Every AI adoption dashboard I look at this week is optimizing the wrong thing. Eli Goldratt explained why in a 1984 novel about an industrial machinery plant: every productive system has exactly one constraint binding total throughput, and local improvements anywhere else just lengthen the queues. AI is the new faster machine. The novel still reads. Until your AI investment moves the real constraint, adding more generation upstream just grows the queue forming in front of whoever approves the work today.

The token KPI, the internal Claude usage ranking, the Cursor adoption panel: each one measures somewhere your team already has spare capacity. None of them measures where the flow actually jams.

The constraint moves; it does not vanish

Alex Wang published an article on May 15 that went viral this week with over 550 reactions. Wang authors the “Learn AI Together” newsletter with 548K subscribers. The core observation, verbatim: “AI increases output, but workflow capacity decides whether that output becomes progress.” When review, integration, and approval capacity does not move at the same speed, the bottleneck does not disappear; it shifts toward whoever now has to validate five times the material in the same hours. Wang names the symptom. The cure exists and is 42 years old.

Eliyahu Goldratt published The Goal in 1984 as an operating novel about Alex Rogo, a plant manager fighting to save a factory from imminent closure. The book sold ten million copies and is still assigned in MBA programs because it explained a simple mechanism most companies ignore when they buy new technology. Every productive system has a single constraint that binds total throughput. Everything else has spare capacity. If you optimize stations that already had spare capacity, you do not lift throughput; you only lengthen the queues forming in front of the station that is genuinely limited. Goldratt called that station the system’s Herbie, after the slowest scout in a Boy Scout troop on a hike: the entire troop walks at Herbie’s pace no matter how much faster the others could go.

In an enterprise AI rollout, the constraint is almost never code generation. It is one of four specific places. Pull request review. Integration testing. Decision authority to merge to production. Or rework, when AI-generated code passes shallow review but breaks something two sprints later. A mid-sized company where one senior engineer reviews everything on Friday afternoon has a Herbie with a name. Cursor or Claude Code adoption is not going to relieve that Herbie; it is going to bury him.

The Five Focusing Steps applied to AI

Goldratt prescribed a five-step process that remains, to this day, the only operating recipe inside Theory of Constraints with measured results behind it. Each step applies directly to an AI investment decision your team is about to make this quarter.

One. Identify the constraint. Do not trust the dashboard. Watch where the queue piles up in the standup. The phrase you have to hear is “blocked on review” or “blocked on approval.” The constraint is the person, process, or system that keeps showing up in that phrase week after week. If five different standups name the same person as the blocker, you do not need more analysis. You found Herbie.

Two. Exploit it. Before you add capacity, before you subcontract, before you buy more tooling, make sure the constraint is working only on what only it can do. The rule on a factory floor: one hour lost at the constraint is one hour lost across the whole system. AI behaves identically. Reduce the noise the senior reviewer has to process. Raise the quality of AI output before raising the volume. The tokens-per-shipped-feature KPI we covered two days ago is a step-two metric, not a step-one metric: it only matters once you already know the constraint sits in review, not in generation.

Three. Subordinate everything else. This is the step most companies skip. It means upstream generation paces itself to the constraint’s capacity, not the other way around. When you see an internal usage ranking like Meta’s reported Claudeonomics dashboard measuring tokens consumed per team and putting them in a public order, you are looking at the exact anti-pattern Goldratt warned about in 1984. Rewarding token consumption is rewarding upstream generation without verifying that the downstream constraint can process the queue. The number goes up, queues lengthen, throughput stays flat. Wang reported the Meta case this week alongside several parallels at other large companies; the observation matters more as a category than as an anecdote.

Four. Raise capacity at the constraint. Only after steps one through three are exhausted. Here you have two options, both expensive. Hire more senior capacity, a six-to-nine-month cycle in the current talent market. Or contract specialized review capacity by the hour. IQ Source’s Technology Partner service exists specifically for this step: we absorb the review and architecture load that does not need to live inside your senior team, while the senior team keeps the product and direction calls.

Five. Go back to step one. When you add capacity, the constraint does not disappear, it moves. Today it is the senior reviewer. In six months it may be the QA cycle, the production approver, the security team signing off every deploy. AI adoption is not a project; it is an identify-raise-identify cycle that repeats every quarter. The company that understands this pays for the next AI phase with margin. The one that does not pays for architecture it never ships.

Why this hits small and mid-sized companies harder

Most of the clients we work with at IQ Source are small or mid-sized companies with one to three senior engineers. The constraint is structurally identifiable because there is only one candidate: the single reviewer. At a large company there is always the option to hire another principal engineer; at a smaller one that move takes six to nine months in the current talent market and is the first commitment to drop when a soft quarter arrives. That is why step two, exploit the constraint, matters more at a mid-sized company than at a Big Tech company. A Big Tech company can absorb hidden waste. A mid-sized firm cannot.

Three concrete moves a mid-sized-company CTO can run this week, with no extra budget and no new hires.

One. Sit in on a standup with a notebook. Write down every time someone says “blocked on.” The phrase that appears most often points to your real constraint, not the one your intuition suggests.

Two. Take one queue in front of the identified constraint. Reduce its input before optimizing anything else. If the senior reviewer is jamming on poorly prepared pull requests, raise the self-review bar before raising the generation volume. One hour spent on a new PR template buys more than ten hours of Cursor configuration.

Three. Audit one AI adoption dashboard. If the metric counts upstream generation (tokens consumed, prompts sent, lines suggested), pull it from the executive report. Replace it with a constraint metric: review wait time, pull requests queued at end of Friday, rework rate from past defects.

What we do at IQ Source

AI Maestro is the two-month discovery program that operates steps one and two for you. The first deliverable is a Process Reality Map that names the constraint by name, not as a concept. The second is an AI Opportunity Score that orders the available bets by whether they raise capacity at the real constraint or merely optimize noise somewhere else. The Go/No-Go gate at the end of the program decides whether the AI implementation actually makes sense or whether the investment is better spent fixing another part of the flow first.

Technology Partner is step four when the constraint sits in senior review capacity and hiring is not a near-term option. We absorb the review and architecture load that does not need to live in your senior team, while the senior team keeps what only they can do: product decisions, technical direction, customer relationship.

The two services cover both sides of the problem Wang named this week. AI Maestro identifies where the jam sits before you buy one more tool. Technology Partner gives you real capacity to move the jam when the problem is structural.

If you want an hour of your team with me, no attached proposal, to identify your constraint in a mapping session: info@iqsource.ai. You bring the “blocked on” list from your last three standups. I bring The Goal and a notebook.

Goldratt closed the five-step process with an explicit warning: step five is not “finish and celebrate”; it is “go back to step one before inertia drags the previous constraint back in.” 42 years later that is still the hardest discipline. The right AI bet is the one that earns you the right to move the bottleneck somewhere harder. The wrong one just buys you a faster Herbie running the same jam in less time.

Map your bottleneck in 60 minutes