Everyone uses AI. Almost nobody gets returns.

Here are the two numbers that should sit at the top of every AI conversation in your firm this year. First, from 5WPR’s AI at Work Index: 88 percent of organizations now use AI in at least one business function, up from 78 percent a year earlier. Second, from the same index: only 5 percent report that their formal AI investments have produced transformative returns. Everyone uses AI. Almost nobody gets returns. Call it the 5 percent problem — and notice that it is not a technology problem, because the technology is the part that worked.

01Access was never the bottleneck

The index’s quietest finding is its most damning one. Only 40 percent of companies have purchased official AI subscriptions, yet employees at more than 90 percent of organizations regularly use personal AI tools on work anyway. Your staff solved access themselves, with or without procurement. The pipe is already full.

That single fact retires the most common AI strategy in professional services: buy licenses, wait for returns. If tool access were the bottleneck, the bottleneck was removed years ago, at consumer prices, by your own people. When nearly everyone has the tools and almost no one has the returns, the constraint lives somewhere else — in how the work is shaped around the tools.

02What the returns minority does

The index has one more cut, and it is the hinge of this argument. The small band of organizations reporting real earnings impact — roughly 6 percent, those attributing more than 5 percent of EBIT to AI — share one behavior: in the index’s wording, they have “redesigned the underlying work around AI capability.” The work itself changed shape — not the models, not the budgets.

Udacity’s 2026 State of AI at Work arrives at the same place from the people side: “In 2026, AI success depends less on tools and more on people.” The report names that advantage the Human Premium — “the unique advantage created when skilled people work alongside AI.” Two studies, different methods, one sentence underneath: the returns sit on the people-and-process side of the toolbar, which is exactly where almost nobody is looking.

03What “redesigning the work” actually means

“Redesign the work” is easy to say and easy to wave at an org chart. In practice it is three concrete changes, and none of them is a software purchase:

• —Decision rights, written down. What an agent may do on its own, what it may draft for human sign-off, what escalates and to whom. Unwritten, every output gets handled ad hoc — which means your most senior people end up as an unplanned QA department at midnight.
• —Review discipline. Agent output lands in a review state before anything client-facing moves, with a named reviewer and sampled audits. The session — or the analyst — that produced the work is the worst-placed judge of it.
• —The human premium. Your skilled people move up a layer: from producing the draft to interpreting, deciding and answering for it. That judgment layer is the defensible advantage, and it is exactly what Udacity’s report is naming.

This is the third way, now with the market’s own numbers attached: agents execute under constraints, humans own judgment, and the line between them is drawn deliberately — by decision rights and review gates — not by whatever happens at the keyboard.

04A worked example: the monthly client report

Take the most boring recurring workflow in a professional firm and redesign it. (A composite, but every element is one we have seen in the wild.) Before: forty licenses and hope. Each analyst assembles the monthly client report by hand in their own format, it takes most of a day, the partner reworks three of them at night, and nobody can say what changed since AI arrived. Usage is high; the P&L sees nothing.

The redesign is one page and one gate:

• —The agent assembles the draft report from the source systems against a written definition of complete — required fields, checks run, exceptions flagged.
• —Decision rights live on the same page: the agent may fetch, compute and draft; any variance beyond a stated threshold and any recommendation goes to the partner; nothing client-facing sends itself.
• —Every report lands in a review state with a named reviewer, and a sample gets a line-level audit each month. The agent proposes; a human ratifies.
• —The partner spends the recovered hours on the two paragraphs that interpret the numbers — the reason the client keeps the firm.

The arithmetic is six hours down to one. But the arithmetic is not the return. The return is structural: consistent cycle time, an auditable trail, and your most expensive judgment moved to where it bills. That is how an AI line item finally reaches the P&L — not by typing faster.

05The honest caveats

• —“Transformative” is self-reported, and survey respondents flatter themselves. Treat 5 percent as a direction, not a constant; the starker cut is the index’s finding that the other 95 percent report zero measurable impact on profit and loss.
• —The index is a synthesis of more than 80 studies, so its cuts come from different baselines (adoption via McKinsey, returns via MIT’s Project NANDA). The pattern — redesign correlates with returns — is the finding; the decimals are not load-bearing.
• —Redesign is the slow, unglamorous option, and training is part of it, not garnish: the index cites BCG finding that five hours of training plus visible leadership support moves positive sentiment from 15 to 55 percent. The Human Premium is bought with hours, not licenses.

06The Monday-morning version

Pick one recurring workflow — reporting, intake, invoice chasing. Write the decision-rights page: may do alone, may draft for sign-off, must escalate. Stand up the review gate so nothing client-facing ships unaudited. Move your best person up a layer. If you want the before picture first, our AI Work Mix scorecard maps how your team’s work actually splits today — it is free and takes five minutes.

The 5 percent did not find better AI. They built better work for the AI everyone already had.

Sources: 5WPR, “The AI at Work Index” (April 2026; 5wpr.com/research/ai-at-work-index/) — a synthesis of 80+ published studies: adoption at 88%, up from 78% (via McKinsey State of AI 2025); 5% transformative returns and 95% zero measurable P&L impact (via MIT Project NANDA); the ~6% reporting >5% EBIT who “redesigned the underlying work around AI capability” (index FAQ); 40% official subscriptions vs. 90%+ personal-tool use at work (via MIT Project NANDA); training figures (via BCG). Udacity, “State of AI at Work 2026” report page (udacity.com/report/state-of-ai-at-work-report-2026) — “In 2026, AI success depends less on tools and more on people”; the Human Premium, “the unique advantage created when skilled people work alongside AI.”