The promise and the catch

Kettle Co. makes four ranges of kettle and sells them to hotels, offices and shops. Enquiries arrive by email and through a form on its website, forty or fifty in a good week, and two sales representatives deal with them. A representative reads the enquiry, looks up the company, decides whether it merits a quotation, prices it from a spreadsheet, applies a discount by feel, sends a document and intends to follow it up. In a quiet week the more promising enquiries are answered the same day. In a busy one, an enquiry that would have been worth pursuing can wait for over a week, by which time the buyer has often gone elsewhere. The commercial director, Nadia, has been aware of this for some time. What has changed recently is that the reasoning involved in the process, the part that once required a representative's attention, has ceased to be scarce.

The promise

That, put plainly, is the promise. A language model can read an enquiry, establish who the company is, form a view on whether it is a suitable customer and draft a competent covering note, within a minute, at any hour, and for every enquiry rather than for the ones a representative happens to reach. A process that has always rationed reasoning need not ration it any longer. The consequence is not a marginal improvement in productivity but a sales desk of a different kind, one in which every enquiry receives a considered reply on the day it arrives.

The catch

The difficulty is that most attempts at this go wrong, and they tend to go wrong in the same ways. In some cases a model is placed where a rule belonged, so that a machine which is never precise is asked to set prices, and the discounts drift. In others a demo is built that handles the handful of enquiries chosen for the demo, and meets its first real customer only once it is in production. In others again a pilot impresses for a quarter and is then quietly abandoned, because it was never connected to the systems in which the work actually takes place, and because nobody owned it or had budgeted for it. In each case the model attracts the blame, and in each case the model is the least of the problems.

Reasoning, and everything else

The distinction on which the rest of this course rests is between reasoning and everything else. Most of a business process does not involve reasoning at all; it consists of rules, of looking things up, of sending messages, of waiting, and of people deciding matters that must be decided by people. A model belongs in the narrow band of steps that genuinely require reasoning over untidy input, with a rule on either side of it. Placed there, it delivers the promise described above. Placed elsewhere, it becomes an expensive way of being wrong. Throughout this course two words are kept distinct. A model reasons: it reads and weighs what it is given and forms a view. A person exercises judgement: he or she decides, and owns the decision. It is reasoning that has become cheap. Judgement has not, and there is no reason to wish that it had.