Find the Bottleneck First: What "The Goal" Can Teach Every AI Leader

In 1984, Eli Goldratt wrote a novel about a failing manufacturing plant. The manager, Alex Rogo, is given three months to turn it around or the plant gets scrapped. The breakthrough isn't a new machine or a better process. It's a mindset shift: stop optimizing everything and find the one constraint that's choking the entire system.

Forty-two years later, that lesson has never been more relevant — and almost nobody in AI is applying it.

S&P Global Market Intelligence found that 42% of U.S. companies have now abandoned most of their AI initiatives — up from 17% the prior year — with the average organization scrapping 46% of its proof-of-concept projects before they ever reach production. Gartner predicts 60% of AI projects will be abandoned through 2026 due to poor data readiness. And only one in five organizations investing in AI can show measurable ROI.

The pattern is remarkably consistent: the technology works in demos but fails in daily operations. The models aren't the problem. The system around the models is.

That's a Theory of Constraints problem.

The Five Focusing Steps — Applied to AI

Goldratt's Process of Ongoing Improvement is deceptively simple. Five steps. No exceptions.

Step 1: Identify the constraint. Find the single limiting factor preventing your system from generating value. In a factory, it's the slowest machine. In AI, it's whatever sits between your pilot and measurable business impact.

Step 2: Exploit the constraint. Before spending money to fix it, extract every ounce of capacity from what you already have. Don't buy a new machine — run the current one through lunch.

Step 3: Subordinate everything else. This is the hardest step. Slow down every other part of the system so work doesn't pile up in front of the bottleneck. In AI terms: stop launching new pilots if your data foundation can't support the ones you have.

Step 4: Elevate the constraint. Only after you've exploited and subordinated do you invest. Now you buy the machine. Now you hire the team. Now you upgrade the infrastructure.

Step 5: Repeat. The moment you break one constraint, the bottleneck moves. Go back to Step 1.

The reason this framework matters for AI is that most organizations are stuck permanently on Step 4 — throwing money at problems they haven't correctly identified. They're buying GPUs, licensing platforms, and hiring data scientists while the actual constraint sits untouched.

Where the Bottleneck Actually Lives

After working across multiple enterprise AI implementations, I've found the constraint almost never lives where leadership thinks it does. It's rarely the model. It's rarely the compute. It follows a predictable hierarchy:

First constraint: Executive alignment. Not "executive support" — that's easy to get. I mean genuine alignment on what AI is supposed to produce, measured in business terms. Deloitte's 2026 State of AI in the Enterprise found that while worker access to AI rose 50% in 2025, the gap between investment and measurable return keeps widening. Organizations spread budget across dozens of experiments without anchoring any of them to a specific business outcome. The money flows where the excitement is, not where the constraint is.

If leadership can't answer "what specific business outcome does this AI initiative change, and how will we measure it?" — you've found your bottleneck. Stop. Exploit that constraint before doing anything else.

Second constraint: Data readiness. This is the bottleneck Gartner keeps flagging: only 37% of organizations have confidence in their data management practices for AI. A Forrester survey found 73% of enterprise data leaders identified data quality and completeness as the primary barrier to AI success — above model accuracy, compute costs, and talent shortages.

In Goldratt's language, this is the machine on the factory floor that everything else depends on. If your data is fragmented, ungoverned, or undocumented, no model — no matter how capable — will produce reliable output. Subordinate everything to fixing this before scaling.

Third constraint: Workflow integration. RAND Corporation research identified the highest-impact failure cause across 65 enterprise AI practitioners: misaligned incentives and the absence of end-user co-design. The model solves the stated objective. Adoption is still zero. Nobody changed how they work.

This is Goldratt's "local optima" trap — optimizing a single workstation while the overall system throughput stays flat. An AI tool that saves an analyst 30 minutes but adds 45 minutes of verification and reformatting to the downstream team hasn't created value. It's moved the bottleneck.

What to Do First

If you're starting an AI initiative — or trying to rescue a stalled one — apply the five focusing steps in order:

1. Identify your actual constraint. Run a readiness assessment before writing a single line of code or licensing a single tool. Ask: where does value get stuck between "AI produces output" and "the business captures measurable impact"? That's your Herbie.

2. Exploit before you elevate. Before investing in new infrastructure, extract full value from what you have. Most organizations have existing data assets, BI tools, and automation capabilities running at a fraction of their potential. A Capital One/Forrester study found companies with superior data infrastructure registered five times the revenue growth and 89% higher profits — not from AI, but from leveraging the data foundation they'd already built.

3. Subordinate aggressively. Stop launching new AI pilots until your current ones can demonstrate value. S&P Global's data shows the abandonment rate nearly tripled in a single year — that's what happens when organizations keep launching pilots on top of unresolved constraints. If your bottleneck is engineering capacity, stop building and start buying. If your constraint is data, freeze new initiatives and fix the foundation.

4. Elevate with precision. When you do invest, invest in the constraint — not adjacent to it. If your bottleneck is data governance, the answer isn't a better model. It's metadata management, data quality tooling, and semantic layers. Gartner projects AI governance platform spending will hit $492M in 2026 for a reason — organizations are finally investing where the constraint actually lives.

5. Expect the bottleneck to move. Once data is solid, the constraint shifts to workflow integration. Once workflows are redesigned, it shifts to change management. Once adoption is high, it shifts to governance and compliance. This is the ongoing improvement Goldratt was after — not a one-time fix, but a discipline.

The Bottom Line

The organizations scrapping their AI initiatives aren't failing because AI doesn't work. They're failing because they're optimizing non-constraints — pouring money into model capability while the actual bottleneck sits in their data layer, their executive alignment, or their workflow integration.

Goldratt had a line for this: "Tell me how you measure me, and I will tell you how I will behave." If you measure AI success by the number of pilots launched, you'll get a lot of pilots. If you measure it by P&L impact delivered through the constraint, you'll get value.

The five focusing steps aren't a manufacturing relic. They're the missing operating system for enterprise AI. Find the bottleneck. Fix the bottleneck. Then — and only then — move on.

Patrick Phillips is CIO at Vasion and Adjunct Professor at Westminster College. He writes The AI Catalyst to help leaders navigate the real complexities of enterprise AI — not the hype cycle.