The System Layer — a newsletter on systems thinking for technology leaders
The stopwatch was never the point
Most organizations adopting AI agents this year will produce more and deliver about the same. Not because the tools are weak, but because they will be pointed at the wrong part of the system.
We have seen this movie before. Frederick Taylor stood beside workers with a stopwatch, broke jobs into motions, and removed the wasted ones. It worked spectacularly, and a whole management philosophy grew out of it: optimize people, tasks and output, and outcome follows.
But it worked for a specific reason. On a 1910 shop floor the constraint was usually physical labor at the station being timed, and the work was visible. Speed up the worker and you sped up the system.
Even then it did not always hold. Eli Goldratt wrote The Goal in 1984 because factories that optimized every station were still missing their numbers. A faster machine upstream of a slow one does not make more product. It makes a bigger pile.
Modern organizations are almost entirely piles. We just cannot see them, because they live in inboxes, review queues, approval chains and Jira columns.
Time the work, not the worker
The modern time and motion study still uses a stopwatch. It just points it at a different thing: the work item’s journey from request to value, rather than the worker’s motions.
When you do that, a number appears that most leaders have never seen: flow efficiency. It is active time divided by total lead time. In knowledge work, figures of 5–15% are commonly reported. A feature that took six weeks to ship may have had four days of actual work in it. The rest was waiting.
Now do the arithmetic. At 10% flow efficiency, making every person twice as fast shortens lead time by about 5%. Halving the waiting shortens it by about 45%. The industrial instinct to optimize the worker attacks the smallest slice of the problem.
The second tool is older and simpler. Little’s Law says that in any stable system, lead time equals work in progress divided by throughput. Start fewer things and everything finishes sooner, at zero cost. Most organizations do the opposite, because a full board looks productive.
So the unit of study changes. Not “how long does this task take” but “where does this work item wait, for whom, and why.”
AI is an accelerant. It accelerates into the constraint.
Generative AI made production cheap: code, drafts, proposals, analyses, test cases. Almost nobody’s constraint was production.
The constraint is usually downstream. Human review. Approval. Integration. A customer’s decision. A specialist everyone needs. When you multiply output upstream of an unchanged constraint, you do not get more outcome. You get a longer queue, older work in progress, more context switching, and reviewers who start rubber-stamping to survive.
Google’s DORA research on software delivery reported, in its 2024 edition, that higher AI adoption was associated with slightly lower delivery throughput and stability. Faster typing did not mean faster delivery.
The fix is not to slow AI down. It is to change the order of operations. Goldratt’s five focusing steps translate cleanly:
- Identify the constraint from evidence: trace 20–30 finished work items through their timestamps and find the biggest, oldest queue.
- Exploit it: protect its time and make sure nothing defective reaches it.
- Subordinate everything else to its pace, including how much AI generates upstream.
- Elevate it: only now add agents, automation, people or authority, aimed at that step.
- Repeat, because the constraint moves the moment you widen it.
Notice that AI arrives at step four. Steps one to three are cheaper and make step four pay off.
Where agents actually belong
Put AI on the constraint’s inputs, its quality gates and its waiting time, not on the producers feeding it.
- Deterministic automation removes handoff waits between systems. No human should re-key data or forward a ticket.
- Copilots cut expert touch time at the constraint: pre-digested inputs, diffs, risk flags for the reviewer.
- Agentic workflows absorb wait states: triage, routing, chasing approvals, first-pass review against spec, overnight verification so a cross-timezone handoff stops costing a day.
- Decision rights encoded as policy widen the most common constraint of all, leadership latency. Below a threshold, an agent or a junior decides.
- Platforms and process mining make the constraint visible by timestamping every state change, continuously.
One warning: agents are a new source of work in progress. Give every agent workflow an owner, a WIP limit and a measured acceptance rate, like a team.
Retire the vanity dashboard
The industrial metrics measured activity. The system needs flow.
Stop reporting
Start reporting
Utilization, hours
Flow efficiency
Tickets, PRs, proposals sent
Throughput of accepted outcomes
Velocity, story points
Lead time, p50 and p85
AI adoption %, agents deployed
Change in constraint throughput; cost per outcome
Features shipped
% used at 30 days
AI-generated volume
Human review queue age
Six numbers about the system, none about individuals. That is how output finally starts to equal outcome.
Before you fund your next AI initiative, ask one question: which constraint does this widen, and where will we see it in end-to-end lead time within 60 days? If nobody can answer, you have found your first bottleneck. It is in the room.
What is the oldest item waiting in your system right now, and who is it waiting for?
Read the full issue on LinkedIn →
The System Layer publishes on LinkedIn. Subscribe here.