Relief and Buffers

The AI made everyone faster. The same people still work late, and every other team's release still waits for the same small team.

Definition

Relief is freed capacity used to lower the load on overloaded people, and keep it lower. A buffer is spare capacity held on purpose where peaks hit.

Relief has happened when overtime and after-hours work fall, and stay down. A buffer has worked when a peak passes and leaves no backlog for the next week. Between peaks, a buffer may sit unused. That is not waste. It is held for the next peak.

Relief and buffers both look like waste on a utilization report. That report does not show the cost of overload.

The recognition pattern

• Overtime and after-hours work did not fall after the AI arrived. People still send messages after dinner.
• The same few specialists keep several teams waiting. Each team sees only its own delay.
• After each peak, a backlog lasts into the next week. Sometimes the next peak arrives before the backlog clears.

Why freed time does not reach the overloaded

New work consumes the freed time

Work spreads into lunch breaks, then into evenings. For eight months, researchers followed about 200 employees of a US technology company. The employees' work spread the same way. AI did not give them spare time. It widened what they took on (Ranganathan and Ye, 2026, initial findings of a study still in progress). A working paper points the same way: in jobs more exposed to AI, workers put in longer hours and had less leisure (Jiang et al., 2025).

New tasks, wider scope and a faster pace take the freed time, unless someone explicitly protects it for relief or a buffer. So the load on the same people never falls. Time set aside for improvement is taken back the same way, as Capacity to Improve shows.

Under load, people work differently

When more work arrives, people hurry. In hospital patient transport and heart surgery, staff worked faster when their load rose. But sustained overload slowed them down, and in heart surgery it was linked to higher patient mortality (KC and Terwiesch, 2009).

Each additional patient per nurse was associated with about 7% higher odds that a surgical patient died within 30 days. The association held in hospitals in Pennsylvania and in nine European countries (Aiken et al., 2002 and 2014).

Relief can more than pay for itself. When Queensland set minimum nurse-to-patient ratios, researchers estimated that the 27 hospitals under the ratios avoided 145 deaths, 255 readmissions and 29,222 hospital days. The avoided costs came to more than twice the roughly A$33 million spent on the extra nurses over the first two years (McHugh et al., 2021).

Relief for the people every team waits for

In a constructed example, eight product teams send work to the same three security reviewers. The sharing alone does not make them worth relieving; their queue does, because it lasts from week to week and delays all eight teams. Take load off the reviewers, and all eight teams wait less. Faster Delivery shows how much less.

None of the studies here measures the gain to every team that waits for shared specialists in knowledge work. That gain is inferred from the mechanism. The nearest measured case is a model built on five years of data from one hospital's 16 inpatient units. When the bottleneck unit is full, patients who need a bed there stay in beds in other units. Its authors estimated that one more bed in the bottleneck unit would free about four beds' worth of capacity across the hospital. That is enough for about 380 more hospital visits a year. A bed at the bottleneck would free about three times as much as one added to a unit chosen without regard to the bottleneck (Kim, Zheng and Brown, 2024).

Se and Schenk (2026) argue that where freed time sits matters as much as how much there is.

A buffer absorbs the peak

One operating room sits unbooked, held for emergencies. When a children's hospital started doing this, the share of one class of urgent cases that reached surgery within 12 hours rose from 52% to 58%. Cancellations of planned operations because of emergencies fell from 1.5% to 0.7% (Heng and Wright, 2013).

When demand varies, something absorbs the swing: spare capacity, inventory or the customer's wait. Toyota chose inventory. After the 2011 earthquake, the company had suppliers hold two to six months of chips. Early in the 2021 shortage, it was largely unscathed (Shirouzu, 2021). If nothing is held in reserve, the peak lands on the customer's wait or on people's extra hours.

What it costs

Where the load is highest, the late nights continue, and so do errors, burnout and people leaving.

AI made the teams faster, but much of their gain waits in the shared specialists' queue before it reaches a customer.

Boundaries and distinctions

Relief is not absorption. Absorption just happens: work soaks up freed time while nobody decides. Relief is chosen, and someone protects it.

Relief is not a cost cut. The same freed hours can lower the load or lower the headcount, but not both in full. A later page in this series covers when it pays to cut.

More spare capacity is not always better. A buffer belongs at the step where peaks hit. Spread the same hours across every step, and the hours at quiet steps never meet a peak.

The evidence is mostly from hospitals. Read the figures as hospital figures. The mechanism carries over to knowledge work. The numbers do not.

Example

In the security example, the teams and the reviewers both use AI. The eight teams produce about 50% more changes, and the three reviewers read each change faster. The reviewers' saved time goes straight to the longer queue. They still work two late nights a week. Every team's release waits about a week for review.

Two moves change that. First, routine checks move to the teams, each backed by an AI check that the reviewers trust. The teams spend some of their own freed time on the routine checks. Second, a fixed share of the reviewers' week stays unbooked, held for urgent reviews and peaks. The first move is relief. The second is a buffer, and it fits only because the first move takes enough work out of the queue. Without the first move, the unbooked hours would make ordinary reviews wait longer.

Relief has happened when the two late nights stop and stay stopped; the buffer has worked when each peak's backlog clears within the week. All eight teams should then wait less for review.

The pattern holds beyond security review: name the shared people, measure their load and the queue in front of them, send them a share of freed capacity, and protect that share. Someone has to own that protection. Otherwise the unbooked hours fill with ordinary reviews, and the late nights return.

Applied test

1. List the people or teams that several teams wait for. Keep only the people and teams whose queue lasts from week to week.
2. For each, record three measures for the last month and for a month before the AI arrived: their overtime and after-hours work; how long work waits in their queue; and how long each peak's backlog lasted.
3. Check whether the AI saved time in the work those people and teams do, or in the work that reaches them. If it did and the three measures did not fall, find what took the saved time: new work, wider scope or more volume arriving.
4. Choose relief, a buffer or both. Decide the share of freed capacity each gets, who protects each share, and when each outcome will be judged by its own test. Relief has happened when overtime and after-hours work fall and stay down. A buffer has worked when peaks pass without a backlog carried into the next week. Treat the queue wait as a supporting measure, not the test.

If you cannot move capacity yourself, take the three measures to whoever can.

Sources and lineage

Relief and Buffers is part of Applied Enterprise AI, a body of work developed by Curtis Hibbs and Joshua Barnes. The hospital studies provide the strongest measured evidence on load and relief. The page uses their findings to explain the mechanism in knowledge work. It also draws on:

• Ranganathan and Ye, AI Doesn't Reduce Work—It Intensifies It, Harvard Business Review (February 2026).
• Jiang, Park, Xiao and Zhang, AI and the Extended Workday: Productivity, Contracting Efficiency, and Distribution of Rents, NBER Working Paper 33536 (2025).
• KC and Terwiesch, Impact of Workload on Service Time and Patient Safety: An Econometric Analysis of Hospital Operations, Management Science 55(9) (2009).
• Aiken et al., Hospital Nurse Staffing and Patient Mortality, Nurse Burnout, and Job Dissatisfaction, JAMA 288(16) (2002).
• Aiken et al., Nurse Staffing and Education and Hospital Mortality in Nine European Countries: A Retrospective Observational Study, The Lancet 383(9931) (2014).
• McHugh et al., Effects of Nurse-to-Patient Ratio Legislation on Nurse Staffing and Patient Mortality, Readmissions, and Length of Stay: A Prospective Study in a Panel of Hospitals, The Lancet 397(10288) (2021).
• Kim, Zheng and Brown, Identifying the Bottleneck Unit: Impact of Congestion Spillover in Hospital Inpatient Unit Network, Management Science 70(7) (2024).
• Heng and Wright, Dedicated Operating Room for Emergency Surgery Improves Access and Efficiency, Canadian Journal of Surgery 56(3) (2013).
• Shirouzu, How Toyota Thrives When the Chips Are Down, Reuters (March 10, 2021).
• Se and Schenk, AI Saved Time. Where Did the Value Go?, Turing Post #7 (July 2026).