For the past two years, the AI narrative has been remarkably consistent:

"Replace expensive human work with AI. Ship software faster. Spend less."

Reality is turning out to be much messier.

Many companies rushed to reduce hiring, freeze engineering teams, or expect existing developers to produce significantly more code with AI assistants. The assumption seemed obvious: if AI writes code in seconds, engineering costs should fall.

Instead, a growing number of enterprises are discovering something unexpected.

Their software engineering payroll hasn’t disappeared. It has simply been replaced by a rapidly growing token bill.

The New Software Tax

Unlike traditional software licences, modern AI coding tools don't simply charge per user.

They charge for usage.

Every prompt.

Every revision.

Every agent.

Every background task.

Every automated code review.

The more successful AI becomes inside an engineering team, the more tokens it consumes.

That is already showing up in the market. Gartner-backed reporting says many teams are seeing AI coding bills rise from around $20 or $100 per developer per month into the $2,000 to $5,000 range, with some extreme cases much higher, and Gartner predicts AI coding costs could overtake the average developer salary by 2028 if usage is not controlled The Register, Computer Weekly.

Ironically, the companies embracing AI most aggressively are often seeing their AI costs grow the fastest.

Recent reporting also suggests many enterprises are introducing usage limits, governance rules, and model-routing strategies after discovering AI spending was rising far faster than expected Computer Weekly. Some organizations are even exhausting annual AI budgets within months because developers and autonomous agents are consuming tokens at a pace finance teams never modeled The Register.

Productivity Didn't Scale the Way People Expected

AI unquestionably makes individual developers faster.

The mistake was assuming faster code generation automatically translates into faster software delivery.

Writing code has never been the bottleneck.

Understanding code has.

Most engineering teams still spend enormous amounts of time on:

• Reviewing pull requests
• Understanding AI-generated code
• Finding hidden bugs
• Learning unfamiliar repositories
• Maintaining software months after it ships

AI accelerates the production of code.

It doesn't automatically accelerate human understanding.

In many cases it actually creates more work.

More Code ≠ Better Software

When code generation becomes almost free, something interesting happens.

Teams naturally create more of it.

Much more.

Large pull requests become common. Repositories grow faster. Reviewers skim instead of reading. Technical debt accumulates quietly.

And the downstream result is often more churn. One 2026 set of developer productivity benchmarks reports that AI-generated code is being rewritten at a significantly higher rate than human-written code, and other AI productivity analyses show code churn rising sharply in AI-heavy environments Larridin, Code Board.

Developers begin trusting AI-generated code because it looks correct.

But looking correct and being understood are two very different things.

The result?

More hidden bugs.

More fragile systems.

More knowledge trapped inside AI conversations instead of engineering teams.

Token Costs Are Becoming a Board-Level Conversation

This isn't speculation anymore.

Industry analysts are warning that AI coding costs may eventually exceed developer salaries if organizations don't actively govern token usage and model selection. Gartner predicts that by 2028, AI coding costs could overtake the average software developer's salary without proper controls.

Even large consulting firms are responding.

Accenture recently instructed employees to avoid using AI for simple tasks after executives observed rapidly escalating token spending that wasn't producing proportional business value. Similar cost controls are reportedly appearing across several major technology companies.

Meanwhile, UBS reports that roughly 60% of enterprise customers they interviewed are actively throttling AI usage because token costs have become a material budgeting concern.

This isn't an anti-AI story.

It's an AI governance story.

The Real Productivity Metric

The question isn't:

"How much code did AI write?"

It's:

"Does anyone actually understand what was merged?"

Because software isn't expensive to write.

It's expensive to maintain.

Every engineer who inherits AI-generated code six months later pays interest on today’s shortcuts. Every reviewer who approves code they don’t fully understand increases future engineering costs. Every undocumented AI-generated change becomes tomorrow’s tribal knowledge.

And the data increasingly supports that this isn’t just a feeling. DORA’s 2025 reporting found that higher AI adoption is associated with increased software delivery throughput, but also a negative relationship with delivery stability Google Cloud, DORA. In other words, teams may be shipping more, but not necessarily shipping safer.

The Companies That Win Won't Use Less AI

They'll use AI more intelligently.

They'll optimise model selection.

They'll manage token costs.

They'll measure outcomes instead of prompts.

Most importantly, they'll make sure every merge leaves behind understanding - not just more code.

Because the future of software engineering isn't about replacing developers.

It's about making sure humans remain capable of understanding what AI creates.

And that may turn out to be the most valuable engineering capability of all.