The AI Bottleneck Lens

The Bottleneck Lens, two panes. Pane 1: horizontal pipe with five stages, one narrow constraint (ember-tinted) with queue waiting in front; AI deployment markers at each stage labelled inventory, idle capacity, throughput gain. Pane 2: same pipe after constraint lifted, a different section now narrowest — Goldratt’s the-constraint-moves warning visualised.

​

One. What does this organisation produce, and how is throughput measured? Not revenue. Throughput is the rate at which the organisation converts inputs into the thing the customer pays for. For a law firm, matters delivered. For consulting, engagements closed. For a hospital, patient episodes completed. For software, features shipped. The unit varies; the question does not.

​

Two. Where does a single unit of work wait the longest on its way through? Pick one recent piece of work and walk it backwards. Mark the wait states. The constraint is almost never the step that takes the most effort; it is the step with the longest queue in front of it. Wait states almost always dwarf work states.

​

Three. Would lifting this step lift the system, or shift the wait elsewhere? Goldratt’s warning: when you lift the constraint, the constraint moves. The leader’s job is to know where it moves to next, and whether the new constraint is one the organisation can address or one that is structural and will swallow the gains.

​

Three questions, mechanical by design. They do not require an AI strategy. They require the leader to know the actual shape of the work, and then to apply AI at the step the answers point to.

​

Genesis

Eliyahu Goldratt (1984), in The Goal, set out the Theory of Constraints: every system is governed by one constraint at a time, and improvements anywhere else produce more inventory or more idle capacity, not throughput. Michael Hammer (1990), in Harvard Business Review (July–August 1990, ‘Reengineering Work: Don’t Automate, Obliterate’), added the insight that organisational bottlenecks are more often handoffs than tasks, and that the wait states between steps are where most organisational time is lost. The Bottleneck Lens applies both to the AI question: AI deployed at the wait states between steps, rather than spread across the steps themselves, is what lifts throughput.

​

Why it matters now

Four forces pull leaders to AI dashboards and away from constraint analysis. Adoption metrics are easy to measure, vendors instrument them by default. Vendor incentives reward licence count, not your throughput. Adoption is collective and reports as a single number, while constraint analysis is local and looks unstrategic. Loss aversion makes leaders prefer the low-risk programme that touches everyone over the higher-risk programme that touches the team that matters most. The result is green dashboards, unchanged throughput, eighteen months of effort applied everywhere except the place that controls the flow.

​

The moves

Trace one unit of work. Take one recent piece of organisational work and walk it from instruction to delivery, marking wait states with timestamps. The exercise takes a morning and surfaces the constraint by inspection, not by argument.

​

Define the throughput metric and report it monthly. Not weekly, the noise drowns the signal at weekly cadence. The metric is whatever unit answers Question One for this organisation; the cadence is whatever lets the trend become visible above the noise.

​

Apply AI at the constraint, rather than at the average. When the constraint is named, the deployment becomes specific. The licence count may fall, the throughput rises, and the dashboard the board sees changes from a licence number to a throughput trend.

Run the three moves once and the leader has a constraint dashboard; run them quarterly and the constraint dashboard tracks the constraint as it moves.

​

How ChangeSchool applies it with executives

We run senior cohorts through the Constraint Trace: each leader walks one recent piece of organisational work backwards in front of a peer, names the constraint, and commits to one throughput metric for the month ahead.

​

The discipline

A monthly throughput review with the constraint metric as the first agenda item. A quarterly bottleneck audit, because the constraint moves when you address it; the audit re-traces a unit of work to find where it has moved to. A standing constraint conversation with the board, which retrains the board over three or four meetings to ask the right question by default.

​

​

​

‘AI’s organisational impact is decided not by how many people use it, but by whether it lifts the one step that controls the flow.’ 

​

​

​

Viren Lall, Managing Director,

ChangeSchool LDN (2026).

virenlall.com/ai-bottleneck-lens

​

​

​