AI cost per task

What it is

AI cost per task converts scattered spending on models, APIs, vector search and human checking into a single figure per completed job, such as one drafted brief, one enriched record or one resolved support ticket. The denominator matters: only outputs that pass quality review should count. Tracked this way, it exposes workflows that look cheap per call but expensive per finished result.

Why it matters

Teams often budget by subscription or token spend, which hides the real driver of cost, namely retries and human rework. Cost per task makes AI investment comparable with the manual process it replaces and with other channels, so it can be defended in a budget conversation. It also guides where to spend effort: prompt changes, cheaper models, better retrieval or tighter scope.

How it works

Define the task and its acceptance criteria, instrument the workflow to log tokens, tool calls and wall-clock human review time per attempt, then divide total period cost by accepted outputs. Practitioners compare variants by routing a share of tasks to a smaller model or shorter context and watching whether the acceptance rate holds. Caching, retrieval trimming and early failure checks usually move the number more than raw model price.

When it applies

It applies whenever an AI workflow runs at volume and repeatedly, such as content production pipelines, data enrichment, classification, support deflection or internal research, and especially before scaling a pilot.

Examples

• A content team calculates that one published product page costs 1.20 pounds in model calls plus 25 minutes of editor time, and uses that to decide how many pages to commission.
• A support lead compares cost per resolved ticket between an AI first-line assistant and the previous human-only queue.
• An ops team finds that a third of enrichment records fail validation and retry, so the true cost per accepted record is well above the per-call figure.

How it is measured

• Cost per accepted task, with model, tooling and human review costs separated
• Acceptance rate on first attempt, and average retries per completed task
• Human minutes per task, tracked as the workflow matures
• Cost per task versus the manual baseline it replaces, reviewed monthly