I think every HR team with an AI rollout should write down where the hours it frees are meant to go, and should do it before the next performance cycle. New research from The Inclusion Initiative at the London School of Economics and Political Science (LSE) and Protiviti shows what happens when nobody decides: the hours go back into more work, and employees start worrying about how much AI they are seen to use.

The numbers behind the argument

The study surveyed 3,001 professional employees globally from March to May 2026. AI users report saving an average of 19% of their working week, and they reinvest 72% of that saved time in other work. Half of those who saved time say they used it to take on new responsibilities or projects. The researchers call the risk “workload creep”.

The same survey found that 48% of employees worry about being judged for not using AI enough, against 20% who worry about being judged for using it. And it found no consistent association between perceived job performance and AI adoption, training, in-house access or integration into standardized workflows.

Advertisement

HRTech Your brand belongs here. Reach the decision-makers who read HRTech every day. Premium placements across the site and newsletter. Advertise with us

Dr Grace Lordan, founding director of The Inclusion Initiative at LSE, said in the release: “Where those time savings go matters as much as the size of the savings.” I agree, and I think the sentence describes a decision that nobody in particular owns unless HR claims it. Our own coverage of the study is in today’s report on the survey.

The strongest case against me

The objection is a fair one. An employer buys AI tools to get more output for the same payroll. If an employee saves a fifth of the week and fills most of it with other work, the investment is doing what the business case said it would. Reserving part of that time for anything else looks like handing back the return. The survey even offers comfort: 65% of those who saved time say they spend at least moderately more time on higher-value activities.

I take that seriously, and I still think it gets the accounting wrong in three places.

Why the default allocation costs more than it looks

First, the saving itself depends on training. Trained employees in the survey report saving 22% of their working hours, and employees without training report 13%. Adoption is 93% among the trained and 50% among the untrained. Training takes time, and the only place that time can come from is the hours AI has freed. An organisation that lets every saved hour flow into delivery is spending the budget that would have enlarged the saving.

Second, access to that training is already uneven. In the past year 66% of Gen Z employees received AI training, compared with 40% of baby boomers. Left to individual initiative, the people with the most slack will train and the people with the least will fall further behind on a tool their employer expects them to use.

Third, the anxiety buys nothing. If nearly half the workforce is worried about being seen to use AI too little, and AI use shows no consistent link to how employees believe leaders rate their work, then the pressure is producing display and little else. I made a related argument in October, that executive AI usage is a poor proxy for workforce readiness.

Newsletter

Get the week's best tech coverage.

Free. Read by thousands of HR, tech, and business leaders.

What I would put in writing

The policy can be short. I would start with three lines.

  1. A stated share of saved time goes to learning. I would begin with one saved hour in five and adjust after two quarters. That ratio is my proposal and does not come from the study. What matters is that a manager can point to a number when an employee asks whether training time is allowed.
  2. New responsibilities trigger a role review. When someone absorbs a project or a duty with hours AI freed, the manager logs it, and the job description and level are checked at the next cycle. Half of time-savers in the survey are already in this position.
  3. Performance conversations assess the work. If AI use is a criterion for a role, the criterion is written and specific. If it is absent from the criteria, managers say so plainly. This sits alongside the earlier case that HR should write role-level AI expectations before building a skills library.

Matt Duncan, managing director at Protiviti, put the employer’s interest this way in the release: “Simply absorbing those gains into existing workloads leaves much of AI’s potential untapped.” A consulting firm is telling clients the default leaves value behind.

Where this could be wrong

The survey is self-reported and covers professional employees, so the 19% figure is an estimate by the people saving the time. The training comparison is an association. If your own data shows trained and untrained staff saving the same hours, my first point falls away and the learning share should shrink.

That is the reason to measure before the cycle starts. Ask AI users how many hours they saved last week and what replaced them, then split the answers by whether the person has been trained. HR teams that have those two numbers by January can set the learning share from their own data instead of mine.

Source: The Inclusion Initiative at LSE and Protiviti, PR Newswire