Compare commercial models by running the same workload through each one. For an AI SDR agent cost estimate, specify the sales workflow, CRM integrations, contact volume, human review, and what counts as an accepted result. Write down what the seller calls a seat, task, successful outcome, included unit, and overage. Then calculate both the invoice estimate and cost per accepted unit. The examples below use invented numbers solely to show the arithmetic; they are not provider prices or market benchmarks.
Pricing models
- Per seat: a recurring amount per enabled user or licensed seat. Check whether inactive users count, what access each seat includes, and whether consumption limits are separate.
- Per task or transaction: a charge for a defined unit of work. Confirm how retries, partial completion, duplicate requests, and failed tasks are counted.
- Usage based: charges for tokens, tool calls, runtime, storage, or another metered unit. Include the exact input/output categories, cache treatment, service mode, and any minimum commitment.
- Per outcome: a charge tied to an agreed result. Define acceptance, evidence, exclusions, dispute handling, and what happens when a human must finish the work.
- Fixed period or tier: a fixed subscription or a tier with included volume and overage rules. Model both expected usage and the first unit above an included threshold.
Real offers can combine these models. Use the provider’s current official price page, written quote, and contract as the source of truth. For underlying model or platform consumption, consult the applicable OpenAI API pricing or Google Agent Platform pricing page as relevant; those pages do not represent a third-party agent vendor’s full commercial offer.
Examples
Assume one month has 50 enabled users, 12,000 submitted tasks, and 1,000 tasks that meet a separately defined acceptance test. Consider three fictional offers:
| illustrative model | calculation | subtotal | effective cost per accepted task |
|---|---|---|---|
| Seat | 50 seats × $40 | $2,000 | $2.00 |
| Task | 12,000 submitted tasks × $0.20 | $2,400 | $2.40 |
| Outcome | 1,000 accepted tasks × $2.50 | $2,500 | $2.50 |
All amounts above are fabricated. The totals exclude implementation, overages, support, and internal review. The seat row divides the license subtotal by the accepted-task count for comparison; it does not mean seats are billed per accepted result. If acceptance falls to 800 while spend stays the same, the effective costs become $2.50, $3.00, and—if only accepted outcomes are billable—$2.50 respectively. Verify each contract’s definition before using that last assumption.
Pros and cons
Seat pricing can make a steady user population easier to budget, while usage limits or uneven seat utilization can change its effective cost. Per-task or usage pricing makes volume sensitivity visible, while context length, retries, and tool behavior may make each unit vary. Outcome pricing can align the fee with accepted work, while the acceptance test and exception process require careful agreement. Fixed tiers support forecasting within the included band, while crossing a threshold can change the marginal charge. These are structural tradeoffs; none identifies a universally cheaper model.
Calculate at least a low, expected, and high workload using the same period and accepted-work definition. Add setup, implementation, security review, human correction, and exit/portability effort to every option where they apply. The FinOps Foundation’s Unit Economics capability recommends relating technology spend to an appropriate unit and value measure; a product-specific comparison still needs your own definitions and evidence.
Evaluating vendors
Ask for a written definition of billable units and a sample invoice at your forecast volume. Record plan name, billing period, currency, minimums, included units, overage tiers, retries and failures, taxes, renewal, support, implementation, data retention, and exit terms. Test an ordinary case and a long-context or retry-heavy case. For outcome pricing, agree how acceptance is measured and how partial or disputed work is handled before comparing the headline amount.
Keep quoted commercial charges distinct from model/API estimates and from internal allocations. The AI agent cost calculator models user-entered task, step, token, tool, fixed, and human-cost assumptions; the LLM API cost calculator and token cost calculator isolate user-entered token scenarios. They do not fetch vendor offers, calculate seat or outcome contracts, or verify a quote.