Productivity vs Realized Capacity

Suppose the quarterly AI update of a regional water authority carries this line. Its new assistant drafts survey reports, correspondence and work orders for 400 staff, who say it saves each of them about two hours a week. Over a 45-week working year that is 36,000 hours. At a loaded cost of 50 an hour, the update reports 1.8 million in savings.

Now go looking for the 1.8 million. Headcount has not moved, overtime has not moved, and a new property still waits 20 working days for a water connection. The hours were probably real. Staff did finish their drafts faster. But nobody can say what the hours became.

A field study of a US technology company, where AI use was entirely voluntary, found where such hours go: people took on a broader scope, worked at a faster pace and let work run into lunch breaks and evenings, often without being asked [@haas-2026-intensification; @ranganathan-ye-2026]. The hours were spent. They were just not spent on anything management had chosen.

Three claims that sound alike

Much of the confusion about AI productivity comes from using one word for three different claims. They need separate names, because only one of them belongs in a business case.

A productivity gain means the same task needs fewer resources: a survey report that took 60 minutes now takes 40. As AI and Workforce Productivity showed, that task is 50 percent faster, and 50 percent faster frees a third of the time, not half; in general, X percent faster frees X divided by (100 plus X) of the time. Capacity created is the 20 minutes now available for other work. It is an option the organization holds, not money. Realized capacity is what happens when the organization uses those minutes to produce something it can measure: more applications processed, a shorter queue, a hire it did not need to make.

The distance between the second and third layers is a management decision. AI creates the capacity. Management decides whether it becomes value or dissolves into the working day. That is the conviction this chapter argues for: productivity improvement is potential value; realized capacity is value the organization actually converts into useful output, avoided cost or a measured business result.

Where unplanned time goes

Freed time that nobody assigns does not sit idle. It is absorbed, and the absorption is mostly invisible to the dashboard that reported the saving.

The reason is simple. A saved hour has no label. If nobody decides what it is for, the default is whatever work is nearest: a task that had belonged to someone else, a second assistant running in the background, a colleague’s AI-drafted report that now needs checking. One person’s acceleration becomes another person’s workload¹. Some of that extra activity may well be useful. The point is that nobody chose it, nobody priced it and nobody measured what it produced.

The same pattern shows up in the economic records. As AI and Workforce Productivity reported, a study of 25,000 Danish workers found no measurable effect of AI chatbots on earnings or recorded hours two years after ChatGPT’s launch². Tasks got faster; the records did not move. This chapter is about the missing step in between.

Is there work for the hours?

The first filter is demand. Productivity and utilization are different things. Productivity is how efficiently people produce output. Utilization is how much of the available capacity is actually used for productive work. A gain in the first creates value only if there is useful work to absorb it.

A team with a real backlog, such as connection applications waiting for a survey, can turn freed hours into finished work almost immediately. A team whose demand is flat or seasonal cannot, unless someone moves the capacity elsewhere or makes an explicit decision about staffing. So the first question for any productivity case is plain: is this team overloaded, or does it already have slack? The answer tells you whether the hours have anywhere useful to go.

The narrowest step sets the pace

Even where demand exists, the gain can stall inside the workflow. Think of the process as a pipe. Widening one section does not raise the flow if another section is narrower.

Suppose the water authority’s connection process takes 20 working days: 1 at intake, 4 for the site survey, 12 waiting for and receiving engineering approval, and 3 for scheduling. AI makes intake four times faster, so it now takes a quarter of a day. The whole process falls from 20 days to 19.25, an improvement of under 4 percent. The computer architect Gene Amdahl made the general argument in 1967 about computer processors, and it is now known as Amdahl’s law: the part of a job you do not speed up limits how much faster the whole job can get³.

It is worse when the slow step is a true constraint. If engineering approves a fixed number of applications a week, a faster intake simply delivers applications to the approval queue sooner. The theory of constraints, set out by Eliyahu Goldratt in 1984, makes this the central rule of operations: the output of a system is limited by its constraint, and improving anything else does not raise output⁴. The easiest task to automate is often not the constraint. Hours freed in the wrong section of the pipe stay potential value until they are moved to the section that limits the business, for example by having intake staff prepare complete technical packs that spare engineers a round of queries.

Count the net gain, not the gross

Capacity also leaks inside the task. The figure the tool reports is usually the gross saving at the step it touches. What the business gets is the net saving across the whole job.

Suppose the assistant saves a surveyor 30 minutes drafting a report. Twelve minutes go on checking it, because a wrong pipe size in a report is expensive to fix later. Eight more go on re-keying the result into an asset system that does not connect to the assistant. The net gain is 10 minutes, a third of the headline. The checking is not waste; much of it is exactly the human oversight you want. But it must be counted, along with extra meetings, new approval steps and the review time that lands on colleagues. The measure that matters is the net workflow gain: the end-to-end time saved after checking, rework, hand-offs and waiting are subtracted.

Six destinations for released capacity

Once you know how much net capacity a change releases, and where in the workflow, the management decision becomes concrete. Suppose, for illustration, the net gain across the authority adds up to the equivalent of 12 full-time people. That does not mean 12 people can be removed. It means there are six places to send the capacity, and each has a different economic result.

More output clears backlogs and shortens waits. Redeployment moves people from routine work to the constraint or to work that needs judgment. Avoided hiring absorbs growth or a seasonal peak without new staff; as Where AI Reduces Cost showed, avoided cost is real but it is not a lower run-rate. Cost reduction is a lower payroll or a cancelled contract, and it does not happen by itself: it needs an explicit decision about staffing, skills, contracts and transition costs. Quality means more review, fewer errors and fewer complaints. Speed and learning means using the time to improve the process itself.

None of these is morally superior. Redeployment can be worth more than a headcount reduction, and keeping some slack on purpose, for resilience during a drought or a burst-main season, is a legitimate choice. What is not legitimate is reporting the hours as savings while making no choice at all.

Plan the capacity before it appears

The remedy is to decide where capacity will go before the tool is scaled, and to report only what has climbed far enough to be proven.

The capacity ladder gives every AI initiative a place to stand. The first rung is gross time saved. The second is capacity available after leakage. The third is capacity assigned to a named destination, with an owner. The fourth is additional output or avoided cost. The fifth is value measured against a baseline, built the way Baselines, Metrics and Measurement described⁵.

A capacity plan is the document that moves an initiative up the ladder. Before scaling, it answers four questions. Which roles and steps are affected, and how much net capacity will be released, and when? Where will it go, and is that step the constraint? Who owns the benefit? Which measure, read at which date, will show that the capacity was used? In the business case, keep potential and realized value apart; how to discount the first into an expected figure, using a realization rate, is the subject of AI ROI and Value Realization.

Story: the earnings desk that chose its destinations

A well-documented example of capacity planned and realized predates generative AI, which makes it useful: it shows the management logic without the hype.

Before July 2014, reporters and editors at the Associated Press wrote about 300 stories a quarter on US corporate earnings. The work was repetitive and error-prone, and the newsroom disliked it. That month AP began generating earnings stories automatically from structured financial data, using natural-language generation software. By January 2015 it was producing more than 3,000 stories a quarter, about ten times as many as before, and the automated stories contained fewer errors than the manual ones⁶.

The instructive part is what AP did with the capacity. By its own estimate, automation freed about 20 percent of the time its journalists had spent on earnings reports; freeing a fifth of the time is the same as the work becoming 25 percent faster. AP did not book those hours as savings. From the start it framed the project as freeing reporters from data processing for higher-level reporting, and no jobs were cut. The capacity went to two destinations. The first was more output for which there was demand: coverage of companies AP had never had the capacity to write about, which customers welcomed because it covered businesses in their own states and regions⁷. The second was redeployment: the freed time went into breaking news and investigative and explanatory journalism⁶.

The value was also measured outside AP. Because the automated stories were rolled out to companies in stages, three accounting researchers could treat the rollout as a natural experiment. Many of the newly covered firms had previously received little or no media attention, and the automated articles increased their trading volume and liquidity⁸. The extra output reached readers and changed behavior.

The case has limits. The 20 percent is AP’s estimate, not an audited figure, and the quality of the redeployed journalism was not measured independently. But compare it with the water authority’s board pack. Both freed time. AP framed the destinations from the start, chose work that had waiting demand, and could point to what the capacity produced. That is the difference between hours reported and capacity realized.

What this means for leaders

Treat every hours-saved figure as the bottom rung of the ladder: a signal worth having, not a result. Ask for net figures, find the constraint before you fund the speed-up, and require a destination and an owner before an initiative scales. Be neutral about which destination is chosen and strict about whether one has been chosen. And watch for the quiet cost of not choosing: teams that are busier, more stretched and no more productive in any way the business can see.

Check yourself

1. A 30 percent productivity gain means 30 percent lower cost.MythFact
2. Time saved by AI that nobody assigns usually sits idle.MythFact
3. Making one step of a process four times faster can leave the end-to-end time almost unchanged.MythFact
4. The task with the largest time saving is the best AI opportunity.MythFact
5. Redeployed capacity can be worth more than a headcount reduction.MythFact
6. Unused capacity is always waste.MythFact

Reflection: follow the hours

What comes next

AP could count stories, newly covered companies and errors every quarter, years before researchers published the effect on trading. It knew early that the capacity was being used. Some indicators move first, and others follow later and prove the impact. The next chapter, Leading vs Lagging AI Metrics, separates the two.