Model the economics of autonomous agents, net of inference cost.
Built for AI, digital and operations leaders scoping their first production agent program. Autonomy rate and run cost are the levers that decide whether the business case holds.
Your inputs
Benchmarks show typical enterprise ranges — override every field with your own numbers.
Multi-step knowledge-work tasks an agent could own end-to-end
Benchmark: 20k–150k for a single process family
Human end-to-end time today, including tool switching
Benchmark: 8–25 minutes for mid-office knowledge work
Share of tasks completed end-to-end with no human escalation
Benchmark: 35–60% in year one on a scoped process
Human still finishes, but the agent has done the retrieval and drafting
Benchmark: 15–30%
Benchmark: $80k–$120k in US enterprise operations
Includes retries, evaluation calls and tool invocations
Benchmark: $0.15–$0.80 per multi-step run
Agent design, tool integration, evaluation harness, guardrails, change management
Benchmark: $300k–$800k for a production agent program
Labor value returned minus inference and orchestration cost, at your autonomy rate.
Directional estimate. Assumes 1,560 productive hours per worker per year and a fully-loaded cost per successful agent run including retries.
Three-scenario view
Finance reviewers expect a range. These scenarios flex adoption and implementation cost around the model you entered.
Slower adoption, higher integration effort
Your inputs as entered
Strong sponsorship, clean data, phased scale-up
How enterprise leaders use this model
- What makes agentic ROI different from automation ROI?
- Agentic systems act across multiple steps and tools, so value is driven by the share of tasks completed end-to-end without escalation — not by per-transaction time savings alone. Autonomy rate is the single biggest swing factor.
- What autonomy rate should we assume in year one?
- Most enterprise programs land between 35% and 60% fully autonomous completion in the first year on a scoped process family, rising as guardrails, evaluations and tool coverage mature.
- Should we include inference cost?
- Yes. Agentic runs consume far more tokens than single-shot copilots. The model here charges a fully-loaded cost per successful run, including retries and orchestration overhead.
- How do we defend this model to finance?
- Run the conservative scenario first, hold the aggressive case for the second phase, and share the URL — your inputs are encoded in the link so reviewers see the identical model.