Products & systems · 25 October 2026

The economics of AI-native product development.

When software becomes cheaper to produce, deciding what is worth producing—and how best to produce it—matters more than maximizing output.

Agents change the cost of making software. They do not eliminate the cost of choosing, verifying, integrating, or living with it. Two decisions govern the investment: how much the opportunity is worth, and which human–machine route can achieve an acceptable result at the lowest total expected cost.

CommitmentWhat is worth investing?

ExecutionWhat is the best route?

EvidenceWhat actually changed?

01 / Commitment

What is this worth?

Choose the appetite before execution. Hold the commitment fixed; let the scope respond.

An appetite is a value judgment, not an estimate of an unlimited solution. In AI-native work, agent execution, human direction and review, money, and elapsed time require separate limits. Better agent productivity increases what might fit inside a bet. It does not retroactively enlarge the bet.

Explore a fixed illustrative bet

2 people · 6 weeks480 human-hour reference · 2× production factor assumed at commitment · 240 agent-hour cap · 160 human hours and $20,000 separately authorized

Comparable scope480 h

Current agent production factor2.0×

Human review per unit of scope25%

The mandatory quality floor has been verified.

Comparable scope and production factors are illustrative, not universal conversions between people and agents.

Does the current plan fit?

Fits the bet

The estimated execution fits all three limits, and the quality floor is verified.

Machine execution240 / 240 agent-h

Cap set from the initial 2× assumption; use reflects the current factor

Human direction + review144 / 160 h

24 hours for direction + review time on comparable scope

Cash required$19,200 / $20,000

$20 per agent-hour and $100 per human-hour, both assumed

An agent-hour is one hour of aggregate agent execution, not one hour of calendar time or the value of a human hour. A deadline, permissions, and recovery requirements may impose additional constraints.

02 / Allocation

What is the least costly acceptable route?

The cheapest generation is not always the cheapest completed outcome.

A direct agent, a reviewed agent, a deterministic program, and a human specialist have different execution costs, latencies, and failure exposure. First rule out routes that cannot meet the task's safety and quality requirements. Then compare the remaining routes on total expected cost.

Change the task economics

Direct-agent failure probability14%

Loss if a failure escapes$2,000

Human verification cost$100

Risk reduction after review70%

Cost of a day of delay$20

This task can be expressed reliably as deterministic rules.

Compare eligible execution paths

Best eligible route

Reviewed agent

Route	Failure	Expected total	Decision
Agent, direct	14.0%	$306	Risk gate
Agent + review	4.2%	$221	Best fit
Human specialist	1.5%	$330	Eligible
Deterministic	0.8%	$168	Not applicable

Expected total = execution + review + delay cost + failure probability × loss. Routes must meet a 10% maximum modeled failure rate. Review is assumed to reduce the direct-agent failure rate; these probabilities are not measured. Rules-based systems are eligible only for tasks they can express.

Two decisions. One economic discipline.

The first fixes what you're willing to invest. The second asks how to use that investment well. Both remain subordinate to the same test: did we produce the intended behavior, and did it create value in use?

More efficient execution can make a bad idea cheaper. It cannot make it worth doing. What happens in the world—not the volume of work generated—should inform the next bet.

About the models. Every rate, cost, and probability shown here is illustrative. These are decision instruments, not measured productivity forecasts or an empirically validated system. This synthesis builds on Shape Up's appetite and Reinertsen's product-development economics. For the governing values, see The AI-Native Manifesto.