Attention has limited capacity: we cannot process multiple streams of information equally well simultaneously. The bottleneck occurs at different processing stages depending on task demands and practice level. Understanding capacity limitations explains why multitasking degrades performance and why automaticity develops for well-practiced tasks.
From your study of divided attention, you know that performance degrades when people try to do two things at once — but the interesting question is why, and why it degrades more for some task combinations than others. The concept of an attentional bottleneck provides the answer: at some stage of information processing, the system has limited capacity, and two tasks that compete for that stage interfere with each other in proportion to their shared demands.
Early cognitive theorists proposed the bottleneck as a filter that operates at the perceptual stage — only one stream of sensory information can be processed for meaning at a time (Broadbent's early selection model). Later evidence showed this was too early; Treisman's attenuation model and Deutsch & Deutsch's late selection model pushed the bottleneck further along, suggesting that more processing occurs in parallel than early models assumed. The consensus view that emerged — and that fits most of the empirical data — is that the bottleneck location is not fixed. For early sensory discrimination tasks, parallel processing is possible and interference is minimal. For tasks requiring conscious response selection, memory encoding, or executive decision-making, competition for the same processing resource creates a genuine queue — one task must wait while the other is processed. This is the psychological refractory period: when two tasks overlap in time and both require central processing, the second task's response is delayed as if waiting for a single-lane bridge to clear.
Capacity models — particularly Kahneman's resource model — offer a complementary framework. Rather than a single-channel bottleneck, these models posit a pool of general-purpose attentional resources that tasks draw from. Tasks that compete for the same resource pool interfere with each other; tasks that draw from different resources may coexist without interference. For example, a verbal task and a spatial task can often be performed together more successfully than two verbal tasks, suggesting partially separate resource pools. The dual-task cost — performance degradation relative to single-task baseline — measures how much resource overlap exists between two tasks.
Automaticity is what changes with practice. Well-practiced tasks make fewer demands on the central capacity-limited stages, potentially shifting from controlled processing (which requires attentional resources and is slow, serial, and effortful) to automatic processing (which runs in parallel, requires few resources, and is fast but inflexible). The classic demonstration is Spelke, Hirst, and Neisser's study of subjects trained for months to read while taking dictation — initially impossible, eventually nearly seamless. The implication is that the fixed "multitasking ceiling" is not entirely fixed: it is partly a function of skill level. But automaticity has a tradeoff: automatic processes are less amenable to modification, which is why it is so hard to change deeply ingrained habits even when you know they are wrong.