinced: What's Eroding Beneath Your Strongest Engagement Numbers
Watch Now
Retention Is a Communication Problem: New Data from 6,000+ Employees
Watch Now
"The fully loaded cost is rarely showing up, and that omission is where most ROI models break," said Jason Dods, director of information management and technology at Highspring.
What looks like savings is instead a reallocation of effort. A process that previously required 10 hours of human work does not become free when AI handles a portion of it. Instead, it becomes judgment, exception handling and verification.
"If we take an honest assessment of impact, the productivity number gets smaller and far more defensible," Dods said.
Integration carries the same problem. Configuration, permissions management and ongoing maintenance are recurring expenses that don’t get considered. Without a multi-year total cost of ownership model that captures them, "the ROI you are presenting is fiction," Dods said.
Embedded, Not Adopted
For AI to be indispensable, it has to be built deliberately, and that takes longer than most enterprises plan for.
Fast onboarding and generic copilot deployments produce adoption metrics, in Engel’s view. Structural embedding requires AI wired into the mechanics of work: connected to systems, operating within constraints and subject to governance that specifies what the agent handles autonomously, what requires human review and what stays off limits.
Validation means being able to demonstrate, consistently and on demand, that the AI-assisted output meets the same standard as the human. Most deployments cannot do that yet, Engel said
Indispensability at the individual level is not the same thing as indispensability at the organizational level. A tool that thousands of people find personally useful but that does not affect workflow cycle time, quality, risk posture or customer outcomes has not changed the business. "It has changed the desktop," he said.
The question enterprise AI leaders need to be asking is not how many seats are active, but whether the organization would be measurably worse off without the tool, and whether anyone has tested that.
Editor's Note: For other thoughts on what productivity means when AI's in the picture, read:
- Rethinking KPIs as AI Shatters the Curve — AI makes targets easy to hit in a day, but that doesn't necessarily translate to results. Here's how to measure what actually matters now.
- Why Nobody Agrees on AI Productivity Metrics — As AI reshapes work, companies want new productivity metrics. The problem? Nobody can clearly define or fairly measure the alternatives.
- The Real Productivity Effects of Generative AI — AI won’t replace work so much as it will reshape who succeeds, how fast people learn and how organizations adapt. Leadership responses will define the outcome.