AI Coding Costs Could Outpace Developer Salaries by 2028: What Gartner Got Right

The result is a budget that behaves the way cloud compute did back in 2014, before anyone built the tooling to tame it. Usage moves with workload, not headcount. A sprint with heavy agent use can cost more than the sprint before it, even with the same team, the same sprint length, and the same number of story points closed.

Token burn is a scoping decision, not a billing event

This is the part Gartner's headline gets right and most of the response to it gets wrong. The instinct after reading that AI coding costs could exceed salaries is to reach for a cost dashboard. A dashboard tells you what already happened. It does not touch the decision that caused it.

Correlation One's 2026 research into enterprise token economics makes the mechanism explicit: token cost scales with usage and design decisions, not seat counts, which makes it, in their words, an engineering-behavior problem wearing a finance costume. That framing matters because it moves the fix upstream, into how work gets specified and handed to an agent in the first place, rather than downstream into a procurement negotiation.

Consider what actually drives a token bill up. A vague ticket forces an agent to explore a codebase broadly before it can act, and every file it reads is billed. A poorly scoped task triggers repeated retry loops, each pass rebuilding context that could have been cached. An agent given open-ended autonomy over a multi-file refactor will call the model far more times than one given a tight, pre-scoped diff to produce. None of that shows up on an invoice as a line labeled "bad scoping." It shows up as tokens, and the invoice never explains why.

The clearest illustration of what happens when this goes fully ungoverned came from the open-source AI project OpenClaw, whose creator disclosed that roughly 100 autonomous coding agents burned more than $1.3 million in API tokens in a single month. OpenAI absorbed that bill as research support. Most engineering organizations will not have that option, and Gartner's own findings back up how common the underlying pattern is: AI governance firm Exceeds found that 80% of companies exceed their AI budgets because spend stays invisible without structured tracking.

What governed AI development actually looks like

Gartner's own recommendations point in the right direction: build a use-case-driven decision framework, match model choice to task complexity, mandate context engineering, and put usage reviews into the development cycle itself rather than the monthly finance review. In practice, that breaks down into decisions an engineering team makes before an agent is ever invoked.

Model routing is the first lever, and it is the one most teams skip. Not every task needs a frontier model running at premium token rates. Formatting a file, writing a routine test, or completing a well-defined function can run on a smaller, cheaper model without a quality loss anyone will notice. Save the expensive model for the reasoning-heavy work it was built for.

Context discipline is the second lever. Every retrieved document, every line of conversation history, every repeated system instruction an agent carries into a task adds to the token count before the model has done anything useful. Trimming what an agent is allowed to see for a given task is not a limitation. It is scope control applied to a machine that will otherwise read everything it has access to.

The third lever is the one most organizations build last and need first: attribution. Token spend has to trace back to the commit, the ticket, or the feature it produced, the same way cloud spend eventually got tied back to the service that generated it. Without that link, a rising bill is just a rising bill. With it, a rising bill becomes a signal about which parts of the codebase are expensive to touch, and why.

What it actually takes to run this well

None of this is a settings toggle. It is a change to how an engineering team scopes work, and that is where most AI coding governance efforts stall out, because it asks developers to slow down at exactly the moment the tools promise speed.

At EPixelSoft, every engagement that puts AI coding agents into a client's delivery pipeline starts with the same question Gartner's report is really asking: what is this task supposed to produce, and what is the smallest, most tightly scoped version of that task an agent can be handed. A well-specified ticket with a defined boundary costs a fraction of what an open-ended one costs, independent of which model runs it. That discipline is not glamorous. It is also the one lever that works before the invoice arrives instead of after.

The honest reality is that this costs engineering leadership time up front, in specification, in reviewing agent output before it merges, in deciding which tasks even qualify for agent handling versus a developer working directly. Teams that skip that step still get the productivity gains Gartner acknowledges are real. They also get the bill Gartner is warning about, and by the time it shows up, the habits that caused it are already load-bearing.

The forecast is a deadline, not a surprise

Gartner did not predict something unknowable. It described a trajectory that is already visible in the spending data of the leaders it surveyed, and gave the industry a date by which the trend, left alone, closes the gap between tool cost and human cost. The organizations that treat 2028 as a countdown to build governance will spend the next two years turning token discipline into a habit. The ones that treat it as a headline will spend those same two years explaining an AI coding bill that grew faster than anyone budgeted for.

If your engineering team is scaling AI coding agents without a scoping and governance model behind them, EPixelSoft's AI Transformation practice can help you build one before the token bill does it for you.

EPixelSoft is an AI-native software engineering company based in Noida, India. Since 2014, we have shipped 700+ production systems across FinTech, HealthTech, NGO operations, and SaaS.