The first two parts showed what the market can offer, and what users want and can bear. AI agents change something else: how much a person must put in personally to accomplish a task. The yardstick is the amplification factor: the ratio of the effort the principal would need to do the task alone to the effort the principal must still put in after handing it to an agent. The latter has three parts: the portion the agent cannot do and the principal must still do; the effort of specifying the task and checking the result; and the rework after the agent makes mistakes. In finance one more constraint applies: markets are close to efficient, so more accurate prediction by an agent brings limited benefit. Part Three takes this decomposition apart term by term.
The three terms in the denominator differ in nature. The portion not applicable depends on the task itself and sets the upper limit of the amplification factor. Specifying and checking depends on whether the result can be verified easily. Rework depends on how reliable the agent is. How fast the agent generates output does not appear in the denominator.
Whether a task can be handed to an agent depends on three questions, in order. First, is it worth doing: once handed over, how much effort does the principal save over a year, and does it make possible something that could not be done before? Second, can it be done: can the agent get the information and tools it needs, can it act on site, and has it obtained legal and institutional access? Third, does one dare hand it over: how severe are the consequences of an error, and can the result be judged right or wrong in good time? If an earlier question fails, the later ones need not be asked; the third question determines how much of the nominal value from the first can actually be realized.
The three questions map one to one onto the decomposition above. Whether it is worth doing concerns the numerator and the overall factor; whether it can be done determines the size of the portion not applicable; whether one dares hand it over determines how much effort goes into specifying and checking, and how much rework follows errors.
Part Three deals only with the nature of the agent itself: what it can do, how reliably it does it, and how its work is checked afterwards. Each chapter again starts from a concrete situation, cites published experiments and evaluations, and gives the date of each number: AI capabilities change quickly, and the scope of each conclusion is defined by its mechanism.
Once the amplification factor is taken apart, the conclusion fits in one sentence: what an agent saves people depends on how large a scope it can take on completely, how much rework its errors require, and how long checking takes, not on how fast it generates output; in finance, any advantage in prediction is used up by competition, the lasting value lies in cost, discipline, behaviour and accessibility, and errors in execution must be stopped by authority and limits.
These conclusions form the third set of constraints on designing financial products. With the three sets of constraints from markets, users and AI side by side, the final question can be answered: under these constraints, what should a financial product promise, how should it be agreed, how should it be implemented, and how should it earn trust?