The Theory of Constraints in AI: Why Speeding Up the Wrong Step Can Slow You Down

Artificial Intelligence is often sold as a silver bullet for productivity: faster decisions, quicker processes, instant insights. But there’s a catch. The Theory of Constraints (TOC) reminds us that improving a single step in a process doesn’t necessarily improve the process as a whole. In fact, if that step isn’t the bottleneck, you might actually make things worse.

Here’s how it works: every process has a constraint, a step that governs the maximum throughput of the entire system. Think of it like the narrowest section of a funnel. No matter how much you widen other sections, the flow is still dictated by that pinch point. If AI makes one non-bottleneck task lightning fast, you’re simply pushing more work into an already overloaded constraint. The result? The bottleneck gets more fragile, queues get longer, and overall throughput can degrade.

Why This Matters for AI Adoption

In the rush to adopt AI, organisations often target tasks that are easy to automate rather than tasks that are genuinely constraining throughput. A chatbot that handles customer queries faster is great, unless the underlying process for resolving complex issues is still slow. All you’ve done is moved customers from one queue to another.

This is one of three "productivity leakage" components that I talk about in my latest e-book: the lost value between a task being faster and that time being meaningfully redeployed. If the step after your AI-enhanced task is still jammed, those efficiency gains either evaporate or pile up as work-in-progress, which itself has a cost.

The Physics of Bottlenecks

Regardless of what industry you look at be it manufacturing, logistics, software delivery etc. they all consistently shows the same TOC patterns:

• Throughput is limited by the slowest step. Efforts to speed up other steps may just increase idle time or work-in-progress.

• Piling work onto a bottleneck increases lead time. This is explained by Little’s Law, a key idea from queueing theory. In simple terms, the amount of work sitting in a system at any given moment is determined by how quickly new work arrives and how long each piece takes to complete. If you feed more work into a process without shortening the time it takes to finish each item, the total time to get things done will grow. In other words, as work-in-progress builds up, the time needed to finish all of it stretches out.

• Bottlenecks shift. When you successfully ease or remove one constraint in a process, another step will inevitably become the slowest point, taking over as the new bottleneck. This happens because the improved step now pushes more work downstream, exposing other limitations that were previously hidden. In AI-enabled processes, this shift can happen very quickly, sometimes within days, so it’s essential to track performance continuously, reassess the process flow, and be ready to redirect optimisation efforts to the new constraint.

In agile software teams, for example, automating code deployment only helps if QA and testing can keep pace. If QA is the constraint, deployment automation can simply push more features into a longer testing queue.

Finding the Real Bottleneck

TOC’s first rule is to identify the true constraint before investing in optimisation. In AI projects, that means:

• Mapping the end-to-end process to see where work slows or accumulates.

• Using data (e.g., time-to-complete, queue lengths) rather than assumptions.

• Talking to the people who live in the process daily; they often know exactly where things jam.

Once the bottleneck is found, AI should be applied there first. This maximises impact because any improvement in the constraint improves the throughput of the whole system.

AI, Bottlenecks, and Fragility

There’s another subtle effect that’s important to understand: AI can unintentionally make bottlenecks more fragile. When you increase the amount of work flowing into a constrained step without also increasing its capacity, you intensify the pressure on that part of the process. In people-driven workflows, this added strain can result in fatigue, burnout, mistakes, and even higher staff turnover, all of which reduce the step’s efficiency further. In system-driven workflows, such pressure can trigger outages, instability, slower performance, or degraded quality. In both cases, the constraint becomes weaker, not stronger, and the overall process suffers.

How to Apply TOC to AI Initiatives

1. Identify the Constraint. Use process mapping, metrics, and stakeholder interviews.

2. Exploit the Constraint. Ensure it’s always working on the highest-value tasks. AI can help with triage and prioritisation here.

3. Subordinate Everything Else. Align other steps to the pace of the constraint, avoiding overproduction.

4. Elevate the Constraint. Apply AI to increase its capacity or reduce its load.

5. Repeat. Once the constraint moves, start again.

A Quick Example

Imagine an AI tool that drafts legal contracts in minutes instead of hours. Sounds great, but if your legal review team can still only handle five contracts a day, you’ve just created a bigger backlog on their desk. This is a classic case where the actual constraint is not in drafting but in review. If you accelerate drafting without improving review capacity, you overload the constraint and lengthen the total process time. The smarter approach would be to apply AI to the review process itself, reducing the time per contract and lifting the true bottleneck. Once that happens, the constraint might shift to another part of the process, such as client approval or final sign-off, and you would then need to repeat the analysis and reapply AI or process improvements there too.

The Payoff of Targeted AI

When AI is aimed at the true constraint, benefits compound: shorter cycle times, less work-in-progress, improved customer satisfaction, and better use of human capital. When it’s aimed elsewhere, at best you get localised improvements; at worst, you slow the whole system.

The lesson is clear: before you unleash AI to make things faster, make sure you’re making the right thing faster. Because in a world of constraints, speed in the wrong place is just another form of waste.

Using Process Mining in ServiceNow to Identify Constraints

In my view, Process Mining in ServiceNow is one of the most underrated tools for making AI effective. It ties directly into everything discussed above, because finding and focusing on the true constraint is at the heart of the Theory of Constraints. By analysing event logs from across your workflows, Process Mining shows you how work actually moves through the system in real life, not just how it’s meant to on paper. This clarity makes it much easier to pinpoint the steps where delays build up, rework happens, or queues grow, all classic signs of a bottleneck. When you pair these insights with TOC, you can aim your AI investment exactly where it will improve overall throughput, then measure the results and adapt as the constraint inevitably moves elsewhere.