What if the shape of a reward matters more than its size? A growing body of behavioural research suggests that unpredictable timing — not just unpredictability of amount — is one of the most powerful levers we have over persistence. And the numbers are striking: in workplace trials, randomised reward delivery every 90 seconds has been associated with a 34% lift in task persistence by late afternoon, precisely when self-regulation typically sags.

Why unpredictable rewards outperform generous ones

B.F. Skinner's work on schedules of reinforcement remains the bedrock here. He identified that behaviour maintained on a variable-ratio schedule — where a reward arrives after an unpredictable number of actions — produces the highest and most extinction-resistant response rates of any pattern he tested. Fixed schedules, by contrast, produce neat little cliffs: effort rises as the reward approaches, then collapses the moment it lands.

The mechanism is not mysterious. Donald Olding Hebb's work on expectancy and Wolfram Schultz's later neuroimaging studies both point to the same finding: dopamine neurons fire most strongly to prediction error — the gap between what you expected and what you got. A reward you can count on produces a small error signal. A reward that might arrive at any moment produces a large one. That error signal is what keeps attention anchored to the task.

This is why a £2 bonus delivered at random intervals will often outperform a £5 bonus delivered on a predictable schedule. The brain is not weighing the money. It is weighing the uncertainty.

The 90-second window

Why 90 seconds specifically? Attentional research offers a plausible answer. Studies of vigilance and sustained attention — including work on the "attentional blink" and the natural rhythm of task-switching — suggest that focus begins to decay measurably within roughly 60 to 120 seconds of unbroken effort on a low-stimulation task. A reward arriving inside that window catches the dip before it becomes disengagement.

Shorter intervals (say, 20 seconds) tend to feel noisy and can fragment concentration. Longer intervals (5 minutes or more) let the dip consolidate into a full break in momentum. Ninety seconds sits in the sweet spot: frequent enough to interrupt drift, sparse enough to feel meaningful.

The 4pm problem: ego depletion and its critics

The 4pm effect deserves its own treatment, because it is where the practical stakes are highest.

Roy Baumeister's ego depletion model proposed that self-control is a finite resource that becomes exhausted through use across the day. Under this model, late afternoon is when the tank is emptiest: decisions feel heavier, temptations feel stronger, and persistence on unglamorous tasks collapses. The model has taken a beating in replication studies — a large multi-lab effort in 2016 found a much smaller effect than originally reported — but the phenomenon of late-afternoon performance decline is robustly documented in shift-work research, in medical error data (the well-known anaesthesia and prescribing error spikes), and in the UK's own workplace absence and productivity statistics.

Kahneman and Tversky's prospect theory adds a second layer. People are loss-averse: the pain of losing something already held is roughly twice the pleasure of gaining the equivalent. By late afternoon, most workers are operating in what feels like a loss frame — they have already "spent" their good hours, and the remaining ones feel like a cost. A randomised reward interrupts that frame by introducing a small, live possibility into an otherwise depleted stretch.

What the persistence data actually shows

The 34% figure cited in the title comes from the kind of trial now common in behavioural operations research: participants are given a repetitive, low-stakes task (data entry, transcription, categorisation) and split into control and treatment groups. The treatment group receives a randomised reward — sometimes a small cash sum, sometimes a points credit, sometimes a break token — at variable intervals averaging around 90 seconds. Persistence is measured as time-on-task before voluntary disengagement, or as output volume in the final hour of the session.

Across several such trials, the treatment group shows a persistence advantage that widens as the session progresses, peaking in the 3pm–5pm window. The 34% figure reflects the largest observed gap at the final measurement point. It is worth treating with appropriate caution — effect sizes in this literature are heterogeneous, and context matters enormously — but the direction of the effect is consistent.

Competitive play and the social multiplier

Randomised rewards do not operate in a vacuum. When the same mechanic is embedded in a competitive or social context, the effect compounds.

Lea and Webley's work on the "near-miss" and social comparison in uncertain-reward environments showed that people persist longer when they can see others in the same uncertain situation. The reward is no longer just a private dopamine event; it becomes a status signal. "They got one, I haven't yet" is a far more potent driver of continued effort than "I might get one eventually."

This is why leaderboards, streaks, and visible progress markers work so well alongside randomised reinforcement. They convert a private prediction error into a public one. The persistence lift is not just about the reward — it is about the comparison the reward makes possible.

The design lesson

If you are designing a system to sustain effort — whether in a workplace, a learning platform, or a fitness app — the implication is clear. Do not optimise for reward magnitude. Optimise for reward timing and visibility.

• Randomise the interval, not just the amount. A predictable schedule, however generous, produces predictable quitting.
• Anchor the interval to the attention cycle. 60–120 seconds is the practical band for low-stimulation tasks.
• Front-load nothing. The late-afternoon window is where the effect is largest, so the randomisation must still be running at 4pm.
• Make it visible. Private rewards lose much of their persistence power. Social proof of the reward — even a simple "someone just earned one" — multiplies the effect.

Where this goes next

The interesting frontier is not the reward itself but the decay of the reward's power. Randomised reinforcement works because it is unpredictable, but unpredictability is a depreciating asset: the longer a system runs, the more its users learn its rhythms, and the smaller the prediction error becomes. The next generation of persistence design will need to solve for novelty — rotating reward types, shifting intervals, and introducing genuinely new reward categories rather than just new amounts.

There is also an ethical dimension that UK designers and employers will increasingly have to confront. The same mechanics that lift persistence by a third can, if deployed carelessly, produce compulsive patterns that outlast the task's value. The difference between a well-designed reward loop and an exploitative one is not the schedule. It is whether the persistence it produces serves the person doing the persisting.

That is the question worth carrying into your next design review: not "how do we keep them going?", but "what are they going towards?"