Artificial intelligence was supposed to enable companies to develop software more quickly and cheaply. However, Gartner warns that by 2028, the costs of AI-assisted coding could exceed the average programmer’s salary if the growing consumption of tokens is not linked to the value of the work performed.
This forecast sounds like a prediction of the moment when employing a human will become cheaper than using a model. In practice, the risk is more complex. A company does not choose between a programmer and an agent, but pays simultaneously for the specialist’s labour, the tool licence, model usage and the subsequent review of the generated code. AI does not eliminate the cost of the team. It adds a new, variable layer to it, the cost of which depends on how the work is organised.
From a fixed licence to an agent’s bill
The first AI tools for developers were sold like traditional SaaS software. A company would purchase a specific number of user licences and know the monthly cost well in advance. This model provided predictability for the finance department, even if it was difficult to measure the benefits precisely.
Coding agents are changing this economic model. Increasingly, the per-user fee covers only a specific usage allowance, with further work billed on a pay-as-you-go basis. GitHub already combines Copilot licences with AI credits, charging additional fees once a company’s allocation has been used up. A long session with an agent working on multiple files and using a more complex model costs more than a brief query directed at a lighter model.
For businesses, this marks a shift from a per-role cost to a cost-per-task model. The number of developers is no longer sufficient to forecast expenditure. Two employees with identical licences can generate completely different bills if one mainly uses autocomplete, whilst the other runs agents that analyse entire repositories.
This is a model well known from the cloud. A server or service may be cheap on paper, but the bill depends on the architecture, traffic volume and usage patterns. In AI coding, the number of calls, context length, model choice and level of autonomy play a similar role.
It is not writing the code that costs the most, but the entire operational cycle
A simple assistant responds to a command or suggests a snippet of code. An agent, however, is given a task and independently carries out a series of operations: it analyses the project structure, locates the necessary files, plans a solution, generates changes, runs tests, interprets errors and retries.
Each such step may involve another model invocation and the resending of a large portion of the context. If the agent gets stuck in a loop, selects too broad a scope of data, or repeatedly attempts to fix the same problem, the cost rises without commensurate progress. Gartner reports that the transition from traditional assistants to agent-based coding has, in some cases, led to bills as much as a hundred times higher.
The token price is therefore only one element of the equation. The design of the process is of greater importance. An agent using an expensive model for a few minutes can be cost-effective if it solves a valuable problem. A cheap agent running for hours, constantly sending the same files and producing rejected changes, remains a cost, even if each individual call appears insignificant.
This relationship shifts responsibility for the budget from the procurement department to the architecture of the development environment. Negotiating licence prices is not enough when the size of the bill is determined by the rules for selecting models, the size of the context and the scope of autonomy granted to the agent.
Cheaper models do not guarantee lower expenditure
The cost of using artificial intelligence is falling rapidly. The Stanford AI Index indicates that between November 2022 and October 2024, the cost of inference for a model reaching the GPT-3.5 level fell by more than 280-fold.
However, this does not necessarily mean a reduction in a company’s bills. Cheaper models pave the way for running more processes, assigning longer tasks to agents and integrating AI into further stages of work. A light user quickly becomes a heavy user once the tool is integrated into their daily workflow.
This effect is similar to that observed in the cloud. The fall in the cost of a single computing unit did not halt the rise in total expenditure, as companies began to process more data and run more services. In the case of coding agents, the increase in consumption may outpace the fall in token prices.
However, productivity does not increase at the same rate across every organisation
A high AI bill is not a problem in itself if it delivers greater value growth. The difficulty lies in the fact that developers’ productivity cannot be reduced to a single universal figure.
Field experiments involving 4,867 developers showed an average 26 per cent increase in the number of tasks completed after an AI assistant was made available. Less experienced staff saw greater benefits.
A study by METR, conducted amongst experienced programmers working on their own open-source projects, yielded different results. In an experiment conducted in early 2025, the use of AI increased task completion times by 19 per cent, although the participants themselves were convinced that the tools had sped them up. More recent data from 2026 suggests an improvement in AI capabilities, but the authors emphasise that it is not yet possible to reliably determine the scale of productivity gains.
This discrepancy does not mean that one of the studies is incorrect. It shows that the outcome depends on the type of task, the user’s experience, the quality of the repository and the maturity of the process. DORA defines AI as an enhancer: the technology amplifies the strengths of efficient organisations, but at the same time exacerbates the effects of chaos, poor testing and technical debt.
For a company, the conclusion is uncomfortable but important. Simply making agents available does not create productivity. In an environment where the code is well documented, tasks are clearly defined, and tests quickly catch errors, AI can reduce the workload. In a chaotic system, an agent generates further problems more quickly whilst consuming more tokens.
The cost does not end with the generation of the code
The cost of tokens is the easiest to measure, as it appears on the invoice. It is much harder to quantify the time spent checking, correcting and maintaining the results of the AI’s work.
In a Stack Overflow survey, 66 per cent of developers cited AI responses that are almost correct but still contain errors as their main frustration. For 45 per cent, debugging AI-generated code took longer.
This is important because an agent may improve their local productivity metric whilst simultaneously worsening the outcome of the entire process. More generated code looks good in activity statistics, but it can increase the number of reviews, defects, fixes and incidents. The time saved during the creation of the first version is then passed on to testers, architects, security and maintenance teams.
Comparing the cost of AI solely with the time spent writing code therefore gives an incomplete picture. The economics begin with the instruction given to the agent, but only end when the change is working in production and does not generate disproportionate maintenance costs.
A token is not a measure of value
The biggest mistake would be to manage agents solely by limiting the number of tokens. A token is the provider’s billing unit, not a measure of business impact.
A company can significantly reduce consumption by routing every task to the cheapest model. If, in doing so, the quality of responses drops, the number of retries increases and developers’ working time is extended, the apparent saving will actually increase the total cost. It is equally misguided to grant all agents access to the most expensive models in the belief that a higher price automatically means a better result.
A far more important metric is the cost of an accepted change: how much the model’s operation cost, how much time people spent on it, how long implementation took, and whether the code required subsequent corrections. Only by combining this data can we see whether an agent is creating value or merely increasing activity.
This shifts the conversation from the question ‘how much are we spending on AI?’ to ‘what do we get for every zloty spent?’. Without such a shift, token reduction remains an accounting optimisation, whilst increasing the budget is a gamble based on declared productivity.
FinOps enters the software development process
Controlling agent expenditure is beginning to resemble cloud cost management. Organisations need visibility into consumption at the team and process levels, rules for routing tasks to different models, and mechanisms to detect unusual spikes in consumption.
This shift is already evident in FinOps practices. According to the State of FinOps 2026 report, 98 per cent of organisations surveyed manage their AI expenditure, compared with 31 per cent two years earlier. Managing AI costs has also become the most important new capability being developed by FinOps teams.
This does not mean simply transferring control over development tools to the finance department. An effective model combines a financial perspective with engineering expertise. FinOps can identify where costs are rising, but it is up to the technical team to assess whether the increase stems from a valuable task, a poorly chosen model or an uncontrolled agent loop.
Gartner’s forecast therefore does not conclude that AI-assisted coding will become uneconomical. It shows that agents are ceasing to be a low-cost add-on to a developer’s role and are becoming a fully-fledged production resource. The greater the proportion of the work they take on, the less relevant the licence price becomes, whilst the economics of the entire process take on greater significance.
The greatest risk is not the token price, but scaling agents without knowing how much a single accepted and implemented software change costs.
