Glossary

The Agent Cash Conversion Cycle (ACCC) is the time between when an AI agent performs work and when that work is recognized in the enterprise's financial systems. In most organizations the cycle is the length of a billing month or longer, because spend doesn't surface until cloud invoices arrive. Settlement and Integration, which is Layer 6 of the Six-Layer Framework, compresses the cycle by writing each Autonomous Job into ERP and billing systems as it happens.

Agent Debt is the economic obligation generated by autonomous agent decision-making that existing financial controls can't see. It compounds silently and drives the Iceberg Effect, where visible model costs mask the deeper downstream spend. The concept is analogous to Technical Debt because it builds quietly, compounds over time, and becomes painful when it surfaces, but Agent Debt accumulates in real time and comes due the moment the next cloud bill arrives. In the Agent Economy, the diagnostic is simple: if production agents call paid APIs, run off-hours, and trigger expenses across multiple budgets but only show up in monthly bills, the organization is carrying significant Agent Debt.

Agent Drift is logic that evolves or degrades over time, changing an agent's cost profile without warning. A small parameter change, a model upgrade, or an updated prompt can shift behavior and spend significantly without anyone noticing until the bill arrives, contributing to Stochastic Spend at the workflow level. Agent Drift is distinct from Model Drift in that it covers the agent's overall behavior and economics, not just the underlying model's output quality.

AI Bankruptcy is the outcome facing organizations that scale autonomous agents without economic governance: compounding Agent Debt, abandoned projects, and governance retrofits that cost more than building controls right the first time. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 for the exact reasons economic controls are designed to prevent, namely escalating costs, unclear business value, and inadequate risk controls.

An AI Economic Control System (ECS) is an architectural layer purpose-built to govern autonomous agent spending, implemented in production through the Six-Layer Control Framework. It sits between AI models and the applications they serve, moving governance into the execution path itself so that spend is authorized or denied before it occurs. Where observability tools report what agents spent after the fact and FinOps tools reconcile it weeks later, an ECS evaluates expected cost against expected value at machine speed on every agent transaction, before resources are committed.

AI FinOps is used interchangeably with FinOps for AI. It is the application of financial management practice to AI workloads, including real-time usage and spend tracking, accurate cost allocation, resource optimization, and business-value measurement. AI FinOps uses KPIs specific to AI (such as cost per inference, cost per API call, token cost efficiency, GPU utilization, and AI ROI) rather than the cloud-era metrics built for predictable infrastructure. Agentic FinOps and Reactive vs. Pre-Execution Optimization are two frontier concepts within AI FinOps that push the discipline toward autonomous agent workloads.

AI Observability is the practice of observing AI applications to understand how they work and to diagnose issues when they arise. It extends the traditional observability pillars (logs, metrics, traces) with AI-specific signals like token usage, Model Drift, response quality, agent reasoning paths, and tool call chains. Where traditional APM was built for deterministic software that fails predictably, AI Observability tracks probabilistic outputs that fail in unpredictable ways. Economic Observability builds on top by adding cost, value, and business outcomes as first-class signals.

AI ROI is return on investment for AI projects, calculated as (business value generated minus AI spend) divided by AI spend, multiplied by 100. Accurate AI ROI depends on Cost Attribution to pinpoint which projects or workflows generated the spend, and on Cost-Per-Outcome to translate spend into business results. Without those foundations, projects risk becoming Zombie Projects: positive reviews, busy dashboards, and no evidence the spend is justified.

Anomaly Detection Rate is the percentage of unexpected cost spikes and misconfigurations a monitoring system detects, measured against the total that actually occurred. It is a coverage metric: detecting 18 of 20 monthly anomalies is a 90% rate. Anomaly Detection Rate is paired with Mean Time to Detect (MTTD) to measure how thoroughly and how quickly an organization catches the cost events that matter.

An Autonomous Job is a discrete economic Transaction that captures cost, context, outcome, and attribution as one auditable unit. It functions as the native unit of measurement for autonomous AI work, analogous to the dollar in ERP, the employee in HRIS, or the deal in CRM, and is captured by Settlement and Integration into enterprise financial systems. Each Autonomous Job binds the agent that executed the work, the business context it operated in, the resources it consumed, the outcome it produced, and the human and budget it's attributed to. A $5 billing inquiry and a $150 churn-prevention action are structured identically.

Batch Processing is a technique where multiple AI requests are grouped together and processed asynchronously rather than one at a time with immediate responses. It is used to maximize hardware (especially GPU) efficiency and reduce operational cost. Most major model providers offer dedicated Batch APIs at lower per-token rates than synchronous endpoints, in exchange for higher latency, making Batch Processing a strong complement to Provisioned Throughput Units (PTUs) for predictable, high-volume workloads.

Circuit Breakers are runtime controls that monitor spending velocity and cascade effects during agent execution and activate at machine speed when operations exceed predicted bounds. They are hierarchical by design: soft switches redirect to cheaper alternatives, hard switches disable expensive operations while maintaining core functionality, and emergency switches halt all activity when cascade failures threaten budgets. Circuit Breakers are the mechanism inside Financial Brakes (Layer 5 of the Six-Layer Framework) that catches Runaway Spend before the first alert fires Monday morning.

Composite Identity is the four-part identity binding an agent action must carry to execute: human budget owner, agent instance, business workflow, and customer context. It replaces anonymous spend on shared credentials with a traceable, attributable record of who authorized what and why. Composite Identity is the mechanism that makes Accountable Identity operationally enforceable in Actor Identity and Attribution, which is Layer 1 of the Six-Layer Control Framework.

Cost-Per-Outcome is the cost incurred to produce a single defined business outcome, such as a loan approval, a resolved support ticket, or a converted lead. It is calculated by tying AI spend directly to the outcome it produced rather than to activity proxies like token count. Cost-Per-Outcome is the measurement foundation for accurate AI ROI, and the metric that turns 'we spent $50,000 on AI last month' into 'we spent $3.78 per approval, at a 13,000% ROI on the workflow.'

Economic Intelligence is the pre-execution evaluation step inside an AI Economic Control System. It calculates expected cost against expected value before any resource commitment, determining whether an action should proceed, proceed at a lower cost tier, or escalate to human review. Economic Intelligence is the 'should this happen at this cost?' decision, made at machine speed, and it's implemented in Layer 4 of the Six-Layer Control Framework as the Cost Truth Engine.

Economic Observability is the discipline of treating cost, value, and business outcomes as first-class signals in the AI observability stack, alongside latency and model output accuracy. Where performance-focused AI Observability asks 'is the agent working?', Economic Observability asks 'is it worth it?'. It connects every AI action to its full cost, margin, and unit economics rather than treating cost as a separate, after-the-fact concern. Economic Observability sits between the model and infrastructure layer and the application layer, capturing the cost of every AI activity alongside the workflow and customer that triggered it.

FinOps for AI is the practice of applying financial management principles originally developed for cloud FinOps to AI workloads. Used interchangeably with AI FinOps, it tracks AI usage and spending in real time, allocates costs accurately to teams, projects, or products, optimizes resource allocation to prevent overspending, and measures the business value AI initiatives generate. Agentic FinOps extends the discipline further, tuning it for the Stochastic Spend patterns of autonomous agent workloads. FinOps for AI is distinct from traditional cloud FinOps because AI workloads scale with tokens (not hours), spike unpredictably, and depend on GPUs that cost an order of magnitude more than standard instances.

Human-in-the-Loop is a workflow design where human review or approval is required at one or more points in an otherwise automated AI process. In an Autonomous Job context, human review steps are tracked as cost events alongside model calls and tool invocations, so the labor cost of supervision shows up in the same ledger as the AI cost it complements. The mix between human and machine effort becomes measurable over time.

The J-Curve of Liability is the pattern in which legal risk from autonomous agent actions scales faster than the value those agents produce. As agents ingest more data, call more services, and make more decisions, regulatory exposure compounds (driven by data provenance failures, unlicensed content use, and rights violations) well ahead of the financial return. The Digital Turnstile, which is Layer 2 of the Six-Layer Framework, is designed to flatten the curve by blocking rights-violating ingestion before it happens.

A Job is a business unit of work performed by AI, whether handled by a single agent (such as resolving a support ticket) or by a coordinated multi-agent workflow (such as processing a loan application). In the Levels of AI Work hierarchy, Jobs group the underlying Transactions and Traces (including tool calls and human review steps) so the job's full cost and final outcome can be read together as one auditable record.

The List Price Illusion is the mistaken assumption that vendor list prices reflect what AI workflows actually cost. List prices ignore the Trust Tax (verification overhead, retry loops, and the compounding cost of multi-agent workflows), so the visible $0.02 LLM call routinely sits on top of dollars of downstream spend that never appears in the model bill. Total Cost of Interaction (TCI) is the metric that corrects for the illusion by normalizing vendor pricing to what the action actually costs in context.

LLM Observability and Agent Observability are two scopes within AI Observability that differ in what they capture. LLM Observability covers a relatively bounded interaction, where a prompt goes in, a response comes out, and key metrics (tokens, latency, model version, cost) can be captured from that single exchange. Agent Observability has to do more, tracing multi-step execution flows, external tool and API calls, retry loops, and decision paths. Agents make autonomous decisions that compound across many steps, so the wider the agent's autonomy, the wider the gap between the two.

Mean Time to Detect (MTTD) is the average time elapsed between when an anomaly occurs and when it's detected. In AI systems, where Runaway Spend can burn through thousands of dollars in hours, MTTD matters more than Anomaly Detection Rate alone, because the faster the catch, the smaller the blast radius.

Model Drift is the change in an AI model's behavior or accuracy over time as input data shifts or the model itself is updated. Originally a model-quality concern, Model Drift becomes an economic concern in agentic systems too, compounding into Agent Drift because shifts in output quality cascade into downstream tool calls, retries, and additional cost.

Model Routing is the practice of directing different requests to different AI models based on task complexity, cost, latency, or output requirements. It works hand-in-hand with Rightsizing (choosing the smallest sufficient model) and Semantic Caching (avoiding redundant inference altogether). Model Routing is one of the core levers for AI cost efficiency, and one of the easiest to get wrong: an overly aggressive routing rule degrades quality, while an overly cautious one wastes money.

Prompt Caching is the practice of storing and reusing common AI requests, or repeated prompt prefixes, so that previously processed content doesn't have to be reprocessed on every call. It sits alongside Semantic Caching as one of the two dominant caching strategies for AI cost optimization, and most major model providers offer native prompt caching that can substantially reduce both cost and latency for workloads with high prompt repetition.

Provisioned Throughput Units (PTUs) are committed-capacity pricing options offered by major AI providers, including Azure OpenAI PTUs and Amazon Bedrock Provisioned Throughput, where you reserve a fixed amount of inference capacity in exchange for lower per-unit pricing than on-demand. PTUs are analogous to reserved instances in traditional cloud FinOps, and pair well with Batch Processing for predictable, high-volume workloads. On-demand remains the right call for spiky or experimental ones.

Rate Limiting is the practice of controlling the number of requests or actions allowed within a set time window. In AI systems, it is typically applied as API calls per minute, tokens per minute, or inference calls per user, often using token bucket algorithms for burst traffic and returning HTTP 429 responses when limits are exceeded. Rate Limiting is a blunt but effective brake on cost and abuse, and complements Circuit Breakers by capping request volume rather than authorizing individual spend decisions.

Reactive vs. Pre-Execution Optimization is the distinction between optimizing AI costs after the fact and intervening before the spend happens. Reactive optimization analyzes AI costs after the fact: looking at last month's invoice, finding the inefficient workflow, and tuning prompts for next time. Pre-Execution Optimization intervenes before the spend happens, by denying, downgrading, or rerouting an agent action in real time based on projected cost versus value. Financial Brakes and the Cost Truth Engine are the runtime layers that make pre-execution optimization possible, catching Runaway Spend before it burns budget rather than after.

Rightsizing is the practice of selecting the appropriate AI model for the task at hand. Using overly complex or expensive models for simple tasks is wasteful, while underpowered models for complex tasks degrade outcomes. The working rule is to default to smaller, cheaper models and escalate only when the task's complexity justifies it. Rightsizing is closely tied to Model Routing, which operationalizes rightsizing decisions at request time, and is the single highest-leverage cost optimization for most AI workloads.

Runaway Spend describes recursive error loops, configuration cascades, and multi-agent feedback that burn budget geometrically. The canonical example is two agents stuck in an undetected conversation loop, burning through $47,000 in 11 days. Every individual action looks rational, but the cumulative cost is not. Financial Brakes (Layer 5 of the Six-Layer Framework) is designed to catch Runaway Spend at execution time, which is the difference between Reactive vs. Pre-Execution Optimization: catching the loop before it burns budget rather than after.

Semantic Caching is a technique that stores LLM responses based on the meaning (semantic similarity) of queries rather than exact string matches. It sits alongside Prompt Caching as one of the two dominant caching strategies for AI cost optimization. Semantic Caching uses vector databases to retrieve cached answers for semantically similar questions and bypass redundant inference: a model that has answered 'How do I reset my password?' can reuse the response for 'I forgot my login details, how can I access my account again?' because the underlying intent is the same.

Shadow AI describes agents running on shared keys with no accountability, generating anonymous spend on pooled credentials with no traceable owner. It is the agent equivalent of 'shadow IT', meaning real, business-impacting activity that doesn't appear in the systems supposed to govern it. Composite Identity, enforced in Actor Identity and Attribution (Layer 1 of the Six-Layer Framework), is the structural answer to Shadow AI.

The Iceberg Effect is the gap between an organization's LLM bill and what its AI product actually costs to run end-to-end. It's a symptom of the List Price Illusion: the model call is the visible tip, while downstream agent activity (tool calls, third-party API lookups, data enrichment, retry loops, and infrastructure events) is the mass below the waterline that determines actual margin. Correcting for it requires normalizing to Total Cost of Interaction (TCI) rather than trusting per-token list prices. That hidden mass lives in vendor systems disconnected from the workflow generating it, which is why teams routinely discover their real AI cost is seven to ten times the model bill only after it's too late to reprice.

Token-Based Billing is a pricing model where AI model providers charge based on token consumption, typically separately for input and output tokens, with rates that vary by model. It is the dominant pricing approach for foundation model APIs, and the reason raw token count became a default proxy for AI spend during The Unlimited Agents Era. Token-Based Billing is a proxy whose limits become obvious once agentic workflows, downstream tool calls, and outcome attribution enter the picture.

Tokenmaxxing is the practice of maximizing token consumption as a proxy for AI engagement and productivity, borrowed from internet slang where '-maxxing' means pushing something to its limit. It took hold in early 2026 as Meta, Amazon, and others ranked employees on internal leaderboards by token usage. Tokenmaxxing treats raw consumption as evidence of value, creating a measurable target engineers can game by running trivial tasks to inflate counts. Impact Density is the countermeasure, shifting the metric from tokens consumed to outcome produced. Tokenmaxxing is a textbook case of Goodhart's Law: when a measure becomes a target, it stops being a good measure.

Tokenomics is the emerging discipline of converting energy and capital into AI token production, efficient consumption of tokens to enable intelligence, and drive business value.

Total Cost of Interaction (TCI) is the normalized, context-aware price of an agent action that reflects what it actually costs to execute, not what the vendor lists. TCI is produced by the Cost Truth Engine and accounts for job type, trust score, resource requirements, verification overhead, and retry behavior. It's the answer to the List Price Illusion, and the metric that makes 'should this proceed at this cost?' a meaningful question.

A Trace is the end-to-end capture of a sequence of model calls, tool usage, retries, and decision paths that make up an agent's execution. In the Levels of AI Work hierarchy, Traces are the debugging view, while Transactions are the economic view and Jobs are the business unit. Full trace visibility lets teams diagnose where a workflow broke down or where money was wasted.

A Transaction is an individual AI call (a model call, tool invocation, or API hit) treated as a financial event that can be attributed, audited, and governed. Within the Levels of AI Work hierarchy, Transactions are the atomic unit of AI economics, distinct from a Trace (which is the debugging view) and rolling up into a Job (which is the business unit).

The Trust Tax is the hidden cost of verifying, retrying, and reconciling agent actions that vendor list prices ignore. Verification overhead, retry loops, and the cost of confirming an agent did what it claimed all add up, but none of them appear on the model bill. The Trust Tax is the structural driver of the List Price Illusion, and the largest source of the gap between quoted per-token pricing and Total Cost of Interaction (TCI).

Zombie Projects are AI projects that look successful on the surface (adoption climbing, dashboards busy) but on closer inspection don't deliver real business value. They consume budget while no one can point to the outcome they produce, and often coexist with Agent Sprawl in organizations without central governance. The fix is structural: tie every AI project to a measurable Cost-Per-Outcome before launch, and run a regular lifecycle review that removes projects unable to justify their spend.