How to price AI agents when the work is variable

An agent is a workflow, not a single model call

Pricing an AI agent requires a different level of care from pricing a simple chat feature. An agent may plan, retrieve information, call tools, ask a model to reason, recover from failure, and make several decisions before it produces a result. Two tasks that look identical from the customer's perspective can require very different paths through that system.

That variability does not make agent pricing impossible. It means the team must understand the task boundary and the cost distribution behind it. Start by defining the customer outcome you are selling, then measure what it costs to complete that outcome across simple, ordinary, and complex cases. The goal is to price a dependable result, not to expose every internal agent step on an invoice.

Define what counts as a completed task

The commercial unit should be tied to a result that can be verified. Examples include a research brief delivered, a case triaged, a support ticket resolved, a sales account enriched, or an approved workflow completed. An agent run by itself is usually not enough because a run may fail, be retried, or finish without a useful outcome.

Define completion carefully. If human review is part of the value proposition, an approved task may be more honest than an attempted task. If the agent creates a draft, a completed draft with the required source material may be the appropriate unit. Clear definitions protect both the customer and the business when usage, quality, and billing are reviewed later.

Measure the expensive paths before committing to a price

Agent cost can expand through planning loops, tool calls, long contexts, high-reasoning models, fallbacks, and retries. Track the median and p90 cost per completed task, then break that view down by task type, model route, and tool chain. A growing p90-to-average gap usually means a subset of work deserves close attention, either because it is commercially valuable or because the workflow needs redesign.

This is an application of AI unit economics, not a reason to reduce every task to its cheapest route. Some high-cost tasks may be exactly what an enterprise customer values most. The question is whether the plan, commitment, or outcome price recognizes that value. Where complexity is intrinsic, a tiered task allowance or a custom enterprise arrangement may be more honest than pretending all agent work costs the same.

Choose a model that matches the customer promise

Subscription pricing can work when the agent is a persistent part of a team's operating system and ordinary activity fits within a clear allowance. Usage-based pricing can work when customers understand the completed task and its volume varies materially. A hybrid plan often provides the best balance: a base commitment for access, configuration, and governance, plus included tasks and transparent overage for sustained high-complexity work.

Outcome-based pricing deserves special caution with agents. It is compelling when the result is measurable, attributable, and stable, but it becomes risky when external systems, human approvals, or customer inputs determine the outcome. Price the part of the result you can observe and stand behind. Do not turn a vague business aspiration into an invoice trigger.

Protect quality and margin together

Cost controls that degrade a customer's result are not a pricing strategy. When an agent is expensive, investigate the delivery path before reducing its capability. Retrieval quality, tool selection, prompt structure, retry handling, model routing, and stopping conditions can often improve both cost and reliability. The best intervention may be a better workflow rather than a lower limit.

Establish thresholds for task cost, p90 task cost, retries, latency, and outcome quality. Give each threshold an owner and a response path. Our AI cost monitoring guide explains how to set that operating rhythm, while the AI gross margin guide helps teams connect the result to revenue and plan-level profitability.

Continue reading

Questions teams ask