The "Toggle Tax" Is Real, and Your AI ROI Calculation Is Wrong

The productivity promise of workplace AI rests on a flawed accounting method. Companies measure what AI produces; almost none measure what it costs workers to operate it. According to research from HERE Enterprise, employees are spending as much effort managing workplace AI tools as they are benefiting from them — a phenomenon the researchers label the "toggle tax," defined as the cumulative cognitive and time cost of switching between AI interfaces, re-prompting failed outputs, and verifying AI-generated work before it can be used. That one finding should end the conversation about whether AI deployment is an automatic win.

Why does AI create more work instead of less?

The toggle tax is not a UX problem. It is an architecture problem dressed up as a convenience feature. Most enterprise AI is deployed as a layer on top of existing workflows rather than as a replacement for the steps that create friction. Workers toggle into a tool, generate output, toggle back to their core system, edit the output to fit context the AI missed, and then often redo the underlying task anyway. HR Executive (2025-07-27) describes users spending as much time managing the AI as they save through it — meaning the net productivity gain for a large share of deployments is, at best, zero.

This would be a manageable problem if leaders recognized it. Most do not. Research from HR Executive (2025-07-27) finds that despite high levels of self-reported confidence, managers show measurable gaps in AI readiness, change management capability, and employee support — the exact skills required to detect when an AI rollout is generating hidden costs rather than savings. Confidence without calibration is how organizations end up six months into a deployment with no mechanism to ask whether the tool is working.

What does fear of obsolescence cost you operationally?

The toggle tax compounds a second problem that is arriving from the employee side: fear of becoming obsolete — a phenomenon that Education Testing Service (ETS) and The Harris Poll have labeled fear of becoming obsolete (FOBO). HR Dive (2025-07-27) reports that most workers surveyed say upskilling "is no longer a choice," yet simultaneously say it is hard to get employer support for it. That gap is not a training budget question. It is a strategic incoherence question: organizations are deploying tools that change the nature of work while simultaneously failing to invest in the human capacity to work differently.

The operational cost of FOBO is not primarily attrition — it is disengagement while people are still present. A workforce that believes it is becoming obsolete will rationally reduce the discretionary effort it applies to processes it expects to be automated. Pair that with a toggle tax that erodes the time those workers do spend on AI-assisted tasks, and you have an organization that is paying for transformation while receiving less useful output than it had before.

Retention is the variable that makes this urgent

Here is where the calculus becomes board-level serious. In 2025, OpenAI's stock-based compensation averaged roughly $1.5 million USD per employee — unprecedented for a pre-IPO company — and OpenAI still experienced high-profile defections to rivals including Meta, which reportedly extended offers in the hundreds of millions of USD, according to MIT Sloan Management Review (2025-07-27). The implication for non-hyperscaler organizations is not that you should match those numbers. It is that compensation alone does not hold talent when the work environment itself is degraded. If your best technical and knowledge workers are absorbing toggle tax all day while feeling unsupported in their development, the cost of losing them is not offset by the AI subscription you bought to replace their output.

Agentic AI (AI that can take sequences of actions autonomously, such as submitting code or modifying configurations, without a human reviewing each step) is the next phase of this problem. GitLab (2025-07-24) notes that agentic systems break the human review loop that made earlier AI-assisted work auditable. When agents act autonomously, the toggle tax does not disappear — it shifts to verification and governance work that is harder to see, more expensive to undo, and more likely to land on your most senior people.

The real question for any executive team: if you audited the actual time your workers spend managing AI versus being aided by it, would the number justify the investment you have already made — and would you still approve the next phase of deployment on that evidence?

Every technology initiative eventually gets stress-tested by the people who have to use it daily, and the organizations that ask that question now will be in a materially different position than those who wait for attrition data to tell them something went wrong.