Why AI is making employees more productive without moving the P&L

The 80/37 gap, in one line

The headline from McKinsey's 2026 State of AI survey is not that AI is failing. It is that AI is working for individuals and stalling for the business. Across 1,719 leaders, 80% reported that AI had improved their own productivity, only 37% reported any EBIT impact at the enterprise level, and just 6% qualified as AI high performers. The 37% has not moved since 2025. In plain terms: adoption is close to universal and profit impact is close to flat. The tools are in the building. The earnings are not. That is the single most important number for any business deciding how much to spend on AI next year, because it says the limiting factor is no longer access to the technology. It is what the organisation does with it once it is there.

Why personal productivity doesn't reach the P&L

Personal productivity is real. A lawyer drafts a letter faster, an analyst summarises a report in seconds, a manager clears an inbox before lunch. The problem is what happens to that saved time. If the workflow around the person has not changed, the time is quietly absorbed back into the same way of working. It does not turn into more matters closed, more deals won or a lower cost to serve. A faster version of the same process is not a better business.

This is also what IBM found. In its 2026 research with Oxford Economics, more than three in four respondents said isolated AI delivers limited value unless the wider process is redesigned. The tool moves one step. The result depends on the whole chain of steps around it, and in most organisations those steps are untouched.

What the companies seeing profit do differently

The organisations that do see a financial result are not using better models. They are changing the work. In McKinsey's data, 74% of high performers had redesigned at least some workflows to capture value from AI, against roughly a quarter of everyone else. They also tend to use AI for growth, not only for efficiency, and they put senior ownership behind it.

The separator is the operating change, not the technology. Two businesses can buy the same AI and see completely different returns, because one rebuilt the process the AI sits inside and the other bolted it onto the process that was already there.

Where the return actually leaks

Even when a business commits, the return often does not survive the journey to production. IBM, with Oxford Economics, found that only 37% of AI initiatives had delivered the value leaders expected by the end of 2025, and that 21% of AI ROI was lost to friction in systems and processes that were never redesigned. Disconnected tools, data stored for reporting rather than for a live process, and manual handoffs between systems quietly eat the return before the model is ever the problem. For most organisations the constraint is the estate the AI has to work across, not the intelligence of the AI itself. That is why integration and data access are part of the work, not a technical appendix to it.

What to measure before you expect a result

If the gap is an implementation problem, the fix starts with a measurement, not a licence. In McKinsey's 2025 research, tracking well-defined KPIs for AI was the single practice most closely tied to bottom-line impact. In practice that means choosing one workflow, writing down what it costs today in time, money or lost opportunities, and agreeing what "better" will look like before anything is built. Without a baseline and an owner, a project reports activity: logins, messages, hours "saved". With them, it produces a number the business can actually see on the P&L.

What this means for your business

The 80/37 gap is good news, in a way. It means the limiting factor is not access to AI, which is now cheap and everywhere. It is the decision about where AI and technology belong in the work, and the discipline to change the process rather than decorate it.

That is the order we work in. Understand how the work is actually being done. Find where time, money and opportunities are leaking. Improve the underlying workflow. Then decide what technology, whether AI, automation, integration or custom software, is actually required, build or connect it, implement it into the operation, and measure the result against the baseline. Audit, Design, Build, Implement, Improve.

Sometimes the right technology already exists and needs connecting. Sometimes it needs to be built. Either way, the thing that moves the P&L is the change to the work, not the tool sitting on top of it.