Redesigning the floor

A year after the first real run, Kettle Co.'s sales desk is not the old desk made faster. Every enquiry is answered on the day it arrives, including an order for twelve kettles from a guesthouse that no representative would previously have spent ten minutes on, and which now receives a quotation nonetheless. The website undertakes to quote within the hour, and does so at any time of day. The two representatives still work the desk, but the work has changed: they own the workflow, read its runs, take the discount decisions and the difficult cases, and devote the time they once spent on lookups to the customers who repay a conversation. Nadia's report to the board carries three figures and a line she had not anticipated: a segment of the market to which the company had never sold, because serving it had never been worthwhile.

Replacing the engine, redesigning the floor

When factories first acquired electricity, they removed the steam engine and installed an electric motor in its place, driving the same belts and shafts, and output scarcely changed. The gains came a generation later, when it was understood that a small motor could be attached to every machine and the floor laid out around the flow of work rather than around the engine. Kettle Co.'s first month amounted to replacing the engine: the same process, run faster. The guesthouse quotations and the undertaking to reply within the hour were the floor being redesigned, and they came only once the desk had run for long enough for someone to ask what else had become cheap.

The question that opens up the larger possibility is the one posed in chapter 1, asked now of the whole business rather than of a single desk. Almost every business is shaped by the fact that judgement is scarce and expensive. A bank assigns a personal adviser to its wealthiest customers and a form to everyone else. A law firm reviews the large contracts and waves the small ones through. An insurer prices policies in bands because to underwrite each individually would cost more than the premium. None of these is a process problem. Each is an offer, a price and a customer base determined by the cost of judgement, and each may be determined afresh once reasoning brings that cost down.

Industries are reshaped for the same reason. A business whose function is to stand between two parties and exercise judgement on their behalf now holds a weaker position than it did; a business that holds data and relationships but could never afford to act on them one customer at a time holds a stronger one. The most ambitious form of this is the company in which much of the work is conducted by agents dealing with other agents, its own with those of its suppliers and customers. Such a company is not the opposite of what this course describes; it is the same arrangement at scale, since a thousand interacting agents require identity, budgets, records and gates even more than one does. A reinvented business still runs on rules, models and people, and the discipline of the earlier chapters applies to it in full. What changes is the content of what is written down: a new offer, a new price or a new position in the industry, rather than a faster version of last year's process.

Who drives it

Improvement, in Kettle Co.'s case, came from below. Tom knew which steps of the sales desk required reasoning without being told. Left to itself, however, improvement from below produces scattered pilots that never reach the accounts, each with its own passwords and no record of what it did. Reinvention came from above, because no team can decide what the company sells, how it prices or which part of the market it enters; the decision to serve guesthouses was Nadia's, not the desk's. Left to itself, direction from above produces a strategy and a platform that nobody uses, because the strategy was never connected to a process that anyone runs. The arrangement that works is one in which the centre owns the platform, the connections, the rules of governance, the spending limits and the question of reinvention, while the teams own the workflows. The sign that the arrangement has failed is a central AI team writing workflows for processes it has never run.

Five ways it goes wrong

Kettle Co. avoided most of the following, in some cases narrowly. They are set out here as a summary of the course and as a reminder for the next process.

Automating stupidity. The first projects in most organisations use a model to cope with information that arrived in a worse form than that in which it was created: a vision model reading scanned invoices that were structured data before someone printed them, a model extracting orders from emails that a form would have captured, or a chatbot answering the same five questions because the product's screens or pricing are confusing. Each is a genuine improvement, and each is a patch over something that ought to be repaired. There is no harm in beginning there, as Kettle Co. nearly did with its email enquiries, provided one knows which kind of project one is undertaking and treats the patch as something to be removed in due course.

Pilot purgatory. A convincing demo can be assembled in an afternoon from a model, a prompt and a few chosen inputs. Production requires everything else: the checks, the limits, the connections to the systems in which the work takes place, the cases the demo never encountered, and a person who owns the result and has a budget for it. Most pilots stall in that gap. Kettle Co.'s first real run, described in chapter 3, was that gap being crossed deliberately.

Automating a broken process. Where a process was never written down, its exceptions and its unwritten rules were never visible, and a model asked to learn the process from examples will reproduce the confusion faithfully. Nadia's notebook, in chapter 2, was the remedy, and it brought to light two rules that the representatives had been applying differently for years.

Shadow AI. Where the sanctioned route is slow or absent, staff turn to consumer tools instead, often with customer data, and the organisation finds itself with AI in its processes and no record of what was done or what was disclosed. Prohibition rarely works; a sanctioned route that is at least as convenient usually does, and a workflow that answers an enquiry within a minute is more convenient than a browser tab.

The untracked bill. Model calls are cheap individually and expensive in volume, and a step without limits may spend for hours on a single unfortunate input. The cost of each task must be capped at every model step and read from a ledger from the first week, as chapter 5 describes, so that the bill is a figure someone has been watching rather than a surprise.

What this asks of leadership

Leading such a programme does not require an understanding of how the models work. It requires clarity about what the programme is for, since a programme that is for everything cannot be measured. It requires that the question nobody below the leadership can ask be asked: not how the company might do what it does more cheaply, but what it would do if reasoning were nearly free. And it requires that leadership intervene wherever a boundary is crossed, since a process that runs through sales, finance and support has three owners and therefore none, and the rules governing what an agent may do on the company's behalf are not for any single team to set.

Two habits then carry more weight than any dashboard. The first is to ask, every quarter and of every process that has been entrusted to a workflow, the three questions set out in chapter 5. The second is to open the record of a single run once a month and follow it step by step. It takes ten minutes, and the fact that it is done will be known throughout the organisation.