The arithmetic of replacement, and its weak point

For almost a decade businesses were sold support automation, and almost always it turned out to be a decision tree: a set of prepared branches a person is walked along until they give up and ask for an operator.

Such a system could not be assessed by payback, because it replaced almost nothing. It shifted irritation from the operator to the customer, and the call happened anyway. There was nothing there to compute a saving from.

The calculation takes a support department at roughly $4,000 a month per person, giving $32,000 a month for the team and around $420,000 a year including associated costs. The figures are illustrative, but the structure is typical: salary plus management plus turnover plus training replacements.

It is precisely those last two lines that usually drop out of such sums. The cost of hiring and training someone who leaves within a year dissolves into the general budget and is never charged to support. In this calculation it is included, which makes it more honest rather than more flattering.

Here appears the number that made the article worth reading: about $0.30 per dialogue. That is not an abstract subscription price but the cost of computation multiplied by the length of the conversation.

Everything unfolds from there. At that order of price, serving the same flow of enquiries costs roughly $7,200 a year, and against that the gap with $420,000 looks less like an improvement than a change of unit.

The stated share of enquiries closed without a human is 98.2 per cent. That is the key coefficient of the whole model and worth holding separately in mind: it is what turns a calculation into a result.

The first assumption is the stability of compute prices. Thirty cents per dialogue holds today and for particular models. A change of model, a rise in load, or a change in a provider's rates moves that figure in either direction, and the whole calculation moves with it.

The second and more important is the 98.2 per cent. The difference between 98 and 90 per cent is not «eight per cent» but a fivefold rise in the number of enquiries that reach a human. On that footing the live department has to be kept, and the saving stops being a saving.

The third is the silent assumption that the replaced work is uniform. First-line support genuinely is repetitive. But inside it there are almost always a few per cent of cases where a human is needed not because the bot failed to understand, but because the customer is angry and wants someone alive to hear them out.

Saving is defence. The article fairly notes that attack is more interesting: the same agent can carry a conversation through to a sale, whereas a live operator loses the deal because they went home or did not answer in time.

Counting that is harder, and the article admits as much. A saving can be read off a payroll; additional revenue can only be measured against what did not happen. Such figures need a different kind of check: not a calculation but a comparison of two periods with everything else unchanged.

Not the number but the method. The useful part of the calculation is the habit of breaking the cost of support into components, including the ones usually hidden: turnover, training, enquiries missed overnight.

And any model of this kind is best tested with one question: what share of enquiries actually closes without a human — for you, on your flow. Every other number is derived from that one, and an error in it inverts the whole result.