AI agent utilisation rate measures how much of an agent’s available capacity is spent on productive, billable work versus idle time, overhead, and wasted effort. It is the productivity metric professional services firms already use for humans — now applied to AI.
Professional services firms track human utilisation religiously. A consultant billing 80% of their available hours is productive. One billing 60% is a problem. Yet most firms deploying AI agents have no idea how productive those agents actually are. Early benchmarks suggest 40–60% utilisation is typical for newly deployed AI agents, with significant room for improvement. This guide covers what utilisation means for agents, three formulas to calculate it, what affects it, and how to raise it.
Key Takeaway: Measure AI agent utilisation the same way you measure human utilisation — then use the data to improve productivity and justify AI investment.
What Does Utilisation Rate Mean for AI Agents?
Human utilisation is straightforward: billable hours divided by available hours. Target rates sit between 75% and 85% at most professional services firms. The formula is well understood and universally applied.
AI agent utilisation is different because agents do not have fixed working hours. An agent is technically available 24 hours a day, 7 days a week. It does not take lunch breaks, attend meetings, or go on holiday. This makes “available time” a misleading denominator if taken literally.
Instead of raw availability, firms need to think about capacity utilisation and billing utilisation separately. Capacity utilisation asks: what percentage of the agent’s processing capacity is being used? Billing utilisation asks: what percentage of the agent’s completed work is billable to a client?
There is also the concept of effective utilisation — which factors in output quality and human review overhead. An agent that completes 100 tasks but requires human corrections on 40 of them has a lower effective utilisation than one that completes 80 tasks with no corrections needed.
The distinction matters because a high-capacity agent running at 90% load but producing mostly non-billable internal work has excellent capacity utilisation and poor billing utilisation. You need both numbers.
How Do You Calculate AI Agent Utilisation?
Three formulas work, depending on your billing model and what data you collect. Most firms should start with one and add the others as their AI agent time tracking matures.
Formula 1: Task-Based Utilisation
(Billable tasks completed / Total tasks completed) x 100
This works when your agents handle discrete, countable tasks — document reviews, data extractions, report drafts. If an agent completed 120 tasks last week and 84 were billable to clients, its task-based utilisation is 70%.
Task-based utilisation is simple to track and easy to explain to stakeholders. Its weakness is that it treats all tasks equally. A five-minute formatting task counts the same as a two-hour research assignment.
Formula 2: Time-Based Utilisation
(Time spent on billable work / Total active time) x 100
This works when you track agent execution time per task. If an agent was active for 40 hours last week and spent 28 hours on billable client work, its time-based utilisation is 70%.
Time-based utilisation maps closely to the human metric that firms already understand. Its challenge is defining “active time.” Do you count only execution time, or also queuing time, retry time, and idle time between tasks?
Formula 3: Cost-Based Utilisation
(Billable AI costs / Total AI costs) x 100
This works when cost attribution matters more than time. If your firm spent £2,000 on AI compute last month and £1,300 of that was for billable client work, cost-based utilisation is 65%.
Cost-based utilisation is useful for firms that bill AI costs directly to clients. It captures the financial efficiency of agent deployment. Its weakness is that it conflates model pricing with productivity — a cheaper model running the same tasks will show different utilisation than an expensive one.
What Is a Good Utilisation Rate for AI Agents?
There is no industry standard yet. Early adopters in professional services report 40–60% utilisation for newly deployed agents, rising to 65–80% after six months of optimisation. These numbers will shift as firms refine task routing, prompt engineering, and agent configuration.
For comparison, human utilisation targets of 75–85% took decades to establish. AI agent benchmarks are still forming. The important thing is to start measuring now and track trends over time.
What Factors Affect AI Agent Utilisation?
Six factors drive utilisation up or down. Each one is worth monitoring independently.
Demand variability is the biggest factor. Unlike humans who find other productive work during quiet periods, agents sit idle when no tasks are queued. Uneven demand — peaks on Monday, lulls on Friday — drags down weekly utilisation even if the agent performs well during busy periods.
Task routing efficiency determines whether the right tasks reach the right agent. Poor routing sends complex tasks to simple agents (causing failures) or simple tasks to powerful agents (wasting capacity). Matching task complexity to agent capability directly affects how much productive work gets done.
Error and retry rates consume capacity without producing value. If an agent fails on 20% of tasks and retries each one twice, that is 60% more compute time spent on the same volume of work. High error rates are the silent killer of utilisation.
Human review bottlenecks stall the pipeline. If agents produce outputs faster than humans can review them, a queue builds up. The agent may be technically idle while waiting for approval to start its next task. This is a workflow problem, not an agent problem.
Agent configuration matters too. A poorly prompted agent spends time on low-value processing steps, generates unnecessarily verbose outputs, or makes redundant tool calls. Each of these wastes capacity.
Model selection affects both speed and accuracy. Running a large reasoning model on a simple classification task wastes tokens and time. Running a small model on a complex analysis task causes failures and retries. Right-sizing the model to the task is a direct utilisation lever.
How Do You Improve AI Agent Utilisation?
Improvement starts with measurement. You cannot fix what you do not track. Once you have baseline numbers, six strategies move utilisation upward.
Optimise task queuing. Batch related tasks together. Prioritise high-value billable work during peak hours. Use scheduling to spread demand more evenly across the day and week.
Reduce error rates. Better prompts, clearer guardrails, and validated inputs prevent failures before they happen. Track which task types generate the most errors and fix those first — the return on effort is highest where failure rates are highest.
Streamline human review. Implement tiered review based on risk. Auto-approve low-risk outputs (formatting, data extraction) while routing high-risk outputs (legal analysis, financial recommendations) through full human review. Parallel review — where multiple outputs are reviewed in a batch — is faster than sequential one-at-a-time checking.
Expand agent scope. As confidence grows, assign agents to more task types. An agent that handles three task categories will find productive work more often than one limited to a single category. Expand gradually and monitor quality metrics as you do.
Right-size deployment. Do not provision more agent capacity than demand requires. Excess capacity inflates the denominator without adding productive work. Scale capacity to match actual demand patterns, not theoretical peak load.
Track weekly trends. Utilisation that drops over three consecutive weeks signals a problem — declining demand, rising errors, or a configuration drift. Weekly reporting catches degradation early, before it becomes a quarterly surprise.
How Does AI Utilisation Compare to Human Utilisation?
The two metrics share a purpose — measuring productive output — but differ in important ways.
| Factor | Human Utilisation | AI Agent Utilisation |
|---|---|---|
| Capacity | Fixed (40–50 hours/week) | Elastic (scales with demand) |
| Target range | 75–85% | 40–60% (early), 65–80% (mature) |
| Non-billable time | Admin, meetings, training | Setup, monitoring, retries |
| Measurement | Hours tracked manually | Logged automatically per task |
| Improvement levers | Reduce admin, improve delegation | Better prompts, task routing, model selection |
Humans have fixed capacity. An employee works 40–50 hours per week. AI agents have elastic capacity that scales up or down based on demand and infrastructure. This means utilisation targets for agents should not simply copy human targets.
Human non-billable time includes meetings, administration, training, and internal projects. Agent non-billable time includes configuration, testing, monitoring, error handling, and internal processing tasks. The overhead categories are different, but the principle is the same — reduce overhead and more capacity goes to billable work.
The most useful metric for many firms is blended utilisation — the combined productive output of human-AI teams on a project. If an agent handles 30% of a project’s tasks and a human handles 70%, blended utilisation considers both contributions. This gives a more accurate picture of how AI costs compare to human costs on real client work.
AI agents can also improve human utilisation by handling low-value tasks that would otherwise consume a consultant’s billable hours. If an agent takes over document formatting, data entry, and first-pass research, the human spends more time on high-value billable work. The agent’s utilisation might be moderate, but its impact on human utilisation can be significant.
Tracking cost per task alongside utilisation gives you the full picture — not just how busy your agents are, but how efficiently they convert spend into billable output.
Keito tracks both human and AI agent utilisation rates in a single dashboard — one metric, one view of your entire workforce.
Frequently Asked Questions
What is AI agent utilisation rate?
AI agent utilisation rate measures the percentage of an agent’s capacity that is spent on productive, billable work. It is calculated by dividing billable output (tasks, time, or costs) by total output and expressing the result as a percentage.
How do you calculate AI agent utilisation?
Three formulas work depending on your billing model. Task-based: billable tasks divided by total tasks. Time-based: time on billable work divided by total active time. Cost-based: billable AI costs divided by total AI costs. Each is multiplied by 100 to produce a percentage.
What is a good utilisation rate for AI agents?
Early benchmarks suggest 40–60% is typical for newly deployed agents. After six months of optimisation — better prompts, improved task routing, reduced errors — firms report 65–80%. Industry standards are still forming, so the priority is measuring consistently and tracking trends.
How does AI agent utilisation compare to human utilisation?
Human utilisation targets sit between 75% and 85% at most professional services firms. AI agent targets are lower because agents have elastic capacity and different overhead profiles. The key difference is that humans have fixed working hours while agents scale with demand.
How can you improve AI agent utilisation rates?
Six strategies: optimise task queuing to spread demand evenly, reduce error rates through better prompts and guardrails, streamline human review with tiered approval, expand agent scope to more task types, right-size deployment to match actual demand, and track utilisation weekly to catch degradation early.
Does AI agent utilisation affect billing?
Yes. Low utilisation means you are paying for agent capacity that is not generating revenue. High utilisation with strong billing attribution means more of your AI spend flows through to client invoices. Utilisation data also helps justify AI charges to clients who question the value.
Should you track AI utilisation alongside human utilisation?
Yes. Blended utilisation — the combined productive output of human and AI on a project — gives the most accurate picture of team productivity. Tracking both in the same system avoids siloed metrics and lets you see how AI deployment affects overall firm utilisation.