At four in the afternoon on Saturday, January 31, 2026, HM Revenue and Customs closed its phone lines. They are normally shut at weekends. That day HMRC had opened them at nine in the morning and stretched its webchat to ten times its usual Saturday capacity, because it was the last day to file a tax return online without an automatic £100 fine.
The busiest hour came after the phones went quiet. Between 5 and 6 p.m., 32,982 people filed. Between 11 p.m. and midnight, another 27,456 did. By the end of the day, 475,722 people had filed on the final day. HMRC estimated that about a million missed the deadline altogether.
They had not been short of time. An online return can go in from April 6, the day after the tax year ends, almost ten months before the deadline. On January 5, HMRC reported that of more than 12 million people expected to file, almost 5.65 million still hadn’t. “If you have yet to start,” HMRC’s chief customer officer said, “the clock is ticking.”
One caution. These figures record when people pressed submit, not when they began. A return can be started, saved and reopened as often as needed, and some are filed by accountants on clients’ behalf. A return sent at 11:40 p.m. may be the end of weeks of work. What the numbers do show is where the finish lands. The case has the rest.
The professor’s extra time
The name comes from Eliyahu Goldratt’s 1997 business novel Critical Chain. As a white paper by Scitor Corporation retells the scene, a professor sets an assignment due in two weeks. The students complain that it’s hard and needs more time, and the professor agrees. Looking back afterward, the students admit that, sure they now had plenty of time, they “put off starting until the last minute anyway.”
Goldratt’s point was about projects, not students. In a project, he argued, everyone pads their estimates with safety, and each level of management adds more. Then several habits quietly burn that safety. A review published by the Project Management Institute lists the first of them as “student syndrome or the tendency to wait until the last minute.” The trouble is timing. Spare time spent waiting at the start is gone if a problem turns up near the end.
Goldratt’s numbers deserve some caution. He claimed people estimate a task at around twice its median time. “There are no scientific data to support this claim,” the project management researcher Shlomo Globerson noted in a PMI article, though he added that it “makes sense to many project managers,” and that in his own seminar exercises, managers who wanted to play it safe more than doubled their estimates. The size of the effect is still Goldratt’s claim, not a measurement.
Moving the slack
Goldratt’s remedy, in the PMI review’s summary, was to set tighter estimates for each task and pool the safety taken out of them. Safety from the critical chain, the longest chain of dependent steps, goes into a project buffer at the end that protects the due date. Safety from tasks off that chain goes into “feeding buffers” that protect the critical chain where those tasks join it. Managers then watch the buffers. When more is used up than an agreed threshold, that triggers action.
The aim is to stop each task holding a private margin it can wait on. Pooling doesn’t stop anyone procrastinating by itself, but it is meant to keep the spare time for real trouble, and it lets everyone see how much is left. HMRC’s version is cruder but similar in spirit: a progress report in early January, a count of the millions still to file, and a reminder that a return can be started now and finished later.
The small agency
Imagine a small web agency that promises a client a site in four weeks. Each of the four people on the job pads their own piece by a few days. The copywriter, feeling safe, spends the first week on another job. In week three a photo shoot falls through, and there’s no slack left anywhere. The agency might instead cut each piece to its likely time, put a shared week at the end, and check on Mondays how much of that week has been used. The same week of safety might then cover the photo shoot.
An agent that rushed
In one Anthropic experiment, an AI agent had the opposite problem. When Anthropic set Claude to build a web app across many sessions, it found the agent “tended to try to do too much at once,” often running out of context partway through and leaving a feature “half-implemented and undocumented” for the next session. Later in a project, a fresh session “would look around, see that progress had been made, and declare the job done.”
The fix overlaps with Goldratt’s: make the real finish line visible. A first agent wrote a feature list, more than 200 items for a clone of claude.ai. Each later session worked on one feature at a time and marked it as passing only after careful testing.
The student starts too late. In that experiment, the agent stopped too early. Both lose track of how far the work still is from done. HMRC’s next window opened on April 6, 2026. It closes at 11:59 p.m. on January 31, 2027.