Writing · 5 min read
The factory moved into the inbox
Why building agentic AI is an industrial engineering problem, and what a hike in The Goal has to do with it.
Tiaan van Antwerpen ·
People hear “engineer” and picture machines. Industrial engineering is mostly about everything around the machines.
A company is a system of people and processes that have to cooperate, and the job is to get the most value out of that system. You map how work flows through it, you measure where it slows, and you change the system rather than one part of it.
Industrial engineers have been measuring factories for over a century: how long an order takes, how many units leave a shift, what each one costs, how many come back. Those numbers are the discipline. Without them you are guessing.
Herbie’s pack
The book that made this click for me is Eliyahu Goldratt’s The Goal. It is a novel about a plant manager trying to save his factory, and its clearest lesson does not happen in the factory. It happens on a hike.
Alex, the manager, takes his son’s scout troop on an overnight walk. The line keeps stretching out. However fast the boys at the front go, the troop only arrives when the slowest boy, Herbie, arrives, and pushing the fast ones harder only widens the gaps. So Alex moves Herbie to the front, opens his pack, and shares the heavy things out among the others. The whole troop speeds up.
That is the theory of constraints. Every system has one step that sets the pace for all of it. Improve anything else and nothing changes. The method is to find that step, protect it, and only then spend money on more capacity.
“An hour lost at a bottleneck is an hour lost for the total system.”Eliyahu M. Goldratt, The Goal
The new factory floor
Knowledge work is a production line too. It is harder to see, because the inventory sits in inboxes and queues instead of on the floor. A claim arrives, someone triages it, an expert reviews it, a decision goes out. A fault code comes in, a technician searches the manuals, a repair gets scheduled. Each is a line of stations, with work waiting in front of one of them.
An agent is a new machine on that line. That framing is what I bring to the work. The first question is not which model to use. It is where the work waits, and whether the agent is pointed at that step or at one that was never slowing anything down.
What I measure
The measures barely change between a factory and an agentic workflow. What changes is what you count.
| Measure | On a factory floor | In an agentic workflow |
|---|---|---|
| Lead time | Order received to order shipped. | Request received to request resolved, including the time it sits in a queue. |
| Throughput | Units off the line per shift. | Cases closed per day, not messages sent or tokens used. |
| Cost per unit | Labour, machine hours, scrap. | People, model calls, infrastructure, and the rework when it is wrong. |
| Quality | Defect rate, returns. | Accuracy, escalations, corrections a person has to make. |
| Utilisation | How often the machine sits idle. | How often anyone acts on the output. An answer nobody trusts is idle capacity. |
Where most projects fall short
Most AI projects skip the baseline. Nobody measured the line before the agent arrived, so nobody can say what it changed. The demo looks good, the pilot gets extended, and a year later the honest answer to “did it help?” is a shrug.
My master’s thesis is a small example of measuring instead. I built a framework that routes each query in a text-to-SQL agent to a suitable model, and evaluated it across thirteen of them. The RAG router landed on the cost and quality Pareto front: it beat the most expensive model, Claude Opus 4, on combined quality at 56 times lower cost, $0.31 for the hold-out set against $17.40.
- 56×
- cheaper than Claude Opus 4, with a higher combined score
- $0.31
- a batch, against $17.40
- 13
- models evaluated
The less exciting finding mattered just as much. On accuracy alone, none of the routing beat simply picking a good model. You only learn that if you measure. The full write-up is in Adaptive LLM routing for effective agentic workflows.
Five steps, read for agents
Goldratt’s answer to a constraint is five focusing steps. They carry over almost word for word.
- Identify. Find where the work actually waits. Map the line and count the queues; it is rarely where people say it is.
- Exploit. Get the most out of that step as it is. Often a better handoff or a cleaner input beats any model.
- Subordinate. Make every other step serve the constraint, so the expert never waits for a case that is not ready.
- Elevate. Only now add capacity: the agent, the model, the automation, pointed squarely at that step.
- Repeat. The constraint moves somewhere else. Measure the line again and start over.
Building AI, for me, is industrial engineering on a new floor. The machines changed. The job did not: find the constraint, relieve it, and prove you moved it.