AI may not be reducing the workload; it's just making it invisible.

AI may not be reducing the workload.

It might only be making part of the job invisible.

Last week, an interesting detail emerged about Meta's personal AI agent, Muse.

Muse can call businesses on behalf of the user; schedule appointments, inquire about product availability in stores, and gather price quotes from different service providers. From the outside, it looks like an autonomous AI agent.

But according to internal company correspondence obtained by Reuters, Meta was testing a “human concierge” system that silently outsourced some conversations to human contractors. While the user thought they were assigning a task to an AI agent, a call center employee was conducting part of the conversation.

Employees within the company warned that sensitive information could be shared with human contractors. A manager admitted the test, launched without sufficient explanation, was a mistake, and the feature was withdrawn.

This single case doesn't mean there are humans behind all AI agents. But it illustrates a larger governance problem we're overlooking in the AI transformation:

When we assign a task to AI, human labor doesn't always disappear. Sometimes it's relocated and becomes invisible.

IBM's global research, published the same week, shows this situation on a broader scale. In the study, which included 1,500 CHROs in 21 geographies and 8,800 employees in 28 countries, 80% of CHROs said that the use of AI is creating a new "invisible job" for employees.

What does this invisible work consist of?

Validating the AI's suggestion. Correcting its errors. Adding missing context. Managing exceptions. Explaining a failed transaction to the customer. Taking responsibility for the system's decision.

42% of employees say AI increases their workload or that the extra effort it creates is not recognized. 43% say they are blamed when something goes wrong with AI.

In my opinion, these figures reveal a significant deficiency in how companies measure AI effectiveness.

Today, most AI business case studies are framed around the question, "How many hours have we saved?" If a report is prepared in thirty minutes instead of three hours, we assume that two and a half hours have been saved.

But obtaining the initial output is not the same as producing a reliable result.

If an employee spends forty minutes checking the AI-generated report; seeking support from the finance team to correct an incorrect figure; re-investigating a claim with an unclear source; or waiting for legal approval before sending it to the client, the real savings are not two and a half hours.

More importantly, if this monitoring task is not included in the job description, is not reflected in the performance system, and is not measured in the budget, the organization may overestimate the economic value of AI.

I call this the invisible human toll of AI.

On one side, there are the visible tasks that are becoming automated. On the other, there is the labor of validation, exceptions, compensation, and accountability distributed throughout the system.

True value emerges when these two aspects are viewed together.

The hidden cost of AI is often not the computer, but the human attention that is not written in anyone's job description.

This isn't just a matter of efficiency.

The invisible work is also a matter of trust and leadership. If an employee questions the AI's output and is seen as "resisting change," we may be punishing human judgment that could prevent negative outcomes. When the system makes a mistake, the responsibility falls on the employee, but if the employee doesn't have the authority to change the system's decision, we are separating accountability from authority.

Here's one of the key findings from the IBM study. Organizations that clearly defined which jobs would be managed by humans, which would be AI-assisted, and which would be performed by AI reported lower risk and higher quality.

In organizations where CHRO shares responsibility for determining which decisions remain human, employees feel significantly more secure in challenging or invalidating AI recommendations.

So the issue isn't just about teaching employees how to write better prompts.

Designing the division of labor between humans and AI, including authority, responsibility, and recognition.

Therefore, I think we need to add new items to the AI ROI calculation.

In addition to gross time savings, we need to measure the time spent validating the output. We need to see the cost of errors and rework. We need to track the volume of exceptions delegated to humans. We need to account for customer compensation. And we need to ask whether the person ultimately responsible actually has the time, knowledge, and authority to do the job.

When this calculation is done, we may see that some AI applications create less value than we thought.

Conversely, we may find that some are far more valuable. This is because well-designed AI doesn't just speed up initial output; it reduces the need for validation, moves exceptions to the right person, preserves the decision trail, and frees up human attention to truly more valuable work.

But this result doesn't happen spontaneously.

AI implementation can be a technology project. Eliminating invisible labor, however, is job design.

At your next AI management meeting, select a single critical workflow. Before asking, “How many hours did we save?”, ask these four questions:

Who validates the AI output? Who corrects the errors? Who handles the exceptions? Who is held accountable when the result is wrong?

If the answers to these four questions are distributed among different people and none of them are being measured, your AI may have automated the task.

But your organization hasn't redesigned the job yet.

My AI Invisible Work Map, designed for management teams, makes the difference between gross time savings and actual net human labor visible in a workflow. If you'd like to do this calculation together across three critical AI workflows, I can share the map with you.