Why AI Agent Pricing Is Still Unsolved and What It Means for How You Track Costs

AI agents' pricing remains an unsolved issue with significant implications for cost tracking. Unlike traditional SaaS offerings, AI agents provide uncertain value at fluctuating costs, challenging conventional pricing logic. Each pricing model on the market today is based on a distinct definition of value, leading to various unit-economic problems.

Monitoring the appropriate metrics is crucial, as different pricing models require monitoring different aspects internally. For instance, a product profitable in aggregate may still incur losses on specific features. A case study revealed that fixing a monthly loss of around $800K saved the company a considerable sum. The pricing challenges arise from the nature of AI agents' value and cost, which are not stable or deterministic.

Resolution-based billing models hinge on closed tickets as the primary value indicator, but this can incentivize closing tickets instead of solving problems, resulting in lost revenue when complex cases are escalated. Conversation-based billing models, on the other hand, charge for every interaction, regardless of resolution, simplifying budgeting but assuming all conversations hold value.

Cost-plus pricing, built on LLM costs and a margin, fails to account for value, and seat-based pricing tracks user access rather than the agent's performance. These models are not inherently wrong but are defined by different notions of value, each creating unique unit-economic challenges. Pricing metrics also dictate what internal metrics need monitoring for profitability.

For AI agents billed per resolution, tracking the cost per resolution (total LLM spend divided by resolved tickets) becomes essential, as changes in resolution rates can significantly impact margins. In contrast, AI agents billed per conversation focus on forecasting conversation volume, which is more predictable than resolution rates.

Thus, the pricing model chosen fundamentally influences which metrics to monitor to maintain profitability. Additionally, each AI agent feature possesses its unique model, prompt length, and call pattern, necessitating separate tracking of costs and billing rates. Aggregate numbers may appear profitable, but unit-level analysis might reveal losses within specific cohorts.

Model choice and prompt complexity further complicate cost tracking, as different features incur varying costs depending on conversation context. In summary, AI agent pricing remains an unsettled issue with significant consequences for cost tracking. Each pricing model carries its unique assumptions regarding value and cost, leading to distinct unit-economic challenges.

Monitoring the appropriate metrics is crucial for maintaining profitability, as different models require tracking different internal metrics.