The cheapest unit price is rarely the cheapest system
A per-resolution price is attractive because it is small and easy to compare. It is also easy to misunderstand. With usage-based pricing, success is the thing you are charged for. The better the agent works, the more resolutions it handles, and the larger the bill becomes.
That can be a good deal at low volume. It can be punishing at scale. The right comparison is not unit price. It is three-year cost, including implementation, usage, review labor, monitoring, knowledge maintenance, and switching cost.
The full cost stack
Real agent cost has more layers than the line item. Platform or license fees are only the start. Usage can be priced per resolution, conversation, credit, or overage. Implementation includes discovery, integration, configuration, and policy work.
Custom implementations add model usage, hosting, retrieval infrastructure, and monitoring. Both platform and custom paths include human review. In regulated or high-value workflows, review labor can be one of the largest recurring costs.
Then there is knowledge maintenance, eval upkeep, analytics, compliance review, and switching cost. If prompts, logs, memory, and workflow logic cannot be exported, the future cost of leaving is part of the price of entering.
- Platform fees and seat costs.
- Usage fees, credits, overages, and demand spikes.
- Implementation and integration work.
- Human review and exception handling.
- Monitoring, evals, knowledge maintenance, and migration risk.
The crossover math
Most pricing decisions come down to the crossover between usage-based platform pricing and an owned or managed custom build. Put both on the same three-year horizon. Platform cost is roughly monthly resolutions times 12 times 3 times price per resolution, plus platform fees and review labor. Raise the resolution count each year if volume is growing.
Custom cost is one-time implementation plus three years of model usage, infrastructure, monitoring, support, and review labor. The model-usage and review portions also scale with volume, not just the platform side.
Illustrative example: 8,000 resolutions a month at 1.50 dollars per resolution is 144,000 dollars a year in usage. Add 40,000 dollars a year in review labor and the platform path is about 184,000 dollars a year, or 552,000 dollars over three years.
Now compare a custom path with 120,000 dollars in implementation and 90,000 dollars a year to run, including model usage and review. Year one is 210,000 dollars. Years two and three add about 180,000 dollars. Three-year cost is roughly 390,000 dollars - cheaper here by about 160,000 dollars. But run both sides honestly: a custom build's model and review costs also rise with volume, just more slowly than per-resolution fees. The number that matters is your break-even volume, the monthly resolution count where the two three-year totals meet. Below it the platform wins; above it ownership does.
Use your real numbers
Those numbers are illustrative. At 800 resolutions a month, the platform likely wins because the custom build is hard to amortize. At 80,000 resolutions a month, usage pricing may become unacceptable. The point is to solve for your crossover volume before signing.
The hidden costs buyers undercount
Human review is the most common missing line. It can be small for low-risk drafting and large for regulated or high-stakes work. Integration maintenance is the next one. Connectors, permissions, and data contracts change, and someone has to keep them working.
Migration risk is the quiet one. It costs nothing until you need to leave. Then the lack of portable prompts, traces, memory, and workflow code becomes expensive quickly.
How to run the decision
Collect monthly volume, current human time per run, escalation rate, systems involved, quality requirements, and expected growth. Model platform and custom paths with the same assumptions. Stress-test high volume, high exception rate, and spike months.
Then decide on the curve, not the sticker. If you cannot yet name the workflow owner, monthly volume, success metric, and exception rate, the honest answer may be to wait or run a short audit before buying anything.
Get the terms in writing before the curve stops mattering: how a billable resolution is defined, whether a reopened ticket or clarifying follow-up is billed once or twice, overage rates, volume cliffs, annual caps or floors, and exit rights covering export of prompts, logs, memory, and workflow logic. A low unit price with an elastic definition of resolution, uncapped overages, and no export rights is not a low price.
- Collect volume, labor, systems, risk, and quality inputs.
- Model platform, custom, and hybrid paths over three years.
- Stress-test spike months and exception-heavy cases.
- Choose the cost curve that matches volume, control, and ownership needs.