Notifications
Clear all
Topic starter
18/08/2026 2:59 pm
TL;DR: AI tokenomics in agentic systems must be measured from production telemetry, because one user request fans out into retrievals, retries, tool calls, and sub-agent runs that spreadsheet models systematically miss, according to Fiddler. The practical implication is that teams need span-level cost attribution to connect token spend to the agent, step, tenant, and outcome, not just the invoice.
NHIMG editorial — based on content published by Fiddler: How to Model AI Tokenomics for Agentic Systems in Production
Questions worth separating out
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.
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.
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.
Practitioner guidance
- 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.
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.
👉 Read Fiddler's analysis of AI tokenomics for agentic systems in production →
AI tokenomics in production: why spreadsheets break down?
Explore further
18/08/2026 3:39 pm
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.
A question worth separating out:
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.
👉 Read our full editorial: AI tokenomics needs span-level telemetry, not spreadsheet forecasts
