AI productivity is not the same as business value

A team tells you AI is saving them six hours a week. It is the kind of number that ends a meeting happily. The harder question, the one worth asking before anyone celebrates, is simpler than it sounds: and then what? Where did those six hours go, and did anything the business actually cares about move as a result?

Most of the time, nobody can answer. The time saving is real, and it is also where the story stops. This is the gap at the centre of AI adoption right now, and it is the difference between productivity, which is about tasks, and value, which is about the business.

The uncomfortable evidence

The pattern shows up in the numbers. An MIT Media Lab report, The GenAI Divide: State of AI in Business 2025, reported that despite very high adoption, around 95% of enterprise generative-AI pilots had produced no measurable profit-and-loss impact. That figure deserves caution: it is an industry report rather than peer-reviewed research, its evidence is built from interviews, a survey and a review of public deployments, and the headline number has been contested by other commentators. Treat it as a provocation, not a precise measurement. But the direction it points, plenty of adoption and very little measured value, matches what a great many leaders quietly report.

It is not that the productivity gains are fake. In a rigorous study of customer-support agents, generative AI produced a real 14% average rise in output. The point is that a genuine task-level saving does not travel up to the business on its own. Why it stalls has a name. Brynjolfsson, Rock and Syverson described a "productivity J-curve": when a general-purpose technology arrives, the measured payoff lags, because the value only appears once organisations make the unglamorous complementary changes around it, redesigned processes, new ways of working, retrained people. The tool is the cheap part. The reorganisation is where the value actually lives, and most pilots never get that far.

Four levels of AI value

It helps to see value as a ladder with four rungs. At each step, the gain either travels up or leaks away.

Task efficiency. A specific task takes less time or effort. Drafting the email, summarising the call, cleaning the spreadsheet. This is where almost all AI use starts, and where almost all of it stays. Real, measurable, and on its own, worth very little.

Workflow improvement. The process around the task is redesigned so the saving is not immediately swallowed. If the draft is faster but it still waits three days for the same approval, the workflow ate your gain. Capturing the saving usually means changing a step, not just speeding one up.

Capacity release. The freed time is deliberately pointed at something worthwhile. This is the rung leaders most often skip. Time saved does not redeploy itself; left alone, it quietly refills with more of the same low-value work. Capacity is only released if someone decides, in advance, where it should go.

Business outcome. Something the business actually measures moves: revenue, faster service a client will pay for, better decisions, lower risk, more work handled without more headcount. This is the only rung that shows up in the results, and you reach it only by climbing the three below it on purpose.

Time saved is not value. It is only the chance of value.

What the leak looks like

Picture a law firm where an assistant cuts twenty minutes off drafting a routine document. At the task level, the win is obvious. But if each fee earner simply absorbs that twenty minutes into a slightly longer lunch or a fuller inbox, nothing reaches the business. The same saving becomes valuable only when the firm decides what the freed hours are for, taking on more matters, shortening a turnaround clients notice, or moving senior people onto higher-value advisory work, and then checks whether that actually happened. (An illustrative example, not a specific firm.)

A marketing agency shows the same shape. AI that halves first-draft time is task efficiency. Using the recovered time to pitch for more work, or to give clients faster cycles they will renew for, is a business outcome. Between the two sit the decisions most teams never make.

The honest limits

Two qualifications, because the argument can be pushed too far. First, not every benefit has to land on this quarter's P&L to be real. Reduced risk, better decision quality and relieved pressure on overstretched people are legitimate outcomes, even when they are harder to put a number on. The J-curve is a genuine warning here: early value can be real and still hard to measure, so do not kill a promising pilot just because the spreadsheet has not caught up. The discipline is to distinguish "hard to measure" from "imaginary", and to name the outcome you are aiming at even when you cannot yet price it precisely.

Second, chasing a measurable outcome too aggressively can push teams toward easy metrics that miss the point, volume of output rather than quality of result. A faster route to worse work is not value either. The measure has to track something that genuinely matters, which is a judgement, not a dashboard setting.

What to do next

Take one place where AI is already saving time and trace it up the four rungs. Did the workflow change to capture the saving, or did the next step swallow it? Where did the freed capacity actually go? What business measure was supposed to move, and did it? Then do the more useful thing: before your next pilot, decide in advance where the recovered time will be redeployed and pick the one outcome you will measure. Deciding that up front is most of the work, and it is the part almost everyone leaves out.

The tool

To make that trace concrete, I have built the AI Value Conversion Worksheet: it walks a claimed time saving up the four levels, asks the two questions that decide whether it becomes value, what happens to the freed time and how the outcome will be measured, and leaves you with a defensible answer rather than a hopeful one.

Download the AI Value Conversion Worksheet (PDF)

Converting savings into outcomes is precisely the work of a 90-day adoption build with a team: redesigning the workflow, deciding where capacity goes, and agreeing what to measure. It follows directly from choosing the right use case in the first place, and it explains a lot about why so many promising pilots never scale.

Sources and further reading

• MIT Media Lab (Project NANDA), The GenAI Divide: State of AI in Business 2025. Industry report, not peer-reviewed; the 95% figure has been contested. Used as directional evidence of the adoption-to-value gap.
• Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161, 2023. Independent. Source for the real 14% task-level productivity gain.
• Brynjolfsson, Rock and Syverson, The Productivity J-Curve, NBER Working Paper 25148, 2018. Independent. Source for the lag between technology adoption and measured productivity, driven by intangible complementary investment.
• UK Government, AI Opportunities Action Plan, January 2025. Note its productivity estimates draw on OECD and third-party analysis rather than being independent measurements.