The real cost structure of agentic workflows: hidden bills from tokens, tools, operations, and human review
An insurance company's claims review agent cost $0.11 per execution in proof of concept but $0.94 in production, a ninefold difference. The additional expense falls outside token consumption: tool calls, failed retries, operations overhead, and human review. A complete cost model requires accounting for five components, not tokens alone.
By
Tenten AI 研究團隊
應用 AI
Published
February 4, 2026
Read time
6 分鐘
An insurance company's claims review agent showed $0.11 per execution during proof of concept. The team included this estimate in an executive summary. Three months of production data revealed an actual cost of $0.94 per execution, nearly a ninefold difference.
The difference reflects costs that fall outside the token line item. Most PoCs never account for them.
Agentic workflow costs are never just a token bill
The full cost of a single agent execution comprises model inference, tool calls, iterative retries, operational infrastructure, and human review, not just the token consumption that API reports show. PoCs underestimate agentic workflow costs because they measure only the first component, which in a live system typically represents the smallest portion of the total.
Agents differ from single prompts. A chat completion takes input and produces output in a single exchange. An agent plans, calls tools, reads results, and determines next steps through multiple rounds, typically five to fifteen LLM calls per task. Each round reprocesses the full accumulated context, and token usage grows with the square of the step count, not linearly. This nonlinear growth is the first factor most cost estimates overlook.
Breaking down a single execution's cost
The following cost model applies to deployment and reconciliation across client engagements. The percentages reflect a weighted average from multiple live production systems. While the proportions vary by industry, the structure remains consistent: tokens almost never constitute the majority of total cost.
| Cost Component | % of Per-Execution Cost | Included in PoC? | Why It Gets Missed |
|---|---|---|---|
| Model inference tokens | 12-20% | Usually | Only counts one turn, doesn't account for accumulated context growth across multi-step execution |
| Tool and external API calls | 15-25% | Seldom | Search, vector retrieval, sandboxed execution, third-party lookups all have separate billing |
| Iterative and failed retries | 10-18% | Rarely | Agents loop and retry; failed tasks still burn full token and compute costs |
| Operational infrastructure | 15-20% | Rarely | Vector database, tracing, logging, version control, monitoring amortization |
| Human review and exception handling | 25-40% | Never | High-risk outputs require human approval before release |
The final row reveals what undermines business cases: not model pricing but human review and operations. These line items never appear in demonstrations.
Human review is the biggest invisible bill
An agent that automatically handles 80% of cases offers appeal. The remaining 20% involves outputs affecting money, compliance, or customer standing and requires human approval. The insurance company's agent achieved a 78% auto-pass rate. Each case returned for human review consumed an average of six minutes from a senior claims examiner, who had to trace the agent's reasoning chain, verify sources, and decide whether to accept or reject the result.
Spreading those six minutes of human labor across each execution reveals the cost gap between $0.11 and $0.94. Human review costs do not decline with scale. Multiplying volume tenfold requires hiring proportionally more reviewers or accepting degraded review quality. This contradicts the software engineering principle of writing code once and deploying indefinitely, and it causes failures in most deployed systems.
Operations and retries: costs that never show up in a demo
Operations costs carry a particular challenge because they remain invisible in daily operations. Vector databases charge fees, and each knowledge base update requires reindexing. Observability tools bill per event, and a multi-step agent execution generates dozens of traces. Prompts and tool definitions need version control to track changes. None of these costs appear in PoC spreadsheets, yet they become a fixed monthly charge after launch.
Retries represent an underestimated variable cost. Agents loop, tool calls fail and restart, and malformed outputs trigger re-execution. The key point: a task that ultimately fails still consumes all tokens, tool calls, and compute. Observed failure rates reach 12% in some systems, meaning one complete execution's cost is lost for every successful result. ROI calculations using only successful cases as the denominator produce dangerously optimistic results.
Keeping cost numbers honest
A defensible cost model must use all initiated executions as the denominator, including failures and cases sent back for review. The numerator must include all five components, with human review and operations amortized monthly rather than treated as per-execution margin. A validation check: take one month of production billing and divide by the tasks the system completed without human override. This figure typically runs five to ten times higher than the PoC estimate. Learning this gap early proves far more valuable than discovering it during budget review in Q2.
During a full deployment evaluation, the analysis extends beyond the initial demonstration. Engineers work on-site to connect all five cost categories to actual billing, run a complete cycle, and then determine which steps should reduce context requirements, which exceptions warrant human review, and which tasks should not be assigned to an agent at all. Demonstration success does not constitute proof of viability. Sustained production use of accurate cost models does.
One stuck workflow
is enough to begin
Tell us what the team does today, where it breaks down, and what a better working day should look like.