AI tokenomics needs span-level telemetry, not spreadsheet forecasts

At a glance

What this is: This is an analysis of why production AI tokenomics fails when teams model agentic systems as single-call workloads, and why span-level telemetry is required to attribute spend correctly.

Why it matters: For IAM, NHI, and AI governance teams, the key lesson is that agentic cost control depends on observability across each runtime step, which is also where identity, access, and tool-use controls become enforceable.

👉 Read Fiddler's analysis of AI tokenomics for agentic systems in production

Context

AI tokenomics is the problem of measuring how tokens are generated, consumed, priced, and converted into value across AI systems. In agentic environments, the governance gap is not just cost forecasting, but attribution across the full run, because one request can trigger multiple calls, retries, and tool invocations before any visible answer is returned.

That creates an identity-adjacent control problem as well as a financial one. When agents, sub-agents, and tools are allowed to fan out without per-span telemetry, teams lose the ability to tie spend to a specific agent, workflow, tenant, or outcome. In practice, that makes it harder to govern both cost and the behaviour of the AI system itself.

Key questions

Q: How should organisations control AI costs in agentic environments?

A: Organisations should control AI costs by combining metering, attribution, and enforcement across the full request path. That means tracking model calls, tool usage, and data movement, then applying caps or routing rules when consumption exceeds policy. Without identity-linked attribution, cost control remains reactive and finance cannot trust the numbers.

Q: Why do agentic systems break traditional token cost models?

A: Traditional models assume one request equals one model call, but agents fan out into retrievals, retries, tool invocations, and sub-agent runs. That makes token consumption non-linear and difficult to predict with spreadsheets. The result is hidden cost growth unless telemetry is tied to the full execution path.

Q: How do you know if AI token optimisation is actually working?

A: You know it is working when lower token spend does not reduce the business outcome. Track value per token by pairing cost with a result signal such as a resolved case, completed workflow, or accepted code change. If spend falls but outcomes hold, the optimisation is real.

Q: What should teams do when AI cost attribution is unclear?

A: They should reconstruct the run from telemetry before changing policy. Join gateway logs, agent traces, model pricing, and outcome data so you can see which span, workflow, or tenant created the overage. Without that attribution, cost controls will be too blunt to fix the actual issue.

Technical breakdown

Why spreadsheet tokenomics breaks for agentic systems

A spreadsheet model assumes a mostly linear relationship between requests and cost. Agentic systems are non-linear because a single user request can expand into retrieval, reasoning, tool execution, fallback calls, and sub-agent orchestration. That means the relevant unit of measurement is not the prompt, but the span inside the agentic hierarchy. If teams only model visible input and output tokens, they miss the hidden work that drives the real bill and the real operating profile.

Practical implication: Practitioners should move cost modelling from request-based estimates to span-level traces across the full agent workflow.

Why token value is not the same as token spend

Token spend is the provider bill, while token value is the business outcome produced by that spend. Two runs can consume the same number of tokens and produce very different utility, which is why cost control alone cannot tell you whether an agent is economically justified. The article’s core point is that AI tokenomics only becomes useful when spend is paired with an outcome signal such as a resolved ticket, completed workflow, or merged pull request.

Practical implication: Teams should measure value per token, not just tokens per run, before deciding which workflows to scale.

How production telemetry enables accurate attribution

Production telemetry lets teams capture input, output, and reasoning tokens at the span level, then resolve price at request time rather than later in reporting. That matters because model routing, pricing, and caching can change between billing cycles, silently rewriting history if prices are applied after the fact. Proper instrumentation also has to stitch gateway-side data to agent-side data, otherwise sub-agent loops and tool calls disappear from the accounting model.

Practical implication: Instrument spans with agent, tenant, model, price, and outcome so cost can be reconstructed after the run.

NHI Mgmt Group analysis

AI tokenomics is becoming an identity and governance problem, not just a finance problem. In agentic systems, the same telemetry that attributes cost is also the telemetry that reveals who or what initiated a tool call, how far a delegated task expanded, and where control boundaries were crossed. That makes the model relevant to NHI governance because agent runtimes behave like non-human actors whose actions need traceability, scope, and accountability. Practitioner conclusion: if you cannot attribute a run, you cannot govern it.

Span-level observability is the named concept this market now needs. The article shows why request-level accounting fails once fan-out, retries, and sub-agent loops enter production. Span-level observability gives teams a measurable boundary for cost, behaviour, and control enforcement inside the agentic hierarchy. Practitioner conclusion: treat span granularity as a governance requirement, not an engineering luxury.

Reasoning tokens expose a new blind spot in AI governance. They are billed but often invisible in the response, which means output-only reporting systematically understates real consumption. In governance terms, the system is executing work that is economically material yet operationally obscured. Practitioner conclusion: budget and monitor hidden reasoning cost separately from visible response tokens.

Attribution must be designed into the AI operating model from the start. The article’s production-telemetry approach aligns with broader governance disciplines that insist on evidence, traceability, and outcome linkage. For identity teams, that same principle applies to delegated access: the runtime record must show not only what happened, but which agent, tenant, and step were responsible. Practitioner conclusion: tie AI observability to accountability controls before scale makes the gaps expensive.

What this signals

AI teams should expect cost governance to converge with runtime governance. As agentic systems scale, the organisations that can trace spend to a span, a tenant, and an outcome will also be the ones able to enforce meaningful control boundaries across delegated actions.

Agentic observability debt: the longer teams wait to instrument span-level telemetry, the more expensive it becomes to explain cost overruns, audit behaviour, or prove that optimisation did not damage outcomes. This is now a programme design issue, not a reporting issue.

For practitioners

• Implement span-level cost telemetry Capture input, output, and reasoning tokens for every span, then join those records to the agent, tenant, model, and outcome so cost can be reconstructed after execution.
• Resolve price at request time Store the effective unit price when the run executes, not after the billing cycle closes, so pricing changes and routing shifts do not rewrite historical spend.
• Attribute cost to workflow outcomes Pair run cost with an outcome measure such as ticket resolution, task completion, or pull request merge so teams can calculate value per token rather than only spend per token.
• Stitch gateway and agent telemetry Combine gateway logs with agent-side traces to capture sub-agent loops, retries, and tool calls that do not pass through the same metering point.

Key takeaways

• Agentic AI breaks single-call cost assumptions because one request can expand into many hidden spans.
• Production telemetry is the only reliable way to attribute token spend to agents, steps, tenants, and outcomes.
• Teams need to optimise for value per token, not just lower invoices, or they risk cutting useful work instead of waste.

Key terms

• AI Tokenomics: AI tokenomics is the study of how tokens are generated, consumed, priced, allocated, and converted into value across AI systems. In practice it is a measurement discipline that links model usage to both economic cost and workflow outcome, especially when agentic systems create multi-step consumption patterns.
• Span-Level Telemetry: Span-level telemetry records the smallest meaningful unit of work inside an AI run, such as one model call or tool invocation. It lets teams trace cost, retries, and behaviour back to the exact step that created them instead of relying on aggregate totals.
• Value Per Token: Value per token measures the business outcome produced for each unit of token spend. It is the most useful way to judge whether an AI workflow is economically justified, because it distinguishes expensive but productive runs from expensive runs that add little operational value.
• Agentic Hierarchy: An agentic hierarchy is the layered structure of AI execution, from individual spans to traces, sessions, and application views. It matters because cost, risk, and policy decisions often happen at the span level, while governance decisions are made higher up the stack.

What's in the full article

Fiddler's full blog covers the operational detail this post intentionally leaves for the source:

• The span-level JSON schema for capturing token counts, resolved pricing, and outcome metadata across a run.
• The pseudo-code for rolling up span costs into per-run spend using different token classes.
• The specific dashboard failure modes that undercount hidden reasoning tokens, retries, cache effects, and gateway-only metering.
• The practical sequence for turning value-per-token analysis into a repeatable optimisation workflow.

👉 The full Fiddler post includes the telemetry model, cost roll-up logic, and optimisation workflow in detail.

Deepen your knowledge

NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, machine identity security, and agentic AI identity. It helps practitioners connect runtime observability to access control, accountability, and lifecycle governance across modern identity programmes.