In October, Elon Musk described xAI's Grok Bot agent as the fastest-growing product he had ever seen, noting that people were using it more each week, not less. Within weeks, as users complained that their usage limits were running out quickly, he wrote that "the cost of serving Grok Bot was unsustainably high relative to revenue", Forbes reported.
It is a vivid illustration of a challenge many enterprises are about to meet on their own scale. An employee asks an AI assistant a question and receives an answer. An AI agent receives an assignment and starts working.
That difference may look small on a screen. For the business paying for the work, it can change the economics considerably.
One assignment, many inference calls
An agent reviewing a supplier agreement might retrieve documents, compare clauses, consult internal policies, draft findings and check its own conclusions. What appears to be one request can involve repeated model calls, growing context and several rounds of tool use. Workflows branch into subtasks, repeat steps when results are incomplete, and several agents may run in parallel.
Anthropic's published account of its multi-agent research system shows the scale of the difference: in its data, agents typically used about four times as many tokens as chat interactions, and multi-agent systems about fifteen times as many. Those figures describe one implementation rather than a universal forecast, but they show why a budget based on chatbot usage may be a poor guide to an agent deployment.
The number of employees using AI is only part of the calculation. How often workflows run, how complex they are and how they execute can matter just as much.
A bigger bill can still be good business
Higher consumption does not automatically mean worse economics. An agent that uses more tokens but completes a valuable task accurately may be a sound investment; a cheaper workflow that repeatedly fails or sends work back to employees may deliver less.
Price per million tokens remains useful for comparing components, but enterprise decisions also need the cost of an acceptable business outcome – a compliant report, or a customer case resolved correctly and on time, including retries and human review. Without that connection, a low token price can create false confidence, while a rising bill can hide a productive deployment.
Capacity planning starts with the workload
At Nebula Data, where our NEB Compute platform focuses on enterprise AI compute, we see capacity planning as a question that begins with the workload, not the hardware specification.
How many workflows a business can complete, how quickly, and what happens during a demand spike depend on the models being served, context lengths, concurrency, latency requirements and how the application coordinates its calls. An overnight batch process may tolerate queuing; an agent supporting a live customer interaction may not. A long-running process needs continuity across stages and clear handling of failed calls.
That means the infrastructure conversation must include the people who design and run the business process. Technology leaders should test realistic workflows before making large capacity commitments. Finance leaders should ask how consumption changes as deployment expands. Business leaders should define the quality and turnaround that make the automation worthwhile.
Match the purchasing model to the workload
No single purchasing model suits every agent deployment. Usage-based APIs offer flexibility while teams experiment, particularly when workloads are uncertain or depend on proprietary models. Dedicated GPU capacity can make sense for sustained workloads using models that can run on that infrastructure, with economics that depend on utilisation and operational capability. A hybrid approach can combine dedicated capacity for predictable demand with external services for specialist models or peaks.
These options are not interchangeable. Securing GPUs does not provide access to every proprietary model, and owning capacity does not guarantee efficient use. The procurement decision should follow the workload assessment, not precede it.
Design in visibility and limits
A monthly invoice may show that spending rose, but not which workflow caused it or whether the extra work was valuable. Enterprises should be able to attribute consumption to teams, applications or customers, and to see retries, failed tasks and unusually long execution paths.
Practical controls include limits on execution steps, budget thresholds, appropriate model selection and escalation to a person when a task stops making progress. A business needs to know both that an agent can finish its assignment and that it will stop when the work exceeds its authority or budget.
Plan for the work, not just the tool
The attraction of AI agents is that they take on more of a task. That same capability makes their resource needs harder to estimate from a demonstration or a subscription price – as even the fastest-growing agent products are discovering.
Before expanding a deployment, enterprises should establish what a successful task looks like, what it costs to complete and what capacity is needed to deliver it reliably. Businesses that connect workflow value, inference consumption and infrastructure planning will be best placed to turn promising agents into productive services.