In April 1956, a converted tanker called the Ideal X sailed from Newark to Houston carrying fifty-eight metal boxes. The shipping container is about as simple as an invention gets — a steel box with standard corner fittings — and it went on to cut the cost of moving goods by more than 90% and rewire world trade.
But not in 1956. For the first decade, the box changed very little, because everything around it was still built for loose cargo: ports designed for gangs of longshoremen, ships designed for sacks and barrels, paperwork designed for individual crates. Dropping a container into that system saved a little labour at the quayside and nothing anywhere else.
The 90% arrived only when the system was rebuilt around the box — purpose-built ships, cranes instead of muscle, standardised fittings so any crane could lift any container onto any truck, and entire ports relocated to deep water. The container was never the value. The redesign was the value. Economists found the same pattern with electricity: factories replaced their steam engine with one big electric motor and gained almost nothing for thirty years, until someone gave every machine its own small motor, reordered the floor around the flow of work, and productivity finally jumped.
This is worth keeping in mind, because most businesses are currently adopting AI the way ports handled containers in 1957.
Licences are boxes
The default corporate AI motion is to buy the box and change nothing: assistant licences for everyone, an AI feature-tier on the existing software, perhaps a chatbot in front of the knowledge base. Every step of every process stays exactly where it was, with a model now hovering politely beside it.
The result is the productivity equivalent of craning one container onto a ship built for barrels. People save minutes drafting emails while the process that consumes their week — quote, close, reconciliation, scheduling — runs exactly as it did in 2022. Which is precisely what the deployment statistics show: overwhelming adoption, negligible operational change.
The uncomfortable conclusion from both history lessons is that this stage cannot be purchased out of. No vendor sold ports a "containerisation platform" that spared them rebuilding the quay. The gains were locked behind redesigning the work, and they still are.
What redesign actually means
Redesigning a workflow around AI is more surgical than the word "transformation" suggests. It means taking one process and splitting every step into three piles:
- Deterministic work — moving data, comparing values, applying a threshold, filing the result. This becomes plain code. No model required, no model wanted.
- Judgment work — reading an unstructured document, classifying an odd case against past examples, drafting something a person will approve. This is where an agent earns its place.
- Decisions — anything that commits money, touches customers, or carries the company's name. This stays human, with the agent doing the gathering so the human decides in thirty seconds instead of fifteen minutes.
Most workflows, honestly decomposed, turn out to be largely pile one, which is why production agents that work are mostly ordinary code with a model in two or three carefully chosen places. The redesign is in the decomposition, not in the quantity of AI applied.
Choosing which workflow to redesign first matters as much as the redesign. The good candidates share four traits: they run often enough to matter (hundreds of times a month, or touching real money), they follow patterns a person could explain, they force people to gather context scattered across systems — inboxes, spreadsheets, the ERP, someone's memory — and their pain is measurable now, in hours, cycle time, or error rate, so you will know whether the redesign worked.
Redesign the work, not the systems
The instinct that kills more transformations than any other is conflating redesigning workflows with replacing software. They are opposites. Your systems of record took years and real money to bed in, and the fastest way to stall an AI programme is to couple it to a migration.
The container didn't require new cargo; it required new handling. Agents work the same way: they run on top of the systems you already have, reading and acting through the same interfaces your team uses, while the workflow around them is what changes. Two practical disciplines keep this maintainable. First, keep the business rules — thresholds, approval chains, special cases — outside the code, where an operations person can change them without an engineer. Second, keep the layers separate: system of record, rules, intake, and the corrections the agent learns from. A workflow built this way survives even when the software underneath it eventually changes.
The compounding part
Containerisation had a tipping point: once ports, ships, and fittings were standardised, every additional route got cheap, because the infrastructure was shared. Workflow redesign has the same property. The first redesigned workflow is the slowest, because it carries the platform — the integrations, the approval and logging machinery, the eval harness. The second lands in a fraction of the time on the same rails, and by the fourth or fifth, "should we automate this?" has stopped being a project and become configuration.
That is the realistic version of the AI-transformed company. Not a lights-out organisation run by a machine god — a business that has, one workflow at a time, moved the repetitive coordination work onto rails, the way freight moved into the box: deliberately, unglamorously, and then all at once.
The practical starting point is knowing which of your workflows is the Newark-to-Houston run — high volume, patterned, painful, measurable. That is a two-week mapping exercise, and it is exactly what our fixed-price audit produces: the decomposition, the candidate list, and the business case for the first redesign. Buy boxes if you like. The returns are in the rebuild.