Enterprise systems were designed with people in mind who carry out single, predictable tasks. An AI agent works differently: a single command can trigger a series of queries to models, databases, applications and external services. The load no longer increases in proportion to the number of users, but to the number of actions performed autonomously.
According to Google’s‘2026 State of AI Infrastructure’report, 83 per cent of IT leaders believe their infrastructure needs modernising to support AI systems on a larger scale. Only 17 per cent are fully confident that their current technology stack can cope with agents carrying out critical processes.
This does not, however, mean that most companies are facing a complete overhaul of their infrastructure. Agentic AI primarily reveals limitations that were previously invisible, as the pace of work was set by humans. When an agent queries several systems within seconds, repeats operations and launches further tools, local imperfections begin to affect the entire process.
A single prompt can involve hundreds of operations
The chatbot answers a question. The agent plans the next steps, selects tools, retrieves data, interprets the results and assesses whether the task has been completed. Each of these activities utilises the infrastructure and incurs a cost.
Therefore, a pilot project conducted on a limited number of tasks says little about the economics of a production deployment. An agent connected to CRM, ERP, financial systems and analytical tools begins to generate costs in multiple areas simultaneously.
Scaling agents therefore requires a different approach to that of deploying yet another application. The number of users ceases to be the primary measure of load. What becomes more important is the number of steps, model invocations, data queries and retries performed within a single process.
Cheaper models do not guarantee lower bills
The cost of a single model invocation is falling, but agentic AI increases the number of invocations needed to achieve a result. Data transfer, memory, search, API calls, event logging, error handling and human involvement in decision verification all add to the bill.
The token price therefore provides an increasingly poor indication of the solution’s true cost. An agent using a cheap model can be expensive if it performs many unnecessary operations or regularly repeats failed actions.
A fall in the unit price may actually accelerate the rise in total expenditure. Cheaper models make it easier to deploy a larger number of agents and assign them more complex tasks. The savings on a single operation are then offset by the scale of usage.
From a business perspective, the cost of a completed process is more significant: a processed order, a resolved customer issue, a prepared quotation or a closed incident. Such a metric shows whether an agent is actually improving operational efficiency.
An older system is not always a problem
The narrative surrounding agentic AI often leads to the conclusion that legacy environments must be replaced with new, integrated platforms. In practice, a stable transactional system can still perform its function well. The problem arises when an agent uses it via fragile integrations, has no query limits, or works with inconsistent data.
The biggest limitation is often not the age of the technology, but the quality of the interfaces, data and access controls. Poorly documented APIs, differing definitions of the same information and scattered data sources increase the risk of errors far more rapidly than the age of the system itself.
This changes the logic of modernisation. Rather than replacing the entire environment, greater value may be derived from improving the integration layer, organising data and reducing the traffic generated by agents. Modernisation then becomes a targeted intervention in critical paths, rather than a multi-year programme to rebuild everything.
Agent control becomes part of the architecture
The more decisions an agent makes independently, the greater the importance of its identity, scope of authorisation and the ability to trace subsequent actions. An agent using a shared technical account may perform operations that cannot later be attributed to a specific task or decision.
Therefore, governance is no longer merely a compliance layer. It is becoming part of the operational infrastructure. An organisation needs to know what data an agent has used, which systems they have run, how much the task cost, and at what point a human approved the result.
Limits on the number of steps, the budget allocated to a task, and the ability to halt a process and undo operations all help to mitigate risk. Without these mechanisms, autonomy not only increases productivity but also the scope for potential error.
Not every process needs an agent
The pressure to implement agentic AI encourages the replacement of simple automation with more complex systems. Meanwhile, processes based on fixed rules are often still better handled by traditional automation, scripts or RPA.
An agent delivers the greatest value where there is incomplete information, a variable context and a need for interpretation. If the outcome of a process can be described using clear rules, the model’s autonomy may increase costs without improving quality.
The difference between an agent and automation is therefore not simply a matter of the level of technological sophistication. It concerns matching the solution to the nature of the process and the value created by the system’s greater autonomy.
Modernisation focuses on the agent’s course of action
The general slogan of ‘preparing the infrastructure for AI’ easily leads to costly migration that does not solve the problems of a specific process. Analysing the agent’s full workflow provides much more insight: data sources, the number of calls, where delays occur, failed operations and the extent of human involvement.
Only on this basis can one see which element is limiting scalability. In one case it may be integration; in another, data quality, access control or the way tasks are routed to different models.
Agentic AI does not destroy legacy systems. It reveals which elements of the architecture cannot keep up with the machine’s pace of operation. Companies that are able to assign cost, permissions and outcome to each agent action gain the ability to scale AI without rebuilding the entire environment.
The most important change, therefore, does not involve replacing the entire IT stack, but rather measuring and restructuring those pathways through which the agent actually creates business value.
