In June 2013, an administrative assistant at the Phoenix VA Health Care System noticed a chart. It hung, very large, on the wall of Room D641, and it showed a steady decline in the number of days new patients waited for an appointment. The latest figure, for April 2013, was 36.8 days.
He had been asking why a veteran who requested a primary care doctor in mid-May still had no appointment in June. One veteran he named in his email had been waiting more than 60 days. “Why the variance?” he wrote to a group of colleagues, and he suggested someone check how the numbers were calculated before the director reported them outside Phoenix.
A year later, the VA inspector general checked. It took a sample of 226 new-patient appointments from Phoenix in fiscal year 2013. VA’s national data said those veterans had waited 24 days on average for their first primary care appointment.
The inspector general found 115.
Down the chain
The number on the wall started a long way from Phoenix. In fiscal year 2011, according to the Government Accountability Office, the Veterans Health Administration shortened its goal for scheduling appointments from 30 days to 14. Wait-time measures went into the performance contracts of regional network directors and medical center directors, which determine their performance pay.
In 2013 the Phoenix director made access to primary care her “Wildly Important Goal.” Posters around the facility showed each clinic’s success or failure on a local measure: every doctor’s third-next available appointment within 7 days. At one VA location, a scheduler told investigators, each supervisor received a list of the schedulers who had exceeded the 14-day goal.
That is goal cascading: split a big goal level by level, until each person holds a piece they can move. Done well, it is how a sentence at the top becomes work at the bottom. At Intel in 1980, the target of design wins in Operation Crush went down to each salesperson as a quota of roughly one a month. Robert Kaplan, describing companies that used his balanced scorecard well, said they “cascaded the high-level strategy down to the operating business units and the support departments.”
In the VA, the number made it all the way down. Not everything else did.
What the scheduler could move
A 14-day wait depends on how many doctors’ appointment slots exist. A scheduler cannot create slots. What a scheduler could change was where the clock started, and VA counted waits more than one way. For new patients, the inspector general wrote, the wait ran from the day a scheduler created the appointment, or added the veteran to the official waiting list. Most of the gap between 24 and 115 days, it found, came before that, between the day the veteran asked for care and the day the appointment was created. Other measures, the ones the GAO described in 2012, ran from a “desired date,” the day the veteran or provider wanted the appointment. Thirty Phoenix staff told investigators they had used the wrong desired date, producing a false wait of zero days. At the location with the supervisor’s list, schedulers said they changed it to the next open slot to keep their names off.
The GAO had warned in December 2012 that the wait-time data rested on that field and was unreliable. In 2014, VA’s own system-wide audit, which interviewed more than 3,700 staff, concluded the 14-day target “was simply not attainable” given the shortage of provider slots. Imposing it “before ascertaining the resources required,” the audit said, was “an organizational leadership failure.”
A cascade passes a number down at each level. Whether it also passes down the means to hit it, or a route for the level below to say “we can’t,” decides what the person at the bottom will optimize.
The response-time target
Imagine your director hands your six-person support team its share of a company goal: first reply to every ticket within two hours. You could divide it evenly and post six charts.
Before you do, you ask each person what would stop them. Two of them answer the overnight queue alone, and tickets that arrive at 2 a.m. can’t get a human reply by 4. The only lever they have is an automatic “we’ve received your message,” which would count.
So you send that answer back up with the number: two hours by day, or another person at night. The chart may look worse this quarter. It will also be measuring replies.
Splitting work for agents
AI agents cascade too. A lead agent breaks a task into parts and hands each part to a subagent.
In June 2025, Walden Yan of Cognition, the company behind the coding agent Devin, gave an example of how that goes wrong. Suppose the task is “build a Flappy Bird clone.” It becomes two subtasks: a moving background with green pipes, and a bird that moves up and down. In his example, one subagent builds a background that looks like Super Mario Bros. The other builds a bird that doesn’t look like a game asset. Each answered its slice. Neither had the whole.
That was an illustration, not a measured study, and Yan’s conclusion was his own: share full context and agent traces between parts, and prefer a single-threaded agent where you can. He added that copying the original task to each subagent is not enough on its own in a real system, where earlier steps carry decisions too. You can still check the basics yourself. When you split a job among agents, pass each one the original request and any decisions already made, not only its piece. Ask each to restate the overall goal in its first line, and read that line. Then judge the assembled result against the original request, not each part against its own subtask.
The clock that never started
After the inspector general’s interim report in May 2014, the Acting VA Secretary removed the 14-day goal from employee performance contracts.
The same report had found about 1,700 veterans in Phoenix waiting for a primary care appointment who were not on the official wait list at all. Their waits were in none of the numbers, because for them the reported clock had not started.