AI agent cost per task is the total cost of every resource consumed when an agent completes a single unit of work — tokens, inference, tool calls, infrastructure, and human oversight — attributed to that specific task.
An AI research agent can cost £0.02 for a simple lookup or £12 for a complex multi-source investigation. Without per-task measurement, most firms treat both as free. They absorb the costs as overhead, misprice client work, and have no basis for deciding whether an agent or a person should do the job.
Cost per task is the foundational metric for AI agent economics. It tells you what autonomous work actually costs, whether it is cheaper than human labour, and whether you can profitably bill clients for it.
Key Takeaway: Cost per task is the unit economics of AI agents. Without it, you cannot bill accurately or compare agent vs human costs.
Why Is Cost Per Task the Key Metric?
Raw API bills tell you how much you spent in total. They do not tell you how much any individual piece of work cost. Cost per task bridges that gap.
It Connects Spending to Business Outcomes
A monthly API bill of £3,000 means nothing on its own. Knowing that your research agent costs £0.45 per client research task — and runs 6,000 tasks per month — tells you where the money goes. It also tells you whether that spend is worth it.
It Supports Agent vs Human Comparison
A junior analyst costs £35 per hour and completes four research tasks per hour. That is £8.75 per task. If the AI agent does the same work for £0.45, the business case is clear. If the agent costs £12 per task for complex research that the analyst does for £8.75, the calculus flips.
Without AI agent cost per task data, this comparison is guesswork.
It Drives Accurate Client Billing
Firms billing clients for AI-assisted work need to know what each task costs. Whether you pass costs through directly, blend them into hourly rates, or charge a flat project fee, per-task cost data is the foundation.
For more on building a full cost tracking system, see our AI agent cost tracking guide for professional services.
What Goes Into an AI Agent Task Cost?
The total cost of an AI agent task has four components. Most firms track only the first.
Direct Costs
These are the costs directly generated by the agent executing the task:
- Token consumption: Input tokens, output tokens, and reasoning tokens. Output tokens cost three to five times more than input. Reasoning tokens can multiply costs by five to twenty times.
- Inference compute: The GPU time consumed to process the request. Varies by model size and provider.
- Tool and API calls: External services the agent invokes — web searches, database queries, code execution, third-party data sources. A single task can trigger five to twenty tool calls.
Indirect Costs
These are shared costs that apply across many tasks:
- Orchestration overhead: The compute cost of running the agent framework itself — routing, planning, memory management.
- Prompt engineering time: The human hours spent designing, testing, and refining the prompts the agent uses. Spread across all tasks that use those prompts.
- Error handling and retries: When an agent fails and retries, the retry tokens are real costs. A coding agent retrying a failing test ten times burns ten times the expected token cost.
Infrastructure Costs
These are ongoing costs that support agent operations:
- Vector database fees: If agents use retrieval-augmented generation, each query costs money for embedding generation and database lookup.
- Monitoring and logging: Observability platforms, cost dashboards, and audit logs.
- Security and access control: Systems that prevent agents from accessing the wrong client data.
Human Oversight Costs
AI agents in professional services require human review. This is a real cost that should be included in per-task calculations.
- Review time: A senior consultant spending 10 minutes reviewing an agent’s output at £80/hour adds £13.33 to the task cost.
- Correction effort: When the agent gets it wrong, someone fixes it. That correction time is a cost of the agent task, not a separate activity.
- Quality assurance: Spot checks, compliance reviews, and client-readiness checks all add time.
The Total Task Cost Formula
Total task cost = direct costs + (indirect costs / total tasks) + (infrastructure costs / total tasks) + oversight costs
For firms running thousands of tasks per month, the indirect and infrastructure components per task become small. For firms running hundreds, they can be significant.
How Do You Measure AI Agent Cost Per Task?
Three measurement approaches exist. Choose based on your firm’s technical maturity and task volume.
Method 1: API-Level Instrumentation
The most accurate method. Instrument your agent framework to capture token counts, model pricing, and tool call fees on every request. Aggregate all requests that belong to a single task using a task identifier.
This requires:
- Middleware in your agent framework that logs every LLM call
- A task ID that propagates through all agent actions
- A cost calculation layer that applies current pricing to raw token counts
Best for: Firms with engineering capacity to instrument their agent stack.
Method 2: Sampling
Measure a representative batch of 50 to 100 tasks per workflow type. Record the total cost of each. Calculate the mean, median, and 90th percentile.
This gives you a reliable per-task cost estimate without full instrumentation. Update the sample quarterly as models, prompts, and workflows change.
Best for: Firms starting out with cost measurement, or those with limited engineering resources.
Method 3: Time-Bounded Measurement
Track all AI costs within a task’s execution window. If a task runs from 14:00 to 14:03, sum all API calls, tool calls, and compute costs in that window.
This works when tasks run sequentially. It breaks down when tasks overlap or when agents work on multiple tasks concurrently.
Best for: Simple, sequential agent workflows with clear start and end boundaries.
For a detailed token costing methodology, see our AI token cost calculator.
What Do AI Agent Tasks Actually Cost? Benchmarks by Type
Cost per task varies enormously by complexity. The table below shows benchmarks based on professional services workloads using standard-tier models in 2026.
| Task Type | Typical Token Range | Cost Range | Average at Scale |
|---|---|---|---|
| Simple classification or routing | 500–2,000 | £0.01–£0.05 | £0.02 |
| Document summarisation (single doc) | 5,000–50,000 | £0.05–£0.50 | £0.15 |
| Deep research with synthesis | 50,000–200,000 | £0.50–£5.00 | £1.80 |
| Code generation (feature-level) | 10,000–100,000 | £1.00–£10.00 | £3.50 |
| Multi-agent orchestrated workflow | 100,000–500,000+ | £5.00–£50.00+ | £18.00 |
Important context: These benchmarks cover direct costs only. Add 20–40% for indirect, infrastructure, and oversight costs to get the full per-task figure.
The per-task average across professional services firms running mixed workloads sits at £0.10–£0.30 for routine tasks at scale. Complex tasks skew the average upward significantly.
Why the Range Is So Wide
A simple classification task — “Is this email a complaint or an enquiry?” — consumes a few hundred tokens and costs pennies. A multi-agent research workflow — “Analyse these 15 contracts, identify risk clauses, and produce a summary memo” — can consume hundreds of thousands of tokens, trigger dozens of tool calls, and cost £20 or more.
The variance is the challenge. Firms that budget based on averages get surprised by outliers. Budget based on 90th percentile costs, not means.
Model Choice Dramatically Affects Per-Task Cost
The same task can cost 20x more on a frontier reasoning model than on a lightweight model. A document summary that costs £0.02 on a Tier 3 model costs £0.40 on a Tier 1 model. Most professional services tasks do not require frontier reasoning.
Matching model tier to task complexity is the single biggest lever for controlling AI agent cost per task. Route simple classification to lightweight models. Use standard models for drafting and summarisation. Reserve frontier models for genuine multi-step reasoning tasks.
Tracking Cost Per Task Over Time
Per-task costs are not static. They change as models update, prompts evolve, and usage patterns shift. Track the trend, not just the snapshot.
A rising cost-per-task trend signals prompt bloat, model tier creep, or increased task complexity. A falling trend suggests your team is getting better at prompt engineering or that provider pricing has dropped. Either way, the trend informs decisions.
How Do You Use Cost Per Task to Inform Billing and Staffing?
Cost per task data answers two strategic questions: What should we charge clients? And should an agent or a person do this work?
Agent vs Human Cost Comparison
Calculate the human cost per task: divide the hourly rate by the number of tasks a person completes per hour.
| Work Type | Human Cost Per Task | Agent Cost Per Task | Agent Savings |
|---|---|---|---|
| Email classification | £2.00 (2 min at £60/hr) | £0.02 | 99% |
| Document summary | £15.00 (15 min at £60/hr) | £0.15 | 99% |
| Legal research memo | £35.00 (30 min at £70/hr) | £1.80 | 95% |
| Code review | £25.00 (20 min at £75/hr) | £3.50 | 86% |
| Complex strategy advice | £50.00 (30 min at £100/hr) | £18.00+ (often needs human rework) | Variable |
For more on this comparison, see our AI agent vs human cost comparison.
When Are Agents Cheaper?
Agents win on high-volume, repetitive, low-judgement tasks. The savings are dramatic — often 90% or more. The key conditions:
- Task is well-defined with clear inputs and outputs
- Quality requirements are met without heavy human review
- Volume justifies the instrumentation and monitoring costs
When Are Humans Cheaper?
Humans win on novel, ambiguous, or relationship-dependent tasks. When an agent task requires 20 minutes of human review on a 5-minute agent run, the net saving is negative. The agent generated the output, but a human spent four times longer checking it.
Setting Client Billing Rates
Three billing models work for AI agent tasks:
- Cost-plus: Pass through the actual agent cost plus a margin (typically 30–50%).
- Blended rate: Fold agent costs into your standard hourly rate. As agents do more work, margins improve.
- Task-based pricing: Charge per deliverable regardless of whether a human or agent produced it. Your cost per task determines your margin.
Frequently Asked Questions
How do you calculate AI agent cost per task?
Sum all direct costs (tokens, inference, tool calls), add a proportional share of indirect costs (orchestration, prompt engineering), infrastructure costs (vector databases, monitoring), and human oversight costs (review time, corrections). The formula: total task cost = direct + (indirect / tasks) + (infrastructure / tasks) + oversight.
What is the average cost of an AI agent task?
For routine professional services tasks at scale, the average is £0.10–£0.30 per task. Simple classification costs £0.01–£0.05. Complex multi-agent workflows cost £5–£50 or more. The wide range reflects differences in task complexity, model choice, and tool usage.
How does AI agent cost per task compare to human labour?
For repetitive, well-defined tasks, agents typically cost 86–99% less than human labour. A document summary that costs a consultant £15 in time costs an agent roughly £0.15. For complex, ambiguous tasks requiring judgement, the gap narrows or reverses when you factor in human oversight and correction time.
What cost components are included in an AI agent task?
Four categories: direct costs (tokens, inference, tool calls), indirect costs (orchestration, prompt engineering, error handling), infrastructure costs (vector databases, monitoring, security), and human oversight costs (review, correction, quality assurance). Most firms track only direct costs, missing 40–60% of the true total.
How do you track cost per task across multiple agents?
Assign a unique task identifier that propagates through all agent actions, including calls to sub-agents. Aggregate all costs — from every agent involved — under that task ID. This requires instrumentation at the orchestration layer, not just at individual agent endpoints.
Keito measures AI agent cost per task automatically, attributed to clients, projects, and billing codes. Start measuring task costs today.