What Do You Do With the Other 90%?

Tenex co-founder Alex Lieberman recently heard a startling figure from the head of AI at a huge company: Despite extensive training, just 10% of employees account for 90% of the reported AI productivity gains and token usage. The distribution hasn’t budged.

Now the executive is wondering whether the other 90% are overpaid, whether significant layoffs would leave output intact, and whether concentrating resources on the top performers would damage company culture.

Some see the split as evidence that most employees lack the initiative to retrain, but the numbers can’t tell us that. They show how reported AI gains are distributed, not what those employees contribute to the business or whether the company has given them a meaningful opportunity to improve their work.

Before judging the other 90%, management needs to establish what AI is supposed to accomplish, and how it will know whether it’s working.

Before blame, check the math

The executive’s 10/90 split measures the additional productivity employees reportedly gain from AI. It doesn’t measure how productive they were to begin with, or how much value they contribute to the business.

Stanford’s AI Index reviewed the workplace research and found that less-experienced workers often saw the biggest improvements. Gains also varied enormously by task. AI performed best on structured, repeatable work, while in tasks requiring more complex judgment, it sometimes offered little benefit or even made performance worse.

An excellent employee whose work offers few opportunities for AI-driven improvement might show almost no gain. A less-experienced colleague doing more easily automated work might show a substantial one. That’s no basis for deciding who’s overpaid.

Token usage is a bad barometer

Token usage tells you how much AI someone uses, but it doesn’t tell you whether they’re producing anything useful. A May working paper from the National Bureau of Economic Research studied more than 500,000 GitHub developers alongside their AI-tool usage. It found that autonomous coding agents were associated with a cumulative 240% increase in code commits.

But when they looked farther along the production process, the estimated increase fell to 80% for projects and 30% for releases. Across four major software marketplaces, they also found more new applications but no increase in total app usage. In other words, more code was being written, but much less of that additional activity was making its way into released software, and the researchers couldn’t find a corresponding increase in how much people used the resulting apps.

Delivering more isn’t necessarily the same as creating more economic value.

Individual productivity ≠ earnings

McKinsey’s State of AI survey found that 80% of respondents said AI had improved their individual productivity. Yet only 37% reported any positive contribution to their organization’s operating earnings, a figure essentially unchanged from 2025. Those are different, self-reported measures, but the contrast captures the difficulty companies face turning individual gains into financial results.

It’s also worth keeping the scale in view: The Generative AI Adoption Tracker, a nationally representative survey run by economists at the St. Louis Fed, Vanderbilt, and Harvard, puts the time saved by AI at 2.1% of total U.S. work hours as of August, non-users included. Those are self-reported estimates, not measured output.

We’re judging employees by how much more productive AI makes them, often without establishing what improvements are possible in their jobs, whether those improvements are valuable, or how the company intends to capture them.

Asking people to do a second job

Some leaders want people to discover opportunities nobody has fully worked out, experiment with unfamiliar technology, evaluate the results, and potentially reinvent how their jobs get done.

They also, of course, want to see their employees do this while continuing to perform the jobs they were hired for. Someone can be excellent at their job without being particularly good at developing AI applications. Is it reasonable to expect otherwise? Does it make sense that being excellent at your job is insufficient unless you’re also capable of reinventing it?

Support, recognition, and training

A two-year MIT Sloan study followed AI initiatives at a healthcare organization and a law firm. Both provided tools and initial training, and researchers found no meaningful differences in AI readiness, regulatory constraints, or how suitable their work was for AI. Yet the healthcare organization deployed 141 AI solutions, while the law firm deployed three.

The difference was that the healthcare employees received ongoing technical support and help integrating AI into their workflows. Experimentation also counted toward their responsibilities, evaluations, and career advancement. The law firm offered little continuing support or recognition, so more than 80% of its participating domain experts disengaged.

The findings suggest the problem is asking for additional work without the support or incentives to sustain it. In Microsoft’s Work Trend Index, 45% of surveyed AI users said it felt safer to focus on their existing goals than to redesign their work with AI. Only 13% said they were rewarded for reinventing work even when their experiments failed.

There’s also evidence that training alone doesn’t explain adoption. In a paper presented at Brookings in March, researchers found that among workers whose employers provided neither AI training nor tools, 47% used AI when their employers encouraged it, versus 10% without encouragement. Once encouragement and access to tools were accounted for, formal training was no longer independently associated with adoption.

Management catchup

Here’s where the discussion of AI “laggards” gets especially strange. Gartner reported on October 26 that 78% of 297 chief human resources officers surveyed said their leadership teams lacked alignment on where AI creates value. Yet we’re having serious conversations about whether employees who aren’t producing extraordinary AI gains lack agency, need replacing, or deserve smaller salaries.

How many of the executives prescribing an AI transformation have spent meaningful time doing the work they’re asking to transform? Have they established where the technology offers a credible return? Have they provided the time, tools, and support necessary to find out? While employees should contribute, should they also be expected to make up for an absent strategy through individual ingenuity?

A 10/90 split isn’t a layoff plan

The Wharton–GBK AI Adoption Report, The Adaptive Advantage, found that organizations reporting strong AI returns were more likely to have redesigned workflows, changed management structures, monitored AI use, and provided adequate training. They were no more likely than other organizations to report staff reductions or hiring freezes.

Expectations have also been running ahead of outcomes. In McKinsey’s previous survey, 32% expected AI to shrink their workforce. This year, just 14% reported that it had. McKinsey checked the comparison among respondents who participated in both surveys, with similar findings. Yale’s Budget Lab, which has tracked the labor market monthly since last fall, reported in September that occupational churn and unemployment among AI-exposed workers remained within historical ranges or on their pre-AI trends.

And cutting staff can be expensive to reverse. In a February survey of 600 HR leaders whose organizations had conducted layoffs, nearly a third reported losing critical skills and expertise. Among organizations that rehired, about half filled eliminated roles within six months. Only 8.4% said their AI-driven restructuring delivered what was promised and that they would repeat it unchanged.

Layoffs may be justified when AI genuinely changes how much labor a company needs, but a distribution of AI productivity gains doesn’t tell the executive which jobs have become unnecessary, or what the company would lose by eliminating them.

A better approach

Yes, some employees will resist useful improvements, some roles will require fewer people, and companies can reasonably expect employees to learn tools that demonstrably help them perform their jobs. But an AI strategy shouldn’t depend on turning the entire workforce into amateur AI developers.

Leadership should decide which problems are worth solving and provide the resources.

Employees and managers who understand the work should help identify opportunities and test solutions.

1. Find the opportunities where the work happens

Give department leaders and employees a meaningful role in identifying expensive, frustrating, repetitive, or otherwise improvable processes. They understand the details, exceptions, and hidden dependencies that may never appear in a management report. Leadership should use that knowledge to identify promising opportunities and concentrate investment where there’s a credible business case.

That also means being willing to discover that AI isn’t particularly useful for some work. That’s a legitimate finding, and it should affect where the company spends its money.

2. Make experimentation part of the job

Give people approved tools, budgets, technical support, and actual time carved out of their workloads. Recognize valuable experimentation in performance evaluations and career advancement, including rigorous efforts that establish why a proposed application isn’t worthwhile.

And don’t assume the person who identifies a problem must also build the solution.

There’s a useful example in BCG’s study of more than 50 AI front-runners. Greenhouse, the hiring-software company, is introducing an innovation lead within each department to identify workflows worth automating and evaluate tools. Those leads then work with a central team of IT engineers to build the solutions.

This also helps retain the employees driving improvements. If they spend their days fighting for resources, getting managers to cooperate, and squeezing experimentation around their regular jobs, why wouldn’t they eventually look elsewhere?

3. Make leadership accountable for the results.

Measure what actually improves: costs, quality, throughput, revenue, or customer experience. Coordinate changes across departments, remove obstacles, and give successful experiments a path to production. Measure outcomes over activity.

Zapier, another of BCG’s front-runners, evaluates people on impact rather than usage for exactly this reason; its summary of the problem is that “AI is brilliant at activity.” The GitHub study is the cautionary version: commits rose 240% while releases rose 30%. We still don’t know whether the executive’s dashboard captures useful business outcomes or merely reports AI activity.

Track whether time saved turns into additional output, whether automation reduces actual costs, and whether the company is capturing the value it projected. Reconsider roles and staffing when the evidence supports it. And where there’s no credible business case for changing a job, let people keep doing good work.

The executive’s questions, answered

The executive Alex talked to may have found a real performance problem. Context can’t tell us for sure, so let’s hit the questions one by one:

Are the 90% overpaid? AI gains aren’t total employee contributions, and token usage doesn’t measure productivity. Some jobs offer more opportunities for AI than others.

Could the company cut deeply without losing output? Nothing in these figures establishes that; strong AI returns don’t necessarily require staffing cuts.

Should it concentrate resources on the top AI users, even at a cost to culture? The answer to this can only be yes where the investment creates value. Otherwise, the company risks rewarding activity while alienating employees whose existing work remains valuable.

There’s a fourth question the executive hasn’t asked, at least in this account: What exactly is AI supposed to improve at this company? Which workflows, measured how? And what has management done to make those improvements possible?

So what do you do with the other 90%? Identify the work worth changing, give people the resources to change it, and measure the results.

The numbers this exec is working with certainly start to tell a story, but it is not a story about who provides value and how.