TL;DR — Key Takeaways
1. Tokens measure consumption, not business value. Enterprises should evaluate agentic AI using cost per accepted task, including retries, failed attempts, tool charges and human intervention.
2. Multi-agent workflows can multiply costs. Delegation, duplicate research, fallback models and unnecessary tool calls can make even inexpensive AI models costly when measured against completed business outcomes.
3. AI observability needs business context. Connecting usage data, financial records and human-review costs to accepted outcomes gives organizations a more meaningful foundation for evaluating efficiency and ROI.
Techstrong’s September 18 ROI commentary describes a monitoring tax: time spent reviewing, editing and fixing AI output. Its coverage of Splunk’s latest announcements shows observability expanding around agent behavior and token consumption. These developments point toward the same accounting question: can enterprises connect what their agents consume to what the business actually accepts?
A million tokens is a useful vendor billing unit. It is not a business outcome.
A customer refund or compliance review may require retrieval, several model calls, paid tools, retries and human approval. Add cooperating agents and the workflow can branch across vendors and sessions. The cheapest model call can sit inside the most expensive completed task.
The Unit of Account Should Be the Accepted Task
The useful denominator is a completed task that passes the organization’s acceptance criteria. The FinOps Foundation’s unit-economics framework already distinguishes resource-efficiency metrics from business-outcome metrics. Agentic AI makes that distinction operational: usage tells us where money went, not whether the intended work was finished.
For a defined set of comparable tasks, calculate cost per accepted task by dividing total attributable cost across all attempts by the number of unique accepted results. Include failed, abandoned and escalated attempts in the numerator. Count a business result once, not once per cooperating agent. If none passes, report no accepted output rather than a misleading unit cost.
Consider a hypothetical 1,000-task evaluation. One configuration costs $1,000 and produces 800 accepted results; another costs $1,200 and produces 960. Both cost $1.25 per accepted task. The second delivers more usable results, not cheaper ones. These are illustrative numbers, not product benchmarks.
Delegation Can Multiply the Bill
Anthropic’s 2025 engineering account describes early multi-agent research failures including excessive subagent creation and duplicate searches. Those observations concern its own development process, not a universal failure rate. They nevertheless illustrate an economic issue that does not require malicious behavior.
Imagine a procurement agent delegating supplier research to three specialists. Two purchase the same report; a third repeats a failed query; the lead agent pays for another model to reconcile the answers. Every component can appear reasonably priced while their combination produces avoidable expense. Parallel execution may reduce elapsed time while increasing total consumption.
An agent does not have to go rogue to make its workflow uneconomic. Cooperation itself needs a cost model.
Give Every Agentic Workflow a Cost Trace
Every production workflow should carry a business-task identifier, with separate attempt identifiers and recorded relationships between delegated operations. A new agent, session or provider should not sever that connection. The task record should link model usage, retrieval, paid tools, runtime, external charges and measured human-review time to the resulting acceptance decision.
OpenTelemetry supports spans and links between asynchronous operations, providing useful correlation machinery. It does not supply the accounting policy. Teams must still connect usage to prices, supplier receipts, allocations and accepted outcomes. Shared work needs a documented allocation rule; its full cost should not be charged independently to every downstream task.
Avoid double counting, too. A hosted model fee should not be added to an estimate of the same compute already included in that fee. Self-hosted infrastructure, shared services and implementation costs need explicit treatment. Publish marginal operating cost separately from fully allocated cost when both inform the decision.
Stripe and Tempo’s Machine Payments Protocol enables agents and services to coordinate payments programmatically. That makes transaction identifiers part of task accounting. A valid receipt proves expenditure, not business value; it must be linked to the work it supported.
User-facing completion and financial reconciliation should be separate states. A final answer can arrive while delegated jobs continue or supplier charges remain unsettled. Mark the cost provisional, track outstanding work and reconcile later invoices. Unassigned spending should remain visible rather than disappear from the efficiency calculation.
Observability Should Connect Cost to Completion
Splunk’s September release notes list agent observability alongside AI token and cost monitoring. Those are useful building blocks. A buyer should still ask how telemetry joins to task acceptance, invoices and human labor, rather than assume any observability product automatically supplies an end-to-end business ledger.
For a support agent, reducing the model bill is not enough to establish an improvement if unresolved tickets and human cleanup increase. A higher inference bill may be justified by better resolution quality or less rework. Compare configurations on the same workload, using acceptance criteria fixed before testing.
Alongside cost, record completion and acceptance rates, human-review minutes and elapsed time. Show the expensive tail as well as the average, and report what share of spending remains unallocated. Observability has its own cost; accounting need not retain confidential prompts to identify who performed each operation and what it consumed.
Flexible Compute Makes Task Economics More Useful
Google, NVIDIA and Emerald AI’s energy-management alliance focuses on flexible data centers. The application-level implication is conditional: where workloads permit, task deadlines can help schedulers defer work to cheaper or less constrained periods. But deferral is economical only after accounting for longer resource occupancy, retries and missed deadlines. Cheap compute that delivers too late is not a saving.
Benchmark the Workflow, Not the Invoice Line
Representative evaluations should include retrieval misses, duplicate subtasks, tool failures, fallback models, human escalation and delayed charges. These are not implementation noise. They are part of the product. Compare multi-agent orchestration with a simpler configuration on the same task set rather than treating more agents as progress by definition.
Cost per accepted task is a unit-economics measure, not ROI by itself. ROI still requires a credible baseline and evidence of realized benefit, such as lower expenditure, additional usable capacity or incremental margin. Minutes saved are not automatically cash savings. Integration, maintenance and rework costs cannot stay outside the investment case.
The agentic economy does not need to abandon token pricing. It needs a better business denominator above it.
The number that should eventually matter most is simple: what did it cost to complete a task that the business was willing to accept?
