Direct answer: do not expand an agent because its cost per credit looks low. Record four denominators in sequence: the platform billing unit, agent run, accepted work unit, and incremental business outcome. Add consumption, licensing, integration, data, human review, failure, and support. Then calculate cost per released campaign, accepted lead, reactivated customer, or another unit Marketing and Finance recognize. A credit is an input—not a result.
Platforms already bill on incompatible structures. HubSpot publishes units by content item, email, lead, response, resolution, or minute and supports simulation and run limits. Microsoft manages credits by environment and agent through prepaid capacity or pay-as-you-go. Salesforce converts actions and prompts into Flex Credits with different multipliers, including production and sandbox treatment. Adding credits across providers or comparing them as a common commodity creates false precision.
Build the Agent Consumption-to-Outcome Ledger
- Workload — agent, version, environment, objective, owner, and cost center.
- Eligible trigger — contact, campaign, record, or event that justified the run; deduplication and reason.
- Technical consumption — runs, actions, prompts, credits, tokens where available, tools, retries, and effective price.
- Accepted work — released campaign, approved email, corrected record, researched lead, or resolved conversation; quality criterion.
- Exception — rejection, correction, review, error, repetition, timeout, escalation, and human hours.
- Outcome — conversion, pipeline, contribution, retention, or realized capacity; window, comparator, and maturity.
- Decision — retain, repair, cap, reallocate, or stop; date, owner, budget, and rollback.
The ledger preserves quantity and cost. Credits alone omit price; dollars alone omit which action consumed them; runs alone hide loops and retries. Version the effective rate card and currency on the transaction date. Do not overwrite history when a provider changes multipliers.
Use four costs, not one
- Cost per run = variable agent cost ÷ completed runs. Useful for technical diagnosis, not ROI.
- Cost per accepted unit = total operating cost ÷ work that passed quality criteria and was approved. This exposes real operating efficiency.
- Cost per eligible treated = total cost ÷ population that could receive the treatment. This prevents failures and exclusions from disappearing.
- Cost per incremental outcome = incremental total cost ÷ outcomes above the comparator. It exists only when the measurement design supports incrementality.
Total operating cost includes credits and overage, required license, connectors, storage, data, observability, evaluation, human review, agency, support, and incidents. Implementation may be amortized over an approved horizon but must remain visible. Do not treat included credits as free: record usage and the marginal price that scale would trigger.
Allocate cost by why the run happened
One agent may research a lead, update CRM, draft an email, and enroll a journey. Allocate by observed trigger and action—not by the employee who clicked or the department that purchased the license. If one run serves two campaigns, use a documented driver such as eligible volume, actions, or processing time, and keep shared cost separate when allocation would be arbitrary.
Testing, evaluation, and sandbox work may also consume capacity. HubSpot documents no-credit simulations for supported agents; Salesforce publishes sandbox multipliers; Microsoft states that consumption varies with design, frequency, and features. Separate development, evaluation, and production. Saving credits by reducing testing can increase production failure and review cost.
Illustrative US example
A hypothetical midsize B2B company operates three agents for campaign creation, account research, and nurture personalization. In one month it records 12,000 technical units, 1,800 runs, 1,050 outputs, 640 approved work units, and 410 actually used. Illustrative operating cost is $18,000 including consumption, labor, and support. These figures are not a benchmark or MAKINAI client result.
Cost per run is $10; per generated output, $17.14; per accepted unit, $28.13; and per used unit, $43.90. None proves incremental pipeline. The campaign agent loops because briefs are incomplete, while the account-research agent costs more per run but produces a higher accepted-lead rate. The rational decision is not to cut the agent with the highest credit cost; it is to repair intake for the first and retain the second until mature pipeline evidence arrives.
Run a six-week observability pilot
- Week 1 — inventory agents, environments, rate cards, capacity, contracts, owners, and cost centers.
- Week 2 — choose one accepted unit per workload, identify triggers, and freeze volume, quality, and cost baselines.
- Week 3 — join consumption exports, agent logs, CRM, review, and outcome events with IDs and timestamps.
- Week 4 — run in shadow mode or a bounded cohort; reconcile estimated, billed, and executed consumption and explain differences.
- Week 5 — apply an agent limit, pacing alert, exception queue, and comparator; do not hard-stop critical work without fallback.
- Week 6 — calculate four costs, read quality and mature outcomes, and retain, repair, reallocate, or stop.
Set alerts that change a decision
- Consumption — credits per eligible unit or run leave the observed range.
- Quality — acceptance falls, correction rises, or the agent stops abstaining when evidence is missing.
- Operations — retries, timeouts, unavailable tools, or human queues grow.
- Budget — pacing forecasts overage before period end.
- Business — cost per accepted unit worsens without better conversion, speed, or quality.
- Continuity — a hard stop would block a critical campaign or customer workflow without an alternate path.
A hard financial stop may fit reversible experiments; a critical workflow may need an alert and explicit approval instead. Prepaid capacity improves predictability but can strand spend. Pay-as-you-go absorbs peaks but increases variance. One environment cap is easy to administer and can let a noisy agent starve others. Compare volume and quality scenarios, not only the published unit price.
FinOps, Marketing, Data, RevOps, and Engineering need one unit. FinOps normalizes cost and forecast; Marketing defines accepted work and outcome; Data joins IDs and timestamps; RevOps confirms eligibility and state; Engineering observes execution and failure; Procurement versions the commercial rate; the business owner decides priority. Charging cost to an agent without assigning outcome to a decision creates chargeback, not unit economics.
Measure value with attribution limits
Compare quality, speed, and cost with the current process and simplest alternative. For revenue, preserve a holdout, split, or another comparator when feasible and respect the buying cycle. Before-and-after analysis mixes seasonality, offer, media, mix, and team learning. Platform reports prove consumption and activity; they do not prove the agent caused pipeline or margin. When volume cannot support causal measurement, disclose uncertainty and use cost per accepted work unit as the operating metric.
Use https://makinai.co/insights/en/platform-campaign-agent-autonomy-marketing-results to bound what the agent may change, https://makinai.co/insights/en/first-ai-marketing-workflow-cost-complete-budget to build the full first-workflow budget, and https://makinai.co/insights/en/marketing-team-uses-ai-results-unchanged-value-leak to locate the leak between output and result. If consumption is rising and nobody can explain cost per accepted unit, bring MAKINAI at https://makinai.co/services/en/ai-strategy-transformation-consulting the credit exports, one agent log, 30 days of outputs, and the business event. We can scope the instrumentation and pilot without replacing the entire stack.