Unit Economics

What question does this solve?

The central economic question is whether AI could reduce the complete cost of serving an intention enough to make a recurring relationship of consuming advances viable. Cheaper code, on its own, does not answer it. Preparation, interpretation, specification, experimentation, coordination, execution, evaluation, rework and shared capacity all belong in the account, and so do the requests that never lead to construction.

Definition and role

Unit Economics relates attributed revenue or resources, consumption, and results per explicit unit of analysis and period, without assuming that a Delivery Unit is a unit of charge.

Define the unit before calculating. A Delivery Unit is a tracking unit. It can be a reference for analysing consumption, but it is not necessarily the unit of charge or a standard measure of work. An analysis can examine one delivery, a set of demands, a commercial commitment or a period of operation, chosen to match the question. If revenue is agreed for a set of deliveries, dividing it between units requires an explicit convention, and the attributed shares do not become independent transactions. Comparing average costs across units without regard to their nature, dependencies and acceptance can suggest productivity or margin the information does not support.

Provider and buyer perspectives

For the provider, the analysis considers what is received and what is consumed to meet commitments. For the buyer, it considers the resources used to obtain and use the result, including their own participation. The price paid is not the whole of the buyer’s cost: clarification, evaluation, integration and changes to their process can consume more. Internally there may be no commercial revenue; comparing consumption with budget is legitimate, but an internal attribution of value must not be presented as external revenue. Business benefit needs an analysis of its own: accepting software says the delivery was fulfilled, not that use produced an economic return.

How it works

Consumption along the route

Component	Examples of consumption
Preparation	Clarification, reading context, configuration and access.
Production	Human work, agent execution and automation.
Evaluation	Environments, checks, review and interpreting results.
Correction and new attempts	Diagnosis, re-execution and re-evaluation.
Coordination	Routing, follow-up, handling dependencies and decisions.
Continuity	Keeping context, resuming, and later obligations covered by the agreement.
Shared capacity	Resources kept available, tools and structures used by several demands.

The classification must not count the same consumption twice. The observation period matters too: an accepted delivery can carry later obligations covered by the commitment, and stopping the measurement before them gives an incomplete picture.

Attributable and shared costs

Some consumption can be related directly to a demand; some sustains several demands or the general availability of the service. Allocating shared costs needs an understandable criterion, and different criteria can change each unit’s apparent contribution without changing total consumption. A shared cost already incurred does not affect an immediate decision the way additional consumption does, but looking only at additional consumption does not show that the whole operation sustains itself over time.

A bounded numerical example

The example is fictional and uses monetary units; it represents no market prices, productivity or margins. Consider an accepted delivery with revenue of 100 and these consumptions attributed to the same scope and period.

Item	Illustrative value
Revenue from the delivery	100
Preparation and context	12
Production	25
Evaluation	15
Correction and re-evaluation	8
Coordination	10
Total directly attributed costs	70
Balance before shared costs	30
Share of shared costs, under the chosen criterion	18
Balance within the scope analysed	12

The balance of 30 is not the service’s total profit, and the balance of 12 is not a net accounting result either: the example does not try to cover every expense, tax, financing cost or other effect a real analysis might need. Its purpose is to show that what is included has to be declared. A claim of margin without a definition of revenue, cost and period can hide relevant differences.

Changing the execution: sensitivity

In the same fictional example, suppose an automated composition cuts production from 25 to 15 but raises evaluation from 15 to 20 and correction from 8 to 14. With the other components unchanged, the directly attributed cost moves from 70 to 71. The isolated reduction in production does not reduce the total in this scenario. It is not a forecast about agents; it shows how the conclusion depends on the components considered.

Now a favourable configuration: preparation 12, production 12, evaluation 10, correction 6 and coordination 8. The direct cost would be 48; with the same shared allocation of 18, the cost of the scope would be 66, and the balance against the same revenue of 100 would be 34. This only illustrates the mechanism; it is not a measurement of AI efficiency.

Suppose that, for the same scope and period, the highest commercially acceptable price were 75. The configuration in the first table, with a direct cost of 70 and a total of 88 including shared costs, would not sustain itself. The favourable configuration would cost 66, leaving a balance of 9. That crossover, from a configuration that could not sustain the offer to one that could, is the kind of change in viability the hypothesis investigates. It explains the opportunity for new bounded commitments; it does not come from charging in instalments, and fictional numbers do not show that it has happened in any period or market. Omitted costs, real capacity, quality and actual willingness to buy would still need checking, and the scenario where automation raises the direct cost to 71 has to stay in the analysis. A composition can also have a similar cost and meet another important constraint, such as time.

Capacity and utilization

Execution Pools can organize shared resources, but their existence does not show savings. Keeping capacity available can cost money with no work assigned; high utilization can coexist with long waits or no room for exceptions. Occupancy should be related to accepted deliveries and commitments met. Context, integrations, evaluation criteria and configurations can be kept to serve new demands, and the cost of preparing that capacity can be spread over a set, but maintenance and updates continue, and more requests can also require new resources.

The scaling hypothesis is that certain classes of demand could grow without a proportional increase in all human effort. To examine it, track the hours of preparation, execution, decision, review and recovery per accepted delivery, including the effort spent on unsuccessful attempts, read together with quality, scope, computational consumption and the work required of the buyer.

Rework, interruptions and deliveries not accepted

The consumption of attempts that never reached acceptance is part of the operation; excluding those demands gives a favourable picture that does not represent the service. An accepted exploration with a negative conclusion is different from failing to fulfil a committed delivery, though both consume resources. Not every consumption is charged separately, and a charge does not show the cost was economically recovered across the relationship. Rework can come from execution, ambiguous criteria, insufficient context or later changes; attributing it all to the executor can lead to choices that miss the real economic problem.

Revenue, receipt and obligations over time

The moment the service consumes resources can differ from the moment money is received, and a commitment that looks attractive may require resources before any payment. Receiving money in advance does not show obligations have been met. The model sets no accounting criteria; it requires that the scope of the analysis make clear what is measured and which obligations remain outside the consumption observed.

A set of routes, including those that never reach construction

This example is independent of the one above and of the barbershop case; its costs are its own. It covers a fictional period with twelve requests. Ten buy and complete discovery; six of those ten also buy and receive construction; four stop after an accepted discovery; two other requests consume an initial assessment but lead to no purchase. All prices, costs and quantities are illustrative.

Group	Calculation	Attributed revenue	Direct cost
Ten accepted discoveries	Price 20 and cost 14 each	200	140
Six accepted constructions	Price 80 and cost 58 each	480	348
Two requests with no purchase	Initial assessment cost of 10 each	0	20
Total directly attributed	All twelve requests	680	508

The costs of 14 and 58 already include their directly attributable attempts, corrections and coordination; they must not be added again. Shared preparation of 120, attributed entirely to the same period, raises the cost to 628, leaving a balance of 52 before other effects not modelled. Attributed revenue does not mean money received.

The four discoveries without construction contribute revenue of 80 and direct cost of 56, inside the totals. They are not four failed deliveries: those discovery commitments were fulfilled. The two requests with no purchase also have a recognized cost. Counting only the six constructions would hide part of the operation. The set contains ten discoveries and six constructions of different sizes and purposes; dividing all costs by the six constructions may be an analytical convention, but that quotient cannot be compared with the direct cost of one construction. The denominator must be declared.

As a sensitivity, if only three of the ten discoveries went on to construction under the same conditions, revenue would be 440 and cost 454: 140 for discovery, 174 for construction, 20 for the requests without a purchase and 120 shared. The balance would be minus 14. The service can deliver useful discoveries and still depend on a different volume or composition to sustain itself. The answer is to examine offer, price, costs and capacity, never to push customers into continuing when it does not suit them in order to hide that dependence.

What would show a change in viability

The economics of the hypothesis have to be examined at three scales: one commitment, a set of routes, and the operation over time. A commitment can make a positive contribution while the service fails to cover the capacity it keeps; an operation can cut costs while the buyer spends more time clarifying and checking the work.

Comparing a competent conventional operation, an operation with AI assistance and an organization with delegated cycles helps locate where the change happens, as long as scope and quality are comparable, differences of context are recorded, and preparation, evaluation, attempts, decisions, relevant waiting and later obligations are included. A reorganization with better processes can reduce costs without AI; attributing a shift in the boundary to the technology requires separating its effect from reorganization, accumulated experience, commercial subsidy and the selection of demands. When a design only compares complete packages, the result supports the package without isolating its cause.

Signals to observe include the cost of resolving uncertainty, total human effort per accepted delivery, computational cost, time to a usable offer, decision latency and buyer effort. They are not interchangeable: less waiting can be valuable without lower cost, and lower cost without quality is not the same delivery. The hypothesis gains support when a composition makes a class of service repeatedly feasible and sustainable and buyers choose it under the conditions offered. Gains that disappear once review, unconverted demands or maintenance are included weaken the economic explanation.

Connected concepts

• Economics overview
• Previous: Pricing
• Next: Outcomes
• Execution Pools, shared capacity and its costs
• Glossary