TL;DR. A small number of people in your org are already doing the work of two, five, sometimes ten of their old selves. You probably can’t name them. Your seat report can’t see them either, because a seat count measures badges issued. Your consumption report can. Sort usage by person, descending, and the top of that list is where your leverage is sitting. This guide is a lens for reading that list, and a way to start the conversation that gets you the rest of the picture.

There are people on your team right now who are quietly doing the work of an extra hire you didn’t make. They aren’t on a special program. You didn’t pick them. You probably haven’t noticed yet, because the thing they’re doing is invisible from where you sit. They are producing work that used to require somebody else.

That’s the most useful lens I can give you for AI in your organization. Leverage is an extra colleague, sitting at someone’s desk, that you didn’t approve a requisition for. Your job is to figure out who they’re sitting next to.

Seat counts hide leverage; people reveal it

Most leadership reporting on AI is built around the wrong unit. Seats licensed. Seats active. Hours of training delivered. A vendor’s adoption curve. Survey results on whether people feel more productive.

Each of these is a proxy for something a vendor wanted to sell you. None of them measures the thing you actually want to know: who in your organization just got dramatically better at their job.

A per-user usage report gets closer. On metered agent and API work, sort consumption descending. On flat plans, sort activity and then inspect the work the active people are producing.¹ Finance reads that as a forecasting problem. Read it the other way. The report points at individual people rather than at license counts. Seat counts are blind by construction: everyone who has a badge looks identical. Usage has a top and a bottom, and the people at the top are running work through the tool at a volume nobody instructed them to.

That list is a shortlist, not an answer. Consumption is input, not output, and somebody burning tokens on a task they don’t understand looks the same on the meter as somebody clearing a backlog. The last step is still a conversation. The usage report tells you who to speak with first.

The shape you’ll find is consistent wherever anyone measures carefully. Start with the average. A St. Louis Fed paper on the November 2024 survey wave found that workers who had used generative AI in the previous week reported saving 5.4% of their work hours, roughly 2.2 hours a week.² Self-reported, and only among the people already using it. That is the polite number that lands on your slide. Now look underneath it. Enterprise telemetry from 2026 found half of AI users had twelve conversations or fewer while the top 5% had at least 144, and that only 18% of users touched AI in a given week at all.³ OpenAI’s enterprise data shows the same lopsidedness inside a single company: the 95th percentile of users sends about six times the messages of the median employee, and about seventeen times for coding.⁴ It is a lopsided distribution, and the published average describes neither population.

Anthropic’s study of its own staff gives the most detailed look inside one company: 132 engineers and researchers, self-reported Claude use in 59% of their work, an average self-reported productivity gain around 50%, and 27% of the assisted work described as work that would not have been done at all otherwise.⁵ Anthropic says plainly that its engineers get early access to frontier models and that the findings don’t generalize, so read it as a ceiling rather than a forecast. The last number is worth carrying. Leverage moves the existing queue faster and adds work that was never going to happen.

Once you’ve seen the shape, the seat report becomes useful for what it actually is: a count of badges issued. It stops being a measure of the program. The program is in the usage data and in the work.

Leverage appears as work that once needed another person

Look for what you’d notice about a strong new employee in their first quarter. Work that used to require somebody else is suddenly getting done. Backlogs that have been “next quarter” for two years are quietly clearing. The “we’ll get back to you” answer disappears from a part of the business where it used to be standard. A team that was supposed to grow doesn’t ask to.

This shows up most clearly in engineering, because the work is countable and the tools are mature. The leveraged engineer is shipping pull requests at a volume that used to belong to a small team. Old internal tools are getting rewritten on weekends nobody asked for. The backlog is moving in a direction it has not moved in years. The person is working differently. A coding agent now does the part of the job that used to require a meeting.

The same shape appears outside engineering, but the surface looks different. Since spring 2026, the clearest non-engineering example is no longer a person typing at all. A finance lead who used to spend the first week of every month on reconciliation now finds it waiting, already done, every Monday morning. She set an agent to produce it on a schedule, then used the reclaimed week to build the analysis her CFO has been asking for since last year. A marketer ships a campaign end to end without briefing the agency. An operations manager closes a vendor support ticket by writing the script herself instead of waiting for IT. A general counsel notices that outside-counsel spend on routine work is dropping and can’t quite explain why. The reconciliation case is the one to watch, because it shows leverage reaching non-coders: recurring work with a checkable output, produced on a cadence by something somebody set up once.

What ties them together is their relationship to the work. They keep AI open as a collaborator rather than treating it as a weekly document summary or search engine. They have changed how they work to keep it open. They delegate. They iterate. They treat a wrong answer as a draft rather than evidence that the tool is broken. That isn’t in your training curriculum. It is a temperament. The people who have it pick up a real tool in their first week. The people without it don’t develop that reflex, regardless of how many hours of enablement you fund. This is the disposition gap, and no training budget on record has closed it. What changed in 2026 is that a second gap began to matter too: a well-designed workflow can create leverage for people who never developed that temperament.

Designed workflows create a second form of leverage

Everything to this point describes leverage as a person: a temperament, a desk, a colleague you didn’t requisition. That’s the first shape, and until 2026 it was the only one that mattered. A second shape arrived in the middle of that year, and it doesn’t run on temperament at all.

The mechanism is the scheduled task. Someone writes a prompt once, picks a cadence, and an agent runs the work on its own: a Monday status brief assembled from five systems, a reconciliation that flags what didn’t tie out, a competitive scan sitting in the inbox before anyone arrives. No code, no IDE, no daily habit to change. This is the proactive agent pattern, and it’s what the Q3 2026 briefing is about: leverage reaching non-engineers through standing, checkable work.

This separates the person creating leverage from the person receiving it. In the first shape they are the same person: the engineer shipping the pull requests is the one who picked up the tool. In the second shape they are different people. The analyst who reads a finished reconciliation every morning may not use AI at all or even know an agent produced it. Credit belongs to whoever designed the workflow and set it running once.

Your search now has two parts. Look for the person who got dramatically better at their own job and the person who made a process better for people who changed nothing. The second person creates leverage for the entire team and is harder to see than a power user, because their fingerprint is on a workflow rather than a visible body of output. One person who can spot a recurring, checkable task and schedule it lifts everyone downstream. That is why a single designer can matter more than a team full of naturals.

Some work has not changed much yet

Current tools offer the least in work that is physical, relational, or adversarially correct. Field service, line manufacturing, hands-on healthcare. Late-stage M&A, executive coaching, complex relationship sales. Final-pass legal work, audit signoffs, anything where the value is in being right rather than plausible.

This is a current statement about those roles, not a permanent one. If you go hunting for the engineering-style step change in your senior litigators, you won’t find it, and you will conclude that AI is overhyped. The better conclusion is that the leverage is elsewhere. These roles resist compression because the work has no cheap way to check whether a given output is correct. That is the bottleneck that decides which workflows yield to AI and which stay stuck.

Leverage leaves observable signals

These signals are observable in your own organization over the next month. A person is operating at AI leverage if you can see at least one. The first comes from a report. The rest require you to look at the work.

Their consumption is several multiples of their team’s median. Sort the usage report by person, descending. In the telemetry, the top 5% of users run more than ten times the volume of the median user, and the median user is barely there at all. Whoever sits at the top of your own list is running work through the tool that nobody assigned them. That is not proof of leverage, because spend is input and a person can burn tokens going in circles. It is the shortlist, and it costs you one export to get.
They are doing work that used to require another person. A marketer who no longer briefs an agency. An engineer who no longer needs the data team to pull a report. A finance lead who no longer needs the BI queue.
They are doing work that used to be deferred indefinitely. Backlog items getting closed. “We should clean that up someday” things suddenly cleaned up. Reports nobody had time to build appearing in chat.
Their cycle time on a recurring task has dropped by more than half. Not 10% faster. Half or better. If it’s only marginally faster, it’s a small efficiency gain, not leverage.
Their output has shifted in a way peers in the same role haven’t matched. Same level, same tenure, dramatically more shipping. Not longer hours. Not corner-cutting. Just more.
Other people are starting to depend on something they built with AI. A script. A prompt template the team uses. An internal tool that was supposed to be a one-off. This is the strongest signal, because it means the leverage is starting to spread.
They have stopped asking for headcount in places they used to ask for it. Quietly, often without being asked. The hiring request that didn’t come in. This is usually the cleanest financial signal in the entire program, and it appears nowhere on a dashboard.
A recurring task that used to need a person now arrives on a schedule, and you’re not sure who set it up. The Monday report that’s simply there. The reconciliation that’s already done. This signal is different in kind from the seven above, because the person benefiting from it may not be an AI user at all and may not know an agent is involved. The leverage isn’t theirs. Trace the standing job back to whoever designed it and pointed it at a cadence: that person is operating at leverage for the whole team, and they’re the hardest of all to spot, because the only mark they leave is work that quietly keeps happening.

A usage report sorted by person gets you the first signal and narrows the search for the rest. An ROI chart gets you none of them. Weekly business reviews will surface the others quickly.

Four names give you a starting point

Start with four names. Three are your honest guess at the people most likely to be operating at AI leverage right now. They are not necessarily the loudest or most senior. They are the people quietly outproducing what you used to expect from their role. Add a fourth name of a different kind: the person most likely to have built a standing workflow that lifts other people without anyone noticing. The first three are doers. The fourth is a designer and the least likely to show up in a report you already run.

Read the consumption report, sorted by person, against those four names. Where the list agrees, you have evidence. Where it surfaces someone you didn’t write down, you have found a person doing something with the tool that nobody asked them to do.

Thirty minutes with each of them, separately, is enough. Three questions usually suffice.

What are you using, and how do you have it set up?

What work used to take you a day or a week that now takes you an hour?

What would you be able to do if I gave you a budget for tools, time, or people to support this?

What you’ll get from those three conversations is the shape of the program your organization actually has, not the shape on the slide. The gap between them is your real roadmap.

Claude Team, ChatGPT Business, M365 Copilot, and Google Workspace retain flat business or workspace seats. Metered consumption belongs to agent workloads, APIs, credit packs, and Claude Enterprise. Last verified July 29, 2026. ↩
Alexander Bick, Adam Blandin and David Deming, “The Impact of Generative AI on Work Productivity,” Federal Reserve Bank of St. Louis, February 27, 2025, on the November 2024 wave of the Real-Time Population Survey. Workers who used generative AI in the prior week reported saving 5.4% of their work hours, about 2.2 hours per week. Self-reported, and the whole-workforce equivalent is closer to 1.1-1.4%. Last verified August 2026. ↩
LayerX, “State of AI Usage Report 2026,” May 28, 2026, based on enterprise browser telemetry rather than survey response. 18% of enterprise users used AI in a given week; half of users had 12 conversations or fewer while the top 5% had at least 144, at an average of 18 prompts per conversation against an overall average of 2. Last verified August 2026. ↩
OpenAI, “State of Enterprise AI,” December 2025, across more than a million business customers. Workers at the 95th percentile of usage send roughly 6x the messages of the median employee at the same company, 17x for coding messages. This is message volume, not output. OpenAI’s own 2026 follow-up reframes the gap at firm level (frontier firms use about 3.5x as much intelligence per worker) and finds that message volume explains only 36% of the difference. Last verified August 2026. ↩
Anthropic, “How AI Is Transforming Work at Anthropic,” December 2, 2025: 132 engineers and researchers, 53 interviews, 200,000 internal Claude Code transcripts. Self-reported Claude use in 59% of work, average self-reported productivity gain around 50%, and 27% of Claude-assisted work reported as work that would not have been done otherwise. Anthropic notes its staff have early frontier-model access and that the findings do not generalize. Last verified August 2026. ↩

Recognizing Leverage