Your Best People Get 5x From AI. Your P&L Sees 29%. The Gap.

Walk into almost any company in 2026 and you can find them: the AI super-users. The analyst who built a prompt chain that turns a full afternoon of reporting into eight minutes. The engineer whose agent harness ships pull requests while she sleeps. The marketer who runs an entire campaign workflow that used to need three people. Writer's 2026 enterprise AI report puts a number on what these individuals achieve: roughly 5x productivity gains. And yet the same report finds that only about 29% of organizations see significant ROI from generative AI, and just 23% from AI agents. Five times the output at the desk; less than a third of companies feeling it on the income statement. That gap is the whole story of enterprise AI right now.

The Individual-to-Organizational Leak

The instinctive reading of the 5x figure is optimistic: if our best people are five times more productive, surely the company is becoming dramatically more efficient. The data says otherwise, and the reason is a leak between two levels that most leaders never examine. Individual productivity and organizational productivity are not the same quantity, and they do not automatically connect. A super-user's 5x lives in their head, their personal prompt library, their browser tabs, and their muscle memory. None of that is, by default, a company asset. It is a personal one.

When that person solves a problem brilliantly with AI, the organization captures the output of that single instance — one faster report, one quicker PR. What it almost never captures is the method. The prompt that worked, the context that made it work, the failure modes the super-user learned to avoid, the judgment about when to trust the output and when to throw it away — all of that stays uncodified. So the next person who faces the same task starts from zero, reinvents a worse version of the same hack, and the organization pays the discovery cost again. Five times the speed for one person does not compound into five times the speed for the team, because nothing was shared, stored, or systematized.

This is the leak. The wins are real and they are large, but they are leaking out of the organization at the boundary between the individual and the system. The 71% of companies that do not see significant ROI are not staffed by people who are bad at AI. Many of them have super-users too. They simply have no mechanism to turn a super-user's private 5x into a shared, repeatable, owned capability. The gain enters the building through one person and leaves the building when that person closes their laptop.

Four Reasons Super-User Wins Evaporate at Org Scale

The leak is not a single failure but four, and they compound. Each one independently keeps an individual win from becoming an organizational one; together they explain almost the entire gap between 5x and 29%.

1. The Gains Are Trapped in Individuals

The first and largest reason is that super-user gains are encoded in people, not in systems. The prompt that works lives in a personal note. The agent workflow runs on the super-user's machine with their API key. The judgment about when the AI is wrong is intuition built from a hundred painful corrections that were never written down. When the work product is an asset but the method is tacit, the organization owns the fish and not the fishing. Worse, the asset is fragile: when the super-user changes teams, goes on leave, or quits, the 5x leaves with them, and the function reverts to its pre-AI baseline overnight.

2. No Memory or Skills Capture Across the Org

The second reason is the absence of shared memory. A super-user improves run over run because they remember what worked last time. Organizations rarely have any equivalent. There is no shared library of vetted prompts, no repository of reusable agent skills, no record of which approaches succeeded and which failed on which task. Every team is a sealed box that learns nothing from the box next door. The result is that the organization's collective AI competence is not the sum of its super-users — it is closer to the average of everyone, because the best practices never propagate from the few who have them to the many who do not.

3. The Last Mile of Integration Is Missing

The third reason is the hardest and the most under-resourced. Turning a personal hack into a production process is real engineering work: you have to generalize the prompt so it handles cases the super-user never hit, wrap it in error handling, connect it to the systems of record, add the evaluation that catches regressions, and document it so a non-expert can run it. This is the last mile, and almost nobody owns it. The super-user is busy being a super-user; the platform team has a backlog; the process stays a personal artifact because no one is funded to industrialize it. The 5x is achievable but not repeatable, and repeatability is where the P&L lives.

4. Measurement Is Anecdotal, Not P&L-Level

The fourth reason is that the wins are measured in stories, not in numbers that reach the income statement. "This saved me a whole afternoon" is an anecdote. It is true and it is encouraging, but it does not roll up. No one converts the afternoon saved into a cost per outcome, a cycle time, or a revenue-per-employee figure that a CFO can audit. So the organization cannot see its own gains, cannot prioritize which workflows to invest in, and cannot defend the AI budget when the disappointment data starts circulating. What you cannot measure at the P&L level, you cannot manage to the P&L level — and the super-user's 5x stays invisible to the business.

A Worked Example: The Analyst's Eight-Minute Report

Make it concrete. An analyst figures out that a particular prompt chain, fed the right three data exports, produces a board-ready summary in eight minutes instead of an afternoon. That is a clean 5x — for her. Now trace what happens to that win under each of the four leaks. It is trapped in her: the prompt lives in a personal doc, the judgment about which numbers to trust lives in her head. It hits no shared memory: the three analysts on the next team keep spending afternoons because nobody told them, and nobody could have, since there is no library to find it in. It never crosses the last mile: no one generalizes her chain to handle the quarters where the data is messy, so it stays a personal tool that only she can babysit. And it is measured anecdotally: she mentions in standup that AI "saved her a ton of time," which is true and completely invisible to finance.

The organization captured exactly one faster report and learned nothing transferable. Multiply that across every super-user in the building and you get the macro picture the data shows: a lot of real individual 5x, almost none of it reaching the income statement. The leaks are not exotic edge cases. They are the default path that every uncaptured win travels.

Individual Hack vs. Organizational Capability

The distinction at the heart of the gap is the difference between a hack and a capability. A hack is fast to create, lives in one person, and produces a single instance of value. A capability is slower to build, lives in a system, and produces value every time anyone runs it. Most organizations have an abundance of hacks and a scarcity of capabilities, and they confuse the two because both involve AI and both feel like progress.

The move from the left column to the right is exactly the work that produces organizational ROI, and it is exactly the work most companies skip. The 29% that capture real return are the ones who have built the muscle to convert hacks into capabilities — to take what the super-user discovered and turn it into something the whole function runs, measures, and improves. The mechanics of that conversion are the subject of our deep dive on building AI workflows that compound, where the difference between a one-shot prompt and a self-improving loop is laid out in full.

How to Systematize Super-User Wins

Closing the gap is not about finding more super-users or buying more tools. It is about building the pipeline that turns the super-users you already have into organizational capability. There are four components, and they map directly onto the four leaks.

Skills and Playbook Capture

Start by making capture a standing expectation, not a heroic act. When a super-user finds a workflow that delivers, the prompt, the context, and the judgment around it should be written into a shared playbook and, where possible, packaged as a reusable skill the rest of the org can invoke. This is the antidote to the trapped-in-people problem. The goal is that the method survives independent of the person — that a competent non-expert can pick up the playbook and reproduce most of the 5x without rediscovering it from scratch.

Shared Compounding Loops

Capture is static; loops are dynamic. A compounding loop is a workflow that gets better every time it runs because it has memory: it records what worked, feeds successful patterns back in, and accumulates skills over time. Building these at the organizational level — rather than letting each super-user maintain a private one — means the entire function improves run over run, not just one person. This is how an organization's AI competence stops being the average of its people and starts being the best of them, propagated and compounding.

Evaluation and the Last Mile

Someone has to own the industrialization. A workflow cannot be trusted by the whole team until it has an evaluation harness that catches its failure modes — the same scars the super-user earned by hand, encoded into automated checks. Funding the last mile means staffing the work of generalizing, hardening, and instrumenting the hack so it can run unattended at quality. Without evals, every rollout is a gamble; with them, a personal hack becomes a process a CFO can rely on.

Enablement and P&L Measurement

Finally, the wins have to be taught and counted. Enablement spreads the captured playbooks through training so the non-super-users actually adopt them. And measurement has to move from anecdote to the income statement: every systematized workflow should report a cost per outcome, a cycle time, or a revenue contribution that rolls up to the P&L. That is what makes the gain visible to leadership, fundable by finance, and defensible against the disappointment narrative. The same trap of mistaking activity for outcomes is what we dissect in our look at AI productivity theater and vanity metrics.

Measure the Org, Not the Individual

The deepest fix is a shift in what gets measured. Most AI dashboards track individual activity — prompts run, seats active, hours reportedly saved. These metrics flatter the super-user and tell you nothing about the organization. The metric that matters is org-level: did the cost per resolved ticket drop, did the proposal cycle time shorten across the whole sales team, did revenue per employee rise. Those numbers only move when individual wins have been converted into shared capability, which is exactly why they are the right thing to measure. They are immune to the leak, because they only register what actually reached the system.

When you measure the organization instead of the individual, the gap between 5x and 29% stops being a mystery and becomes a backlog. Each workflow where a super-user has a private 5x is a candidate for systematization, and each one you industrialize moves an org-level number. The 5x does not have to stay locked in your best people. It is a preview of what the whole function can do once you build the pipeline to release it.

The Leadership Job Nobody Owns

There is a reason the gap persists in companies full of smart people: closing it is nobody's job. The super-users are not incentivized to slow down and document; they are rewarded for their output. The platform team is buried in tickets. Finance does not know which workflows to instrument because the wins were never surfaced as numbers. The work of harvesting individual wins and turning them into shared capability falls in the gap between every function's mandate — which is exactly why it does not happen by default.

Fixing that requires a deliberate owner with a clear charter: find the super-users, capture their methods, fund the last-mile engineering, stand up the evals, run the enablement, and report the results at the P&L. This is not a tooling purchase; it is an operating capability that has to be staffed and held accountable. The companies in the 29% are not the ones with better AI tools — everyone has the same models. They are the ones who decided that converting individual wins into organizational ones was a real job and gave it to someone. The gap between 5x and 29% is, at bottom, an ownership gap before it is anything else.

Conclusion: Close the Leak, Capture the Gain

The 5x super-user is not the exception that proves AI works — it is the demonstration of what is possible and the indictment of how little of it organizations capture. The gap to 29% is not a technology problem. The technology is delivering. It is an operating problem: the wins are trapped in individuals, propagated by no shared memory, never carried across the last mile, and measured in stories instead of P&L. Each of those is fixable, and the fixes are concrete — capture, compounding loops, evals, enablement, and org-level measurement.

The organizations that close this gap will not do it by chasing more dramatic individual demos. They will do it by treating every super-user win as raw material for a shared capability, and by measuring success where it actually shows up — on the income statement, not at the desk. The 5x is already in your building. The work is to stop letting it leak out.