A new survey of 10,842 working adults found that 69 percent worry AI will reveal parts of their jobs they do not fully understand. By comparison, 41 percent were more concerned about being replaced by AI. Nearly two-thirds said they spend additional time reviewing AI-generated work, and more than half worried about how a manager would react to seeing that AI completed something faster.
I can see why. Most AI adoption dashboards begin with the data that is easiest to count: who has a license, who logs in, how many prompts they send, and how many tokens they consume. Put those numbers next to employee names and people will reasonably assume they are being graded, even if leadership describes the dashboard as an adoption tool.
Once that happens, the measurement system starts changing the behavior it is supposed to observe. Some employees hide their usage. Others produce more visible activity to look engaged. Some avoid AI entirely because they do not know how the data will be interpreted. The company ends up with a cleaner chart and a less accurate picture of how work is changing.
That is not only a culture problem. It is a data-quality problem. A system designed to improve adoption cannot guide good decisions when the people generating its signals are managing their visibility instead of improving the work.
The measurement system changes behavior
Every workplace metric creates an incentive. Sales teams respond to pipeline targets. Support teams respond to resolution metrics. Engineers respond to delivery and reliability goals. AI metrics are no different. If leadership celebrates prompt volume, people learn that visible usage matters. If leadership publishes token leaderboards, people learn that spending more looks like progress.
Neither metric tells the company whether the work improved. A thousand prompts could represent a valuable research process, repeated model confusion, or an employee breaking one simple task into unnecessary interactions. Low activity could mean resistance, or it could mean that a team built one reliable automation that now runs quietly. The number becomes useful only when it is connected to the workflow and its result.
Activity metrics are not performance metrics
Activity data still matters. It can show whether a purchased tool is being used, where adoption is growing, and where costs are concentrating. The mistake is treating activity as the outcome. A login is not a better customer experience. A prompt is not a finished task. A model call is not a decision anyone should trust.
Gallup's workplace research makes the distinction useful. In its July 2026 survey, employees reported the strongest productivity gains from task-specific uses such as coding assistance and automation. Seventy-seven percent of employees using AI for either category said it had a positive effect on productivity. Writing and research were more common uses, but employees reported smaller gains from them.
Gallup also found that employees using AI across more kinds of work were more likely to report productivity gains, while warning that the relationship does not prove that broader usage caused the improvement. That caution is important. Measurement should help a company ask better questions, not turn correlation into a performance score.
Measure workflows, not people
The better unit of analysis is the workflow. A workflow has a purpose, inputs, steps, review points, costs, and a result the business already understands. Looking at that level changes the conversation from whether an employee uses enough AI to whether the work became faster, better, safer, or more economical.
Consider a support team using AI to prepare case summaries. The useful questions are whether handle time changed, whether escalations became clearer, whether reviewers corrected important errors, and whether customers received better answers. Prompt count can help explain the process, but it should not be confused with the result or used as a shortcut for evaluating the employee.
The same approach works in engineering, sales, finance, legal, recruiting, and marketing. Start with the operating result each function already manages. Then use AI activity as one input for understanding how the workflow changed. This keeps the measurement connected to work the company actually values.
Five questions for a useful AI adoption review
- Which workflows changed? Identify repeated work where AI is now part of the process, rather than counting isolated experiments.
- What result moved? Look for changes in cycle time, quality, customer outcomes, revenue, cost, risk, or another metric the team already owns.
- Where did human review remain essential? Record corrections, exceptions, and judgment calls instead of hiding them inside a productivity estimate.
- What did the workflow cost? Connect seats, model usage, supporting tools, and review time to the work that consumed them.
- Can another team reuse the pattern? Find successful workflows that are teachable and transferable without exposing an individual employee's private activity as a leaderboard.
These questions produce a very different management conversation. A team can explain what changed, what the evidence shows, and what remains uncertain. Leaders can fund useful patterns, improve weak ones, and retire waste without pretending that more activity automatically means better performance.
Trust has to be designed into the measurement
Leaders should be explicit about why AI data is being collected, who can see it, how long it is retained, and which decisions it will not make on its own. Employees should know whether the goal is license optimization, security, workflow improvement, cost control, enablement, or individual evaluation. Combining those purposes without clear boundaries is how an adoption program begins to feel like surveillance.
Access should follow the decision being made. Finance may need aggregated spend by vendor and team. Security may need enough detail to investigate risky tools or data paths. A business leader may need workflow outcomes and ownership. A manager usually does not need a ranking of every employee's prompts. More detail is not automatically more useful, especially when it damages the trust required for honest adoption.
The organization also needs a way to challenge the interpretation. If one team has unusually high model usage, the explanation might be waste, a more complex workload, a successful automation, or a measurement artifact. The people closest to the work should be able to add context before a dashboard becomes a conclusion.
Visibility should help people improve the work
The purpose of AI visibility is not to prove that every employee is using the tools. It is to understand where AI participates in work, what it costs, which patterns produce value, where risk appears, and what the organization should improve next. That requires enough detail to manage the system without reducing people to an activity score.
At Proxon, we are building measurement around AI usage, tools, agents, workflows, ownership, spend, adoption, and outcomes. Connecting those signals gives leaders a way to evaluate the work rather than relying on seat counts or prompt volume alone. It can support better management decisions, but it should not be treated as an automatic judgment about an employee's performance.
Companies need enough visibility to learn where AI helps without turning the rollout into a new form of employee surveillance. The test is straightforward: does the measurement help the team understand and improve a workflow, or does it merely make employees feel watched?