Because routing cost is only one part of the workflow. If the selected downstream model is poor for the task, the organisation can save money on the decision step and still lose more through retries, bad outputs, or slower completion later in the chain.
Why the cheapest routing step is not the whole risk calculation
Routing is only a decision point, not the full workload. If a lower-cost router sends a request to a model that is weak for the task, you may save on selection cost and still pay more later through retries, bad completions, manual correction, or slower end-to-end processing. The overall risk depends on total workflow outcome, not the routing fee alone.
That is why routing should be evaluated as part of a chain, not as an isolated optimisation. A cheap decision layer can look efficient while increasing downstream waste, especially when the chosen model has poor task fit, weak instruction following, or inconsistent output quality.
When the task is operationally sensitive, the meaningful comparison is between total cost to achieve an acceptable result and the likelihood that the routed model will need rework. In practice, cheaper routing can simply shift cost from the front of the workflow to the back of it.
What actually drives the hidden cost after routing
The main cost drivers are usually repetition, correction, and delay. A model that produces incomplete, low-confidence, or structurally wrong output can trigger another pass, a fallback model, or human review. Those extra steps often cost more than the original routing savings, particularly at scale or in high-volume agent workflows.
Quality losses also compound. If the routed model is slightly wrong but not obviously wrong, the organisation may not catch the problem immediately. The result can be more expensive than a clean failure because the workflow appears to have succeeded while quietly degrading the final outcome.
In agentic systems, this matters because the selected model may influence tool use, action planning, or exception handling. A poor routing choice can therefore increase not just token spend, but also the chance of mis-execution, wasted tool calls, and slower resolution of the original task.
How to judge whether cheap routing is really cheaper
The right question is whether the routing policy reduces total task cost per successful outcome. That means measuring the full path: routing decision cost, downstream model cost, retry rate, escalation rate, completion latency, and quality of the final result. If those downstream measures worsen, the apparent saving is not a saving at all.
Useful routing policies usually optimise for task fit, not just model price. For some tasks, a slightly more capable model or a more conservative router is cheaper overall because it avoids repeated attempts and manual intervention. For other tasks, low-cost routing is fine because output quality is easy to verify and low stakes.
So the practical test is simple: if you remove the routing discount but the workflow becomes materially more reliable, the cheaper option may have been a false economy. If the cheaper route holds the same success rate and the same completion quality, then the saving is real.
Risk and Threat Considerations
Cheaper routing can hide risk when cost optimisation is separated from quality control. The organisation may believe it has reduced spend, while the real exposure is shifted into retries, degraded decisions, or delayed completion that is harder to notice and harder to attribute.
Failure mechanism: The router selects a lower-cost model whose output quality is insufficient for the task, which increases rework, fallback usage, or human correction and erodes the original savings.
Impact: Total cost rises, service quality becomes inconsistent, and time-sensitive workflows can miss their operational target even though the routing line item looks efficient.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST CSF 2.0, CIS Controls v8 and NIST SP 800-53 Rev 5 set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | GV.RM-01 — Risk Management Strategy | Routing policy should be judged by total workflow risk and cost, not just selection price. |
| Recommendation — Set routing metrics around successful outcome cost, retry rate, and completion latency. | ||
| CIS Controls v8 | CIS-18 — Penetration Testing and Red Team Exercises | Operational validation of workflow choices needs measurable testing, not assumed savings. |
| Recommendation — Test routing policies under realistic task loads and compare rework rates. | ||
| NIST SP 800-53 Rev 5 | RA-5 — Vulnerability Monitoring and Scanning | Task-fit failures act like control weaknesses that should be monitored and corrected. |
| Recommendation — Continuously monitor routing outcomes and fix policies that generate repeated failures. | ||
Practitioner Guidance
What to prioritise: Measure routing on successful task completion, not model selection cost alone. The most useful metric is cost per acceptable outcome, because that captures retries, escalation, and latency in one view.
What to verify: Check whether the cheaper route increases fallback frequency, manual review, or output repair. If any of those rise, the routing policy is probably transferring cost rather than removing it.
Decision rule: If the routed model cannot reliably meet the task quality bar on the first pass, treat the lower price as misleading unless the downstream workflow is designed to absorb the extra failure rate.
Practitioner takeaway: Cheap routing is only valuable when it reduces end-to-end cost, because a low-cost decision that creates more downstream work is still an expensive workflow.