Writing on the economics of production AI.

Artificial intelligence has become agentic work.

More of what organizations run is no longer a single question and a single answer. It is a multi-step agent that plans, calls tools, retries, and branches its way toward an outcome — and the meter never stops. When the meter never stops, the useful question is not how many tokens were spent but what a completed task actually cost, and how much of that cost was avoidable without degrading quality or service.

Agent economics is the study of exactly that: the full cost and value of a completed agentic task, where avoidable cost arises, and what share can be recovered. It is a young discipline. The bodies now convening to define it are only months old — when I attended AGNTCon in San Jose this October, the Linux Foundation's Tokenomics Foundation, barely three months old, was setting out its first findings. The vocabulary is being written now, in its first months, which is precisely why a Canadian contribution, offered openly and held to evidence, can help shape it.

This paper proposes a shared, testable way to measure it. Cost is read at four levels — Canopy (the outcome), Branch (the workflow), Trunk (the routing), and Root (the silicon) — so that the unit of account becomes cost per completed task rather than cost per token. Across those levels, agentic spend leaks through seven recurring mechanisms: model overkill, context bloat, retry storms, round-trip excess, cache misses, verbosity, and batch drag. Each is a hypothesis to test against a real workload, not an automatic verdict.

The number the paper pursues is residual overhead — the waste that remains after the best available mitigations have been applied, measured by running the same workload in two configurations with identical instrumentation. A residual figure is harder to establish than a headline one, and that is the point: it cannot be dismissed with "the stack already fixes that."

One discipline runs through all of it, and it is the same one that runs through everything we build: never confuse what is measured with what is modelled. Missing fields are recorded as "not reported," never as zero; no dollar figure is shown without verified rates; and no saving is promised before measurement. There is no synthetic data in this report.

Against Canada's AI for All push on sovereign compute and adoption, no published Canadian, cross-estate measurement of residual overhead yet exists. This paper names that gap plainly, sets out the research questions it implies, and invites Canadian operators — public sector, regulated industry, and enterprise — to help build the evidence as founding design partners: metadata-only, consent-based, aggregated, and published only under strict disclosure rules. The framework is offered openly for examination and use; the underlying measurement method is the subject of a United States provisional patent application.

The full paper carries the four-layer and seven-element tables, the residual-overhead method, the Canadian research programme, and the complete sources.

About Thompson Kinkade Foundry

Thompson Kinkade Foundry is a Canadian company building the measurement layer for the agent economy. Our philosophy is simple:

Measure before you optimize.

We believe organizations should understand the economics of their AI estate before making optimization decisions. Through evidence-based measurement and collaboration, we help enterprises and public institutions understand how intelligent work creates value — and where greater efficiency can be achieved.

Artificial intelligence is entering a new phase.

For more than a decade, progress in AI has been defined by capability. Researchers built increasingly powerful models. Governments invested in research. Technology companies competed to produce systems capable of reasoning, writing, translating and solving increasingly sophisticated problems.

That work transformed computing. It also changed the question organizations are now asking.

The challenge is no longer simply whether AI can perform useful work. It is whether organizations can operate AI economically, responsibly and at scale. As AI becomes embedded in everyday business operations, the conversation naturally shifts from intelligence to management.

Every technological revolution creates a new economic discipline. Railways transformed transportation but their greatest impact came from the economies they connected. Electricity transformed manufacturing after factories reorganized around electric power. Cloud computing became mainstream only after organizations developed new approaches to governance, financial management and operational oversight. Artificial intelligence is now reaching the same point.

Most people still think about AI as a conversation with a chatbot. Increasingly, organizations use AI to review contracts, investigate fraud, analyse documents, write software, coordinate workflows and support operational decisions. AI is becoming organized intelligent work.

Today's AI industry largely measures activity: tokens, requests, latency and inference. These remain essential engineering metrics but are increasingly insufficient management metrics. Executives invest in business outcomes, not tokens. Governments invest in public value, not inference. The economic unit of AI is shifting from the individual model interaction toward the completed outcome.

To describe this transition, this paper offers a name for the emerging management discipline of intelligent work: Agent Economics. Agent Economics studies how autonomous and AI-assisted systems create, consume and transform economic value. It asks how organizations should measure intelligent work, where operational waste emerges, how public investments should be evaluated, and what should be measured before optimization begins. Scholars have independently begun examining the economics of AI agents — convergence that suggests a coherent field is taking shape. This paper addresses its management dimension: not how agents behave as economic actors, but how organizations should govern the work they perform.

Canada has played an important role in advancing artificial intelligence research. The next opportunity is different. As AI becomes national infrastructure, Canada can help define the institutions, measurement practices and management disciplines that enable intelligent work to scale responsibly.

Every organization deploying AI will eventually confront the same questions: Do we understand the economics of our AI estate? Which intelligent workflows create measurable value? Where is AI creating waste? How should success be measured?

This paper does not attempt to answer every question. Its purpose is to argue that a new management discipline is being forged, to offer that discipline a name, and to invite governments, enterprises, researchers and practitioners to help build it together. Disciplines are not declared into existence. They are built.

About Thompson Kinkade Foundry

Thompson Kinkade Foundry is a Canadian company building the measurement layer for the agent economy. Our philosophy is simple:

Measure before you optimize.

We believe organizations should understand the economics of their AI estate before making optimization decisions. Through evidence-based measurement and collaboration, we help enterprises and public institutions understand how intelligent work creates value — and where greater efficiency can be achieved.